Skip to content

types

Exports simple type definitions used in MEDS torchdata.

BatchMode

Bases: StrEnum

An enumeration of the possible batch modes for the dataset.

Attributes:

Name Type Description
SEM

Subject-Event-Measurement mode. In this mode, data are represented as 3D tensors of sequences of measurements per event per subject, with tensor shapes [batch_size, max_events_per_subject, max_measurements_per_event].

SM

Subject-Measurement mode. In this mode, data are represented as 2D tensors of sequences of measurements per subject, without explicit separation between measurements of different events, with tensor shapes [batch_size, max_measurements_per_subject].

Source code in meds_torchdata/types.py
class BatchMode(StrEnum):
    """An enumeration of the possible batch modes for the dataset.

    Attributes:
        SEM: Subject-Event-Measurement mode. In this mode, data are represented as 3D tensors of sequences of
             measurements per event per subject, with tensor shapes
             `[batch_size, max_events_per_subject, max_measurements_per_event]`.
        SM: Subject-Measurement mode. In this mode, data are represented as 2D tensors of sequences of
            measurements per subject, without explicit separation between measurements of different events,
            with tensor shapes `[batch_size, max_measurements_per_subject]`.
    """

    SEM = "SEM"
    SM = "SM"

MEDSTorchBatch dataclass

Simple data structure to hold a batch of MEDS data.

Can be accessed by attribute (e.g., batch.code) or string key (e.g. batch["code"]). The elements in this tensor can take on several shapes, and keys can be present or omitted, depending on details of dataset configuration. To clarify these shape options, we’ll define the following terms. Most of these terms will also be realized as properties defined on this class for accessing shape variables over the batch for convenience.

  • batch_size is the number of subjects in the batch.
  • max_events_per_subject is the maximum number of events (unique time-points) for any subject in the batch.
  • max_measurements_per_event is the maximum number of measurements (observed code/value pairs) for any event in the batch (across all subjects).
  • max_static_measurements_per_subject is the maximum number of static measurements observed across all subjects in the batch.
  • max_any_measurements_per_event is the maximum number of measurements that are either static for a given subject or observed in any event for a given subject across the batch (e.g., max(max_measurements_per_event, max_static_measurements_per_subject)).
  • max_measurements_per_subject is the maximum number of measurements observed across all dynamic events for any given subject, in total, in the batch.
  • max_any_measurements_per_subject is the maximum number of measurements observed for any subject regardless of whether they are dynamic or static.

There are a few shape “modes” that this batch can be in, depending on the configuration of the source dataset. These include:

  • "SEM": In Subject-Event-Measurement (SEM) mode, the data is represented as a tensor of measurements per-event, per-subject, with missing values padded in all dimensions.
  • "SM": In Subject-Measurement (SM) mode, the data is represented as a tensor of measurements per-subject, with events concatenated in order with neither per-event padding nor explicit separator tokens.

Under each of these modes, different sets of the core attributes take on different consistent shapes.

Under all modes:

  • Static data elements (static_code, static_numeric_value, and static_numeric_value_mask) are of shape [batch_size, max_static_measurements_per_subject].
  • The label tensor, boolean_value tensor is of shape [batch_size].

In SEM Mode:

  • Per-event data (time_delta_days & event_mask) are of shape [batch_size, max_events_per_subject] if static data is not prepended and shape [batch_size, max_events_per_subject + 1] if static data is prepended. time_delta_days will have no zeros at any position save the last event per subject, for which position the time delta to the next event may be unknown, and, in the case where static data has been prepended into the sequence, the first event per subject (which will contain static data and has no time delta).
  • static_mask is of the same shape as the per-event data and will have True at event indices that correspond to the static event (currently only the first event) and False otherwise.
  • Per-measurement data (code, numeric_value, & numeric_value_mask) are of shape [batch_size, max_events_per_subject, max_measurements_per_event] if static data is not prepended and shape [batch_size, max_events_per_subject + 1, max_any_measurements_per_event] if static data is prepended. All measurements in the first event if static data is prepended will be static measurements.

In SM Mode:

All tensors are of shape [batch_size, max_measurements_per_subject] if static data is not prepended and [batch_size, max_any_measurements_per_subject] if static data is prepended.

  • time_delta_days will have zeros at measurement indices that correspond to either static measurements or measurements that do not correspond to the last measurement in an event, or at the last measurement in the sequence if the next time-delta is unknown.
  • static_mask will be of the same shape as the measurement level data and will have True at indices that correspond to static measurements and False otherwise.
  • event_mask is omitted.
  • Per-measurement data (code, numeric_value, & numeric_value_mask) has the same shape given above.

Attributes:

Name Type Description
time_delta_days FloatTensor | None

Tensor of time deltas between sequence elements, in days.

event_mask BoolTensor | None

Boolean tensor indicating whether a given event is present or not.

code LongTensor | None

Measurement code integral vocabulary indices. Equals PAD_INDEX when measurements are missing.

numeric_value FloatTensor | None

Measurement numeric values. No guaranteed value for padding or missing numeric values.

numeric_value_mask BoolTensor | None

Boolean mask indicating whether a given measurement has a numeric value. Values of this mask for padding measurements are undefined.

static_mask BoolTensor | None

Boolean mask indicating whether a given measurement or event is a static measurement/event or a true dynamic measurement/event. Only used when static data is prepended into the dynamic sequence. When the batch is in SEM mode this will correspond to a mask with a True at the first event and false otherwise.

static_code LongTensor | None

Static measurement code integral vocabulary indices. Equals PAD_INDEX when measurements are missing.

static_numeric_value FloatTensor | None

Static measurement numeric values. No guaranteed value for padding or missing numeric values.

static_numeric_value_mask BoolTensor | None

Boolean mask indicating whether a given static measurement has a numeric value.

boolean_value BoolTensor | None

Per-sample boolean labels.

Examples:

>>> batch = MEDSTorchBatch(
...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.2]]),
...     event_mask=torch.tensor([[True, True], [True, False]]),
...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
...     numeric_value=torch.tensor(
...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
...     ),
...     numeric_value_mask=torch.tensor([
...         [[True, False, True], [False, False, False]],
...         [[False, True, False], [True, True, True]] # Note the padding values may be  True or False
...     ]),
... )

The batch is effectively merely an ordered (by the definition in the class, not order of specification), frozen dictionary of tensors, and can be accessed as such:

>>> print(batch["code"])
tensor([[[1, 2, 3],
         [3, 0, 0]],
<BLANKLINE>
        [[5, 6, 0],
         [0, 0, 0]]])
>>> print(batch["event_mask"])
tensor([[ True,  True],
        [ True, False]])
>>> print(list(batch.keys()))
['code', 'numeric_value', 'numeric_value_mask', 'time_delta_days', 'event_mask']
>>> print(list(batch.values()))
[tensor(...), tensor(...), tensor(...), tensor(...), tensor(...)]
>>> print(list(batch.items()))
[('code', tensor(...)), ('numeric_value', tensor(...)), ('numeric_value_mask', tensor(...)),
 ('time_delta_days', tensor(...)), ('event_mask', tensor(...)]
>>> batch["code"] = torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]])
Traceback (most recent call last):
    ...
ValueError: MEDSTorchBatch is immutable!

Though note that if you manually define something in a batch to be None, it will not be present in the keys/values/items:

>>> batch = MEDSTorchBatch(
...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.2]]),
...     event_mask=torch.tensor([[True, True], [True, False]]),
...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
...     numeric_value=torch.tensor(
...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
...     ),
...     numeric_value_mask=torch.tensor([
...         [[True, False, True], [False, False, False]],
...         [[False, True, False], [True, True, True]]
...     ]),
...     boolean_value=None,
... )
>>> print(list(batch.keys()))
['code', 'numeric_value', 'numeric_value_mask', 'time_delta_days', 'event_mask']

The batch can also be accessed by attribute, and has default values for allowed fields:

>>> print(batch.event_mask)
tensor([[ True,  True],
        [ True, False]])
>>> print(batch.boolean_value)
None

The batch has a number of properties that can be accessed for convenience:

>>> print(batch.mode)
SEM
>>> print(batch.static_inclusion_mode)
omit
>>> print(batch.has_labels)
False
>>> print(batch.batch_size)
2
>>> print(batch.max_events_per_subject)
2
>>> print(batch.max_measurements_per_event)
3
>>> print(batch.max_measurements_per_subject)
None
>>> print(batch.max_static_measurements_per_subject)
None

Batches exist in one of several combinations of modes across the “batch mode” and the “static data inclusion mode”. Batch mode can either be BatchMode.SEM/"SEM" or BatchMode.SM/"SM", and static data inclusion mode can be StaticInclusionMode.PREPEND/"prepend", StaticInclusionMode.INCLUDE/"include", or StaticInclusionMode.OMIT/"omit". The batch mode reflects the shape of the batch’s elements (being either organized at an event X measurement level vs. at a measurement level) and the static data inclusion mode reflects how static data is included in the batch.

Note

These modes are determined implicitly by the organization of the data in the batch, not explicitly via passed flags or anything.

The batch comes with a useful print representation function that clearly indicates what modes the batch is in, which we can use below:

Subject-Event-Measurement (SEM) Mode

In SEM mode, the batch is organized as a tensor of measurements per event per subject, indicated by a 3D structure of the batch’s main data elements (codes and numeric values).

Static Data OMIT/"omit" Mode

In this mode, no static data is included.

>>> batch = MEDSTorchBatch(
...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.2]]),
...     event_mask=torch.tensor([[True, True], [True, False]]),
...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
...     numeric_value=torch.tensor(
...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
...     ),
...     numeric_value_mask=torch.tensor([
...         [[True, False, True], [False, False, False]],
...         [[False, True, False], [True, True, True]]
...     ]),
... )
>>> print(batch)
MEDSTorchBatch:
│ Mode: Subject-Event-Measurement (SEM)
│ Static data? ✗
│ Labels? ✗
│
│ Shape:
│ │ Batch size: 2
│ │ Sequence length: 2
│ │ Event length: 3
│ │
│ │ Per-event data: (2, 2)
│ │ Per-measurement data: (2, 2, 3)
│
│ Data:
│ │ Event-level:
│ │ │ time_delta_days (torch.float32):
│ │ │ │ [[1.00, 2.10],
│ │ │ │  [4.00, 0.20]]
│ │ │ event_mask (torch.bool):
│ │ │ │ [[ True,  True],
│ │ │ │  [ True, False]]
│ │
│ │ Measurement-level:
│ │ │ code (torch.int64):
│ │ │ │ [[[1, 2, 3],
│ │ │ │   [3, 0, 0]],
│ │ │ │  [[5, 6, 0],
│ │ │ │   [0, 0, 0]]]
│ │ │ numeric_value (torch.float32):
│ │ │ │ [[[ 1.,  0., -3.],
│ │ │ │   [ 0.,  0.,  0.]],
│ │ │ │  [[ 0.,  0.,  0.],
│ │ │ │   [ 0.,  0.,  0.]]]
│ │ │ numeric_value_mask (torch.bool):
│ │ │ │ [[[ True, False,  True],
│ │ │ │   [False, False, False]],
│ │ │ │  [[False,  True, False],
│ │ │ │   [ True,  True,  True]]]
>>> print(batch.mode)
SEM
>>> print(batch.static_inclusion_mode)
omit
>>> print(batch.has_labels)
False
>>> print(batch.batch_size)
2
>>> print(batch.max_events_per_subject)
2
>>> print(batch.max_measurements_per_event)
3
>>> print(batch.max_measurements_per_subject)
None
>>> print(batch.max_static_measurements_per_subject)
None
Static Data INCLUDE/"include" Mode

In this mode, static data is included as separate keys (the presence of such keys is the indicator that the batch is in this static data inclusion mode).

>>> batch = MEDSTorchBatch(
...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.2]]),
...     event_mask=torch.tensor([[True, True], [True, False]]),
...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
...     numeric_value=torch.tensor(
...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
...     ),
...     numeric_value_mask=torch.tensor([
...         [[True, False, True], [False, False, False]],
...         [[False, True, False], [True, True, True]]
...     ]),
...     static_code=torch.tensor([[10], [9]]),
...     static_numeric_value=torch.tensor([[0.], [0.]]),
...     static_numeric_value_mask=torch.tensor([[False], [False]]),
... )
>>> print(batch)
MEDSTorchBatch:
│ Mode: Subject-Event-Measurement (SEM)
│ Static data? ✓
│ Labels? ✗
│
│ Shape:
│ │ Batch size: 2
│ │ Sequence length: 2
│ │ Event length: 3
│ │
│ │ Per-event data: (2, 2)
│ │ Per-measurement data: (2, 2, 3)
│ │ Static data: (2, 1)
│
│ Data:
│ │ Event-level:
│ │ │ time_delta_days (torch.float32):
│ │ │ │ [[1.00, 2.10],
│ │ │ │  [4.00, 0.20]]
│ │ │ event_mask (torch.bool):
│ │ │ │ [[ True,  True],
│ │ │ │  [ True, False]]
│ │
│ │ Measurement-level:
│ │ │ code (torch.int64):
│ │ │ │ [[[1, 2, 3],
│ │ │ │   [3, 0, 0]],
│ │ │ │  [[5, 6, 0],
│ │ │ │   [0, 0, 0]]]
│ │ │ numeric_value (torch.float32):
│ │ │ │ [[[ 1.,  0., -3.],
│ │ │ │   [ 0.,  0.,  0.]],
│ │ │ │  [[ 0.,  0.,  0.],
│ │ │ │   [ 0.,  0.,  0.]]]
│ │ │ numeric_value_mask (torch.bool):
│ │ │ │ [[[ True, False,  True],
│ │ │ │   [False, False, False]],
│ │ │ │  [[False,  True, False],
│ │ │ │   [ True,  True,  True]]]
│ │
│ │ Static:
│ │ │ static_code (torch.int64):
│ │ │ │ [[10],
│ │ │ │  [ 9]]
│ │ │ static_numeric_value (torch.float32):
│ │ │ │ [[0.],
│ │ │ │  [0.]]
│ │ │ static_numeric_value_mask (torch.bool):
│ │ │ │ [[False],
│ │ │ │  [False]]
>>> print(batch.mode)
SEM
>>> print(batch.static_inclusion_mode)
include
>>> print(batch.has_labels)
False
>>> print(batch.batch_size)
2
>>> print(batch.max_events_per_subject)
2
>>> print(batch.max_measurements_per_event)
3
>>> print(batch.max_measurements_per_subject)
None
>>> print(batch.max_static_measurements_per_subject)
1
Static Data PREPEND/"prepend" Mode

In this mode, static data is prepended to the beginning of the sequence of dynamic data. They will not be separated out into their own keys, and some static data specific properties will raise errors, as determining their values are not currently supported in these modes (please raise an issue if you need this functionality). This mode is indicated by the presence of the static_mask tensor in the batch. Time-deltas for static events will be 0, and the event_mask will be True.

>>> batch = MEDSTorchBatch(
...     time_delta_days=torch.tensor([[0.0, 1.0, 2.1], [0.0, 4.0, 0.2]]),
...     event_mask=torch.tensor([[True, True, True], [True, True, False]]),
...     static_mask=torch.tensor([[True, False, False], [True, False, False]]),
...     code=torch.tensor([[[10, 0, 0], [1, 2, 3], [3, 0, 0]], [[9, 0, 0], [5, 6, 0], [0, 0, 0]]]),
...     numeric_value=torch.tensor(
...         [[[0., 0., 0.], [1., 0., -3.], [0., 0., 0.]], [[0., 0., 0.], [0., 0., 0.], [0., 0., 0.]]]
...     ),
...     numeric_value_mask=torch.tensor([
...         [[False, True, False], [True, False, True], [False, False, False]],
...         [[False, True, False], [False, True, False], [True, True, True]]
...     ]),
... )
>>> print(batch)
MEDSTorchBatch:
│ Mode: Subject-Event-Measurement (SEM)
│ Static data? ✓ (prepended)
│ Labels? ✗
│
│ Shape:
│ │ Batch size: 2
│ │ Sequence length (static + dynamic): 3
│ │ Event length: 3
│ │
│ │ Per-event data: (2, 3)
│ │ Per-measurement data: (2, 3, 3)
│
│ Data:
│ │ Event-level:
│ │ │ time_delta_days (torch.float32):
│ │ │ │ [[0.00, 1.00, 2.10],
│ │ │ │  [0.00, 4.00, 0.20]]
│ │ │ event_mask (torch.bool):
│ │ │ │ [[ True,  True,  True],
│ │ │ │  [ True,  True, False]]
│ │ │ static_mask (torch.bool):
│ │ │ │ [[ True, False, False],
│ │ │ │  [ True, False, False]]
│ │
│ │ Measurement-level:
│ │ │ code (torch.int64):
│ │ │ │ [[[10,  0,  0],
│ │ │ │   [ 1,  2,  3],
│ │ │ │   [ 3,  0,  0]],
│ │ │ │  [[ 9,  0,  0],
│ │ │ │   [ 5,  6,  0],
│ │ │ │   [ 0,  0,  0]]]
│ │ │ numeric_value (torch.float32):
│ │ │ │ [[[ 0.,  0.,  0.],
│ │ │ │   [ 1.,  0., -3.],
│ │ │ │   [ 0.,  0.,  0.]],
│ │ │ │  [[ 0.,  0.,  0.],
│ │ │ │   [ 0.,  0.,  0.],
│ │ │ │   [ 0.,  0.,  0.]]]
│ │ │ numeric_value_mask (torch.bool):
│ │ │ │ [[[False,  True, False],
│ │ │ │   [ True, False,  True],
│ │ │ │   [False, False, False]],
│ │ │ │  [[False,  True, False],
│ │ │ │   [False,  True, False],
│ │ │ │   [ True,  True,  True]]]
>>> print(batch.mode)
SEM
>>> print(batch.static_inclusion_mode)
prepend
>>> print(batch.has_labels)
False
>>> print(batch.batch_size)
2
>>> print(batch.max_events_per_subject)
3
>>> print(batch.max_measurements_per_event)
3
>>> print(batch.max_measurements_per_subject)
None
>>> batch.max_static_measurements_per_subject
Traceback (most recent call last):
    ...
ValueError: This is not supported in PREPEND mode as it requires a computation
Subject-Measurement (SM) Mode

In SM mode, the batch is organized as a tensor of measurements per subject, indicated by a 2D structure of the batch’s main data elements (codes and numeric values).

Static Data OMIT/"omit" Mode

In this mode, no static data is included.

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[1, 2, 3, 3], [5, 6, 0, 0]]),
...     numeric_value=torch.tensor([[1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0]]),
...     numeric_value_mask=torch.tensor([[True, False, True, False], [False, True, False, True]]),
...     time_delta_days=torch.tensor([[1.0, 0.0, 0.0, 2.0], [4.0, 0.0, 0.0, 0.0]]),
... )
>>> print(batch)
MEDSTorchBatch:
│ Mode: Subject-Measurement (SM)
│ Static data? ✗
│ Labels? ✗
│
│ Shape:
│ │ Batch size: 2
│ │ Sequence length: 4
│ │
│ │ All dynamic data: (2, 4)
│
│ Data:
│ │ Dynamic:
│ │ │ time_delta_days (torch.float32):
│ │ │ │ [[1., 0., 0., 2.],
│ │ │ │  [4., 0., 0., 0.]]
│ │ │ code (torch.int64):
│ │ │ │ [[1, 2, 3, 3],
│ │ │ │  [5, 6, 0, 0]]
│ │ │ numeric_value (torch.float32):
│ │ │ │ [[ 1.,  0., -3.,  0.],
│ │ │ │  [ 0.,  0.,  0.,  0.]]
│ │ │ numeric_value_mask (torch.bool):
│ │ │ │ [[ True, False,  True, False],
│ │ │ │  [False,  True, False,  True]]
>>> print(batch.mode)
SM
>>> print(batch.static_inclusion_mode)
omit
>>> print(batch.has_labels)
False
>>> print(batch.batch_size)
2
>>> print(batch.max_events_per_subject)
None
>>> print(batch.max_measurements_per_event)
None
>>> print(batch.max_measurements_per_subject)
4
>>> print(batch.max_static_measurements_per_subject)
None
Static Data INCLUDE/"include" Mode

In this mode, static data is included as separate keys (the presence of such keys is the indicator that the batch is in this static data inclusion mode).

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[1, 2, 3, 3], [5, 6, 0, 0]]),
...     numeric_value=torch.tensor([[1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0]]),
...     numeric_value_mask=torch.tensor([[True, False, True, False], [False, True, False, True]]),
...     time_delta_days=torch.tensor([[1.0, 0.0, 0.0, 2.0], [4.0, 0.0, 0.0, 0.0]]),
...     static_code=torch.tensor([[10], [9]]),
...     static_numeric_value=torch.tensor([[0.], [0.]]),
...     static_numeric_value_mask=torch.tensor([[False], [False]]),
... )
>>> print(batch)
MEDSTorchBatch:
│ Mode: Subject-Measurement (SM)
│ Static data? ✓
│ Labels? ✗
│
│ Shape:
│ │ Batch size: 2
│ │ Sequence length: 4
│ │
│ │ All dynamic data: (2, 4)
│ │ Static data: (2, 1)
│
│ Data:
│ │ Dynamic:
│ │ │ time_delta_days (torch.float32):
│ │ │ │ [[1., 0., 0., 2.],
│ │ │ │  [4., 0., 0., 0.]]
│ │ │ code (torch.int64):
│ │ │ │ [[1, 2, 3, 3],
│ │ │ │  [5, 6, 0, 0]]
│ │ │ numeric_value (torch.float32):
│ │ │ │ [[ 1.,  0., -3.,  0.],
│ │ │ │  [ 0.,  0.,  0.,  0.]]
│ │ │ numeric_value_mask (torch.bool):
│ │ │ │ [[ True, False,  True, False],
│ │ │ │  [False,  True, False,  True]]
│ │
│ │ Static:
│ │ │ static_code (torch.int64):
│ │ │ │ [[10],
│ │ │ │  [ 9]]
│ │ │ static_numeric_value (torch.float32):
│ │ │ │ [[0.],
│ │ │ │  [0.]]
│ │ │ static_numeric_value_mask (torch.bool):
│ │ │ │ [[False],
│ │ │ │  [False]]
>>> print(batch.mode)
SM
>>> print(batch.static_inclusion_mode)
include
>>> print(batch.has_labels)
False
>>> print(batch.batch_size)
2
>>> print(batch.max_events_per_subject)
None
>>> print(batch.max_measurements_per_event)
None
>>> print(batch.max_measurements_per_subject)
4
>>> print(batch.max_static_measurements_per_subject)
1
Static Data PREPEND/"prepend" Mode

In this mode, static data is prepended to the beginning of the sequence of dynamic data. They will not be separated out into their own keys, and some static data specific properties will raise errors, as determining their values are not currently supported in these modes (please raise an issue if you need this functionality). This mode is indicated by the presence of the static_mask tensor in the batch. Time-deltas for static events will be 0, and the event_mask will be True.

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[10, 1, 2, 3, 3], [9, 5, 6, 0, 0]]),
...     numeric_value=torch.tensor([[0., 1., 0., -3., 0.], [0., 0., 0., 0., 0.]]),
...     numeric_value_mask=torch.tensor(
...         [[False, True, False, True, False], [False, False, True, False, True]]
...     ),
...     time_delta_days=torch.tensor([[0., 1., 0., 0., 2.], [0., 4., 0., 0., 0.]]),
...     static_mask=torch.tensor(
...         [[True, False, False, False, False], [True, False, False, False, False]]
...     ),
... )
>>> print(batch)
MEDSTorchBatch:
│ Mode: Subject-Measurement (SM)
│ Static data? ✓ (prepended)
│ Labels? ✗
│
│ Shape:
│ │ Batch size: 2
│ │ Sequence length (static + dynamic): 5
│ │
│ │ All [static; dynamic] data: (2, 5)
│
│ Data:
│ │ [Static; Dynamic]:
│ │ │ time_delta_days (torch.float32):
│ │ │ │ [[0., 1.,  ..., 0., 2.],
│ │ │ │  [0., 4.,  ..., 0., 0.]]
│ │ │ code (torch.int64):
│ │ │ │ [[10,  1,  ...,  3,  3],
│ │ │ │  [ 9,  5,  ...,  0,  0]]
│ │ │ numeric_value (torch.float32):
│ │ │ │ [[ 0.,  1.,  ..., -3.,  0.],
│ │ │ │  [ 0.,  0.,  ...,  0.,  0.]]
│ │ │ numeric_value_mask (torch.bool):
│ │ │ │ [[False,  True,  ...,  True, False],
│ │ │ │  [False, False,  ..., False,  True]]
│ │ │ static_mask (torch.bool):
│ │ │ │ [[ True, False,  ..., False, False],
│ │ │ │  [ True, False,  ..., False, False]]
>>> print(batch.mode)
SM
>>> print(batch.static_inclusion_mode)
prepend
>>> print(batch.has_labels)
False
>>> print(batch.batch_size)
2
>>> print(batch.max_events_per_subject)
None
>>> print(batch.max_measurements_per_event)
None
>>> print(batch.max_measurements_per_subject)
5
>>> batch.max_static_measurements_per_subject
Traceback (most recent call last):
    ...
ValueError: This is not supported in PREPEND mode as it requires a computation

Note that labels can also be included

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[1, 2, 3, 3], [5, 6, 0, 0]]),
...     numeric_value=torch.tensor([[1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0]]),
...     numeric_value_mask=torch.tensor([[True, False, True, False], [False, True, False, True]]),
...     time_delta_days=torch.tensor([[1.0, 0.0, 0.0, 2.0], [4.0, 0.0, 0.0, 0.0]]),
...     static_code=torch.tensor([[1], [5]]),
...     static_numeric_value=torch.tensor([[1.0], [0.0]]),
...     static_numeric_value_mask=torch.tensor([[True], [True]]),
...     boolean_value=torch.tensor([True, False]),
... )
>>> print(batch.has_labels)
True
>>> print(batch["boolean_value"])
tensor([ True, False])
>>> print(batch)
MEDSTorchBatch:
│ Mode: Subject-Measurement (SM)
│ Static data? ✓
│ Labels? ✓
│
│ Shape:
│ │ Batch size: 2
│ │ Sequence length: 4
│ │
│ │ All dynamic data: (2, 4)
│ │ Static data: (2, 1)
│ │ Labels: torch.Size([2])
│
│ Data:
│ │ Dynamic:
│ │ │ time_delta_days (torch.float32):
│ │ │ │ [[1., 0., 0., 2.],
│ │ │ │  [4., 0., 0., 0.]]
│ │ │ code (torch.int64):
│ │ │ │ [[1, 2, 3, 3],
│ │ │ │  [5, 6, 0, 0]]
│ │ │ numeric_value (torch.float32):
│ │ │ │ [[ 1.,  0., -3.,  0.],
│ │ │ │  [ 0.,  0.,  0.,  0.]]
│ │ │ numeric_value_mask (torch.bool):
│ │ │ │ [[ True, False,  True, False],
│ │ │ │  [False,  True, False,  True]]
│ │
│ │ Static:
│ │ │ static_code (torch.int64):
│ │ │ │ [[1],
│ │ │ │  [5]]
│ │ │ static_numeric_value (torch.float32):
│ │ │ │ [[1.],
│ │ │ │  [0.]]
│ │ │ static_numeric_value_mask (torch.bool):
│ │ │ │ [[True],
│ │ │ │  [True]]
│ │
│ │ Labels:
│ │ │ boolean_value (torch.bool):
│ │ │ │ [ True, False]

The batch will automatically validate tensor shapes, types, and presence vs. omission. code is the only structurally required tensor (the batch’s mode and shape are derived from it); every other dynamic field — numeric_value, numeric_value_mask, time_delta_days, event_mask — is optional and gated by MEDSTorchDataConfig.include_numeric_value, include_time_delta, and the batch mode (see issues #46 and #47):

>>> batch = MEDSTorchBatch()
Traceback (most recent call last):
    ...
ValueError: Required tensor code is missing!
>>> batch = MEDSTorchBatch(code="foobar")
Traceback (most recent call last):
    ...
TypeError: Field 'code' expected type <class 'torch.LongTensor'>, got type <class 'str'>.
>>> batch = MEDSTorchBatch(code=torch.tensor([1.]))
Traceback (most recent call last):
    ...
TypeError: Field 'code' expected type <class 'torch.LongTensor'>, got type <class 'torch.Tensor'>.

numeric_value and numeric_value_mask are treated as a pair — both must be present, or both must be omitted. Providing one without the other is rejected before any shape check:

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[1, 2, 3]]),
...     numeric_value=torch.zeros((1, 1, 3), dtype=torch.float32),
... )
Traceback (most recent call last):
    ...
ValueError: numeric_value and numeric_value_mask must both be provided or both be
None, but got numeric_value=present and numeric_value_mask=None.
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[1, 2, 3]]),
...     numeric_value_mask=torch.ones((1, 1, 3), dtype=torch.bool),
... )
Traceback (most recent call last):
    ...
ValueError: numeric_value and numeric_value_mask must both be provided or both be
None, but got numeric_value=None and numeric_value_mask=present.

In addition, the shapes of the tensors must be consistent. To begin with, the code tensor’s shape must correctly align with one of the allowed modes (SEM or SM):

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([1]),
...     numeric_value=torch.tensor([1.]),
...     numeric_value_mask=torch.tensor([True]),
...     time_delta_days=torch.tensor([1.]),
... )
Traceback (most recent call last):
    ...
ValueError: Code shape must have length either 2 (SM mode) or 3 (SEM mode); got shape torch.Size([1])!

If the code shape is in SM mode, the remaining tensors must have the correct shapes for that mode:

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[1]]),
...     numeric_value=torch.tensor([1.]),
...     numeric_value_mask=torch.tensor([True]),
...     time_delta_days=torch.tensor([1.]),
... )
Traceback (most recent call last):
    ...
ValueError: Expected shape (1, 1) for time_delta_days, but got torch.Size([1])!
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[1]]),
...     numeric_value=torch.tensor([1.]),
...     numeric_value_mask=torch.tensor([True]),
...     time_delta_days=torch.tensor([[1.]]),
... )
Traceback (most recent call last):
    ...
ValueError: Expected shape (1, 1) for numeric_value, but got torch.Size([1])!
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[1]]),
...     numeric_value=torch.tensor([[1.]]),
...     numeric_value_mask=torch.tensor([True]),
...     time_delta_days=torch.tensor([[1.]]),
... )
Traceback (most recent call last):
    ...
ValueError: Expected shape (1, 1) for numeric_value_mask, but got torch.Size([1])!
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[1]]),
...     numeric_value=torch.tensor([[1.]]),
...     numeric_value_mask=torch.tensor([[True]]),
...     time_delta_days=torch.tensor([[1.]]),
... )

You also can’t provide an event mask in SM mode:

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[1]]),
...     numeric_value=torch.tensor([[1.]]),
...     numeric_value_mask=torch.tensor([[True]]),
...     time_delta_days=torch.tensor([[1.]]),
...     event_mask=torch.tensor([[True]]),
... )
Traceback (most recent call last):
    ...
ValueError: Event mask should not be provided in SM mode!

If the code shape is in SEM mode, the remaining tensors must similarly have the correct shapes for that mode, and you must provide an event mask:

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([1.]),
...     numeric_value_mask=torch.tensor([True]),
...     time_delta_days=torch.tensor([1.]),
... )
Traceback (most recent call last):
    ...
ValueError: Event mask must be provided in SEM mode!
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([1.]),
...     numeric_value_mask=torch.tensor([True]),
...     time_delta_days=torch.tensor([1.]),
...     event_mask=torch.tensor([True]),
... )
Traceback (most recent call last):
    ...
ValueError: Expected shape (1, 2) for time_delta_days, but got torch.Size([1])!
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([1.]),
...     numeric_value_mask=torch.tensor([True]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([True]),
... )
Traceback (most recent call last):
    ...
ValueError: Expected shape (1, 2) for event_mask, but got torch.Size([1])!
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([1.]),
...     numeric_value_mask=torch.tensor([True]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([[True, True]]),
... )
Traceback (most recent call last):
    ...
ValueError: Expected shape (1, 2, 2) for numeric_value, but got torch.Size([1])!
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
...     numeric_value_mask=torch.tensor([True]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([[True, True]]),
... )
Traceback (most recent call last):
    ...
ValueError: Expected shape (1, 2, 2) for numeric_value_mask, but got torch.Size([1])!
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([[True, True]]),
... )

If you provide static data explicitly, you must provide both the static code and numeric value tensors:

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([[True, True]]),
...     static_code=torch.tensor([1, 2]),
... )
Traceback (most recent call last):
    ...
ValueError: Static numeric value and mask must both be provided if static codes are!

You can’t provide static numeric values without static codes:

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([[True, True]]),
...     static_numeric_value=torch.tensor([1., 2.]),
... )
Traceback (most recent call last):
    ...
ValueError: Static numeric value and mask should not be provided without codes!

You can’t provide both static codes/values (for include mode) and static masks (for prepend mode):

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([[True, True]]),
...     static_mask=torch.tensor([[True, True]]),
...     static_code=torch.tensor([1, 2]),
...     static_numeric_value=torch.tensor([1., 2.]),
...     static_numeric_value_mask=torch.tensor([True, True]),
... )
Traceback (most recent call last):
    ...
ValueError: Static mask should not be provided if static codes are!
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([[True, True]]),
...     static_mask=torch.tensor([[True, True]]),
...     static_numeric_value=torch.tensor([1., 2.]),
...     static_numeric_value_mask=torch.tensor([True, True]),
... )
Traceback (most recent call last):
    ...
ValueError: Static numeric value and mask should not be provided with static mask!

Static data tensors must also be provided with consistent shapes, both internally and with respect to the other tensors in that the batch size must be conserved.

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([[True, True]]),
...     static_code=torch.tensor([1, 2]),
...     static_numeric_value=torch.tensor([1.]),
...     static_numeric_value_mask=torch.tensor([True, False, True]),
... )
Traceback (most recent call last):
    ...
ValueError: Expected 2D static data tensors with a matching batch size (1), but got static_code shape
            torch.Size([2])!
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([[True, True]]),
...     static_code=torch.tensor([[1, 2]]),
...     static_numeric_value=torch.tensor([1.]),
...     static_numeric_value_mask=torch.tensor([True, False, True]),
... )
Traceback (most recent call last):
    ...
ValueError: Expected shape (1, 2) for static_numeric_value, but got torch.Size([1])!
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([[True, True]]),
...     static_code=torch.tensor([[1, 2]]),
...     static_numeric_value=torch.tensor([[1., 0.]]),
...     static_numeric_value_mask=torch.tensor([True, False, True]),
... )
Traceback (most recent call last):
    ...
ValueError: Expected shape (1, 2) for static_numeric_value_mask, but got torch.Size([3])!
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([[True, True]]),
...     static_code=torch.tensor([[1, 2]]),
...     static_numeric_value=torch.tensor([[1., 0.]]),
...     static_numeric_value_mask=torch.tensor([[True, False]]),
... )

Similarly to static data, if labels are provided, they must be of shape (batch_size,):

>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([[True, True]]),
...     boolean_value=torch.tensor([[True, False], [True, False]]),
... )
Traceback (most recent call last):
    ...
ValueError: Expected shape (1,) for boolean_value, but got torch.Size([2, 2])!
>>> batch = MEDSTorchBatch(
...     code=torch.tensor([[[1, 2], [3, 0]]]),
...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
...     time_delta_days=torch.tensor([[1., 2.]]),
...     event_mask=torch.tensor([[True, True]]),
...     boolean_value=torch.tensor([True]),
... )
Source code in meds_torchdata/types.py
 353
 354
 355
 356
 357
 358
 359
 360
 361
 362
 363
 364
 365
 366
 367
 368
 369
 370
 371
 372
 373
 374
 375
 376
 377
 378
 379
 380
 381
 382
 383
 384
 385
 386
 387
 388
 389
 390
 391
 392
 393
 394
 395
 396
 397
 398
 399
 400
 401
 402
 403
 404
 405
 406
 407
 408
 409
 410
 411
 412
 413
 414
 415
 416
 417
 418
 419
 420
 421
 422
 423
 424
 425
 426
 427
 428
 429
 430
 431
 432
 433
 434
 435
 436
 437
 438
 439
 440
 441
 442
 443
 444
 445
 446
 447
 448
 449
 450
 451
 452
 453
 454
 455
 456
 457
 458
 459
 460
 461
 462
 463
 464
 465
 466
 467
 468
 469
 470
 471
 472
 473
 474
 475
 476
 477
 478
 479
 480
 481
 482
 483
 484
 485
 486
 487
 488
 489
 490
 491
 492
 493
 494
 495
 496
 497
 498
 499
 500
 501
 502
 503
 504
 505
 506
 507
 508
 509
 510
 511
 512
 513
 514
 515
 516
 517
 518
 519
 520
 521
 522
 523
 524
 525
 526
 527
 528
 529
 530
 531
 532
 533
 534
 535
 536
 537
 538
 539
 540
 541
 542
 543
 544
 545
 546
 547
 548
 549
 550
 551
 552
 553
 554
 555
 556
 557
 558
 559
 560
 561
 562
 563
 564
 565
 566
 567
 568
 569
 570
 571
 572
 573
 574
 575
 576
 577
 578
 579
 580
 581
 582
 583
 584
 585
 586
 587
 588
 589
 590
 591
 592
 593
 594
 595
 596
 597
 598
 599
 600
 601
 602
 603
 604
 605
 606
 607
 608
 609
 610
 611
 612
 613
 614
 615
 616
 617
 618
 619
 620
 621
 622
 623
 624
 625
 626
 627
 628
 629
 630
 631
 632
 633
 634
 635
 636
 637
 638
 639
 640
 641
 642
 643
 644
 645
 646
 647
 648
 649
 650
 651
 652
 653
 654
 655
 656
 657
 658
 659
 660
 661
 662
 663
 664
 665
 666
 667
 668
 669
 670
 671
 672
 673
 674
 675
 676
 677
 678
 679
 680
 681
 682
 683
 684
 685
 686
 687
 688
 689
 690
 691
 692
 693
 694
 695
 696
 697
 698
 699
 700
 701
 702
 703
 704
 705
 706
 707
 708
 709
 710
 711
 712
 713
 714
 715
 716
 717
 718
 719
 720
 721
 722
 723
 724
 725
 726
 727
 728
 729
 730
 731
 732
 733
 734
 735
 736
 737
 738
 739
 740
 741
 742
 743
 744
 745
 746
 747
 748
 749
 750
 751
 752
 753
 754
 755
 756
 757
 758
 759
 760
 761
 762
 763
 764
 765
 766
 767
 768
 769
 770
 771
 772
 773
 774
 775
 776
 777
 778
 779
 780
 781
 782
 783
 784
 785
 786
 787
 788
 789
 790
 791
 792
 793
 794
 795
 796
 797
 798
 799
 800
 801
 802
 803
 804
 805
 806
 807
 808
 809
 810
 811
 812
 813
 814
 815
 816
 817
 818
 819
 820
 821
 822
 823
 824
 825
 826
 827
 828
 829
 830
 831
 832
 833
 834
 835
 836
 837
 838
 839
 840
 841
 842
 843
 844
 845
 846
 847
 848
 849
 850
 851
 852
 853
 854
 855
 856
 857
 858
 859
 860
 861
 862
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
@dataclass
class MEDSTorchBatch:
    """Simple data structure to hold a batch of MEDS data.

    Can be accessed by attribute (e.g., `batch.code`) or string key (e.g. `batch["code"]`). The elements in
    this tensor can take on several shapes, and keys can be present or omitted, depending on details of
    dataset configuration. To clarify these shape options, we'll define the following terms. Most of these
    terms will also be realized as properties defined on this class for accessing shape variables over the
    batch for convenience.

      - `batch_size` is the number of subjects in the batch.
      - `max_events_per_subject` is the maximum number of events (unique time-points) for any subject in the
        batch.
      - `max_measurements_per_event` is the maximum number of measurements (observed code/value pairs) for any
        event in the batch (across all subjects).
      - `max_static_measurements_per_subject` is the maximum number of static measurements observed across all
        subjects in the batch.
      - `max_any_measurements_per_event` is the maximum number of measurements that are either static for a
        given subject or observed in any event for a given subject across the batch
        (e.g., `max(max_measurements_per_event, max_static_measurements_per_subject)`).
      - `max_measurements_per_subject` is the maximum number of measurements observed across _all_ dynamic
        events for any given subject, in total, in the batch.
      - `max_any_measurements_per_subject` is the maximum number of measurements observed for any subject
        regardless of whether they are dynamic or static.

    There are a few shape "modes" that this batch can be in, depending on the configuration of the source
    dataset. These include:

      - `"SEM"`: In Subject-Event-Measurement (SEM) mode, the data is represented as a tensor of measurements
        per-event, per-subject, with missing values padded in all dimensions.
      - `"SM"`: In Subject-Measurement (SM) mode, the data is represented as a tensor of measurements
        per-subject, with events concatenated in order with neither per-event padding nor explicit separator
        tokens.

    Under each of these modes, different sets of the core attributes take on different consistent shapes.

    Under all modes:

      - Static data elements (`static_code`, `static_numeric_value`, and `static_numeric_value_mask`) are
        of shape `[batch_size, max_static_measurements_per_subject]`.
      - The label tensor, `boolean_value` tensor is of shape `[batch_size]`.

    In SEM Mode:

      - Per-event data (`time_delta_days` & `event_mask`) are of shape `[batch_size, max_events_per_subject]`
        if static data is not prepended and shape `[batch_size, max_events_per_subject + 1]` if static data is
        prepended. `time_delta_days` will have no zeros at any position save the last event per subject, for
        which position the time delta to the next event may be unknown, and, in the case where static data has
        been prepended into the sequence, the first event per subject (which will contain static data and has
        no time delta).
      - `static_mask` is of the same shape as the per-event data and will have `True` at event indices that
        correspond to the static event (currently only the first event) and `False` otherwise.
      - Per-measurement data (`code`, `numeric_value`, & `numeric_value_mask`) are of shape
        `[batch_size, max_events_per_subject, max_measurements_per_event]` if static data is not prepended and
        shape `[batch_size, max_events_per_subject + 1, max_any_measurements_per_event]` if static data is
        prepended. All measurements in the first event if static data is prepended will be static
        measurements.

    In SM Mode:

      All tensors are of shape `[batch_size, max_measurements_per_subject]` if static data is not prepended
      and `[batch_size, max_any_measurements_per_subject]` if static data is prepended.

      - `time_delta_days` will have zeros at measurement indices that correspond to either static measurements
        or measurements that do not correspond to the last measurement in an event, or at the last measurement
        in the sequence if the next time-delta is unknown.
      - `static_mask` will be of the same shape as the measurement level data and will have `True` at indices
        that correspond to static measurements and `False` otherwise.
      - `event_mask` is omitted.
      - Per-measurement data (`code`, `numeric_value`, & `numeric_value_mask`) has the same shape given above.


    Attributes:
        time_delta_days: Tensor of time deltas between sequence elements, in days.
        event_mask: Boolean tensor indicating whether a given event is present or not.
        code: Measurement code integral vocabulary indices. Equals `PAD_INDEX` when measurements are missing.
        numeric_value: Measurement numeric values. No guaranteed value for padding or missing numeric values.
        numeric_value_mask: Boolean mask indicating whether a given measurement has a numeric value. Values of
            this mask for padding measurements are undefined.
        static_mask: Boolean mask indicating whether a given measurement or event is a static
            measurement/event or a true dynamic measurement/event. Only used when static data is prepended
            into the dynamic sequence. When the batch is in SEM mode this will correspond to a mask with a
            `True` at the first event and false otherwise.
        static_code: Static measurement code integral vocabulary indices. Equals `PAD_INDEX` when measurements
            are missing.
        static_numeric_value: Static measurement numeric values. No guaranteed value for padding or missing
            numeric values.
        static_numeric_value_mask: Boolean mask indicating whether a given static measurement has a numeric
            value.
        boolean_value: Per-sample boolean labels.

    Examples:
        >>> batch = MEDSTorchBatch(
        ...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.2]]),
        ...     event_mask=torch.tensor([[True, True], [True, False]]),
        ...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
        ...     numeric_value=torch.tensor(
        ...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
        ...     ),
        ...     numeric_value_mask=torch.tensor([
        ...         [[True, False, True], [False, False, False]],
        ...         [[False, True, False], [True, True, True]] # Note the padding values may be  True or False
        ...     ]),
        ... )

    The batch is effectively merely an ordered (by the definition in the class, not order of
    specification), frozen dictionary of tensors, and can be accessed as such:

        >>> print(batch["code"])
        tensor([[[1, 2, 3],
                 [3, 0, 0]],
        <BLANKLINE>
                [[5, 6, 0],
                 [0, 0, 0]]])
        >>> print(batch["event_mask"])
        tensor([[ True,  True],
                [ True, False]])
        >>> print(list(batch.keys()))
        ['code', 'numeric_value', 'numeric_value_mask', 'time_delta_days', 'event_mask']
        >>> print(list(batch.values()))
        [tensor(...), tensor(...), tensor(...), tensor(...), tensor(...)]
        >>> print(list(batch.items()))
        [('code', tensor(...)), ('numeric_value', tensor(...)), ('numeric_value_mask', tensor(...)),
         ('time_delta_days', tensor(...)), ('event_mask', tensor(...)]
        >>> batch["code"] = torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]])
        Traceback (most recent call last):
            ...
        ValueError: MEDSTorchBatch is immutable!

    Though note that if you manually define something in a batch to be `None`, it will not be present in
    the keys/values/items:

        >>> batch = MEDSTorchBatch(
        ...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.2]]),
        ...     event_mask=torch.tensor([[True, True], [True, False]]),
        ...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
        ...     numeric_value=torch.tensor(
        ...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
        ...     ),
        ...     numeric_value_mask=torch.tensor([
        ...         [[True, False, True], [False, False, False]],
        ...         [[False, True, False], [True, True, True]]
        ...     ]),
        ...     boolean_value=None,
        ... )
        >>> print(list(batch.keys()))
        ['code', 'numeric_value', 'numeric_value_mask', 'time_delta_days', 'event_mask']

    The batch can also be accessed by attribute, and has default values for allowed fields:

        >>> print(batch.event_mask)
        tensor([[ True,  True],
                [ True, False]])
        >>> print(batch.boolean_value)
        None

    The batch has a number of properties that can be accessed for convenience:

        >>> print(batch.mode)
        SEM
        >>> print(batch.static_inclusion_mode)
        omit
        >>> print(batch.has_labels)
        False
        >>> print(batch.batch_size)
        2
        >>> print(batch.max_events_per_subject)
        2
        >>> print(batch.max_measurements_per_event)
        3
        >>> print(batch.max_measurements_per_subject)
        None
        >>> print(batch.max_static_measurements_per_subject)
        None

    Batches exist in one of several combinations of modes across the "batch mode" and the "static data
    inclusion mode". Batch mode can either be `BatchMode.SEM`/`"SEM"` or `BatchMode.SM`/`"SM"`, and static
    data inclusion mode can be `StaticInclusionMode.PREPEND`/`"prepend"`,
    `StaticInclusionMode.INCLUDE`/`"include"`, or `StaticInclusionMode.OMIT`/`"omit"`. The batch mode reflects
    the shape of the batch's elements (being either organized at an event X measurement level vs. at a
    measurement level) and the static data inclusion mode reflects how static data is included in the batch.

    > [!NOTE]
    > These modes are determined _implicitly_ by the organization of the data in the batch, not explicitly via
    > passed flags or anything.

    The batch comes with a useful print representation function that clearly indicates what modes the batch is
    in, which we can use below:

    ### Subject-Event-Measurement (SEM) Mode
    In SEM mode, the batch is organized as a tensor of measurements per event per subject, indicated by a 3D
    structure of the batch's main data elements (codes and numeric values).

    #### Static Data `OMIT`/`"omit"` Mode
    In this mode, no static data is included.

        >>> batch = MEDSTorchBatch(
        ...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.2]]),
        ...     event_mask=torch.tensor([[True, True], [True, False]]),
        ...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
        ...     numeric_value=torch.tensor(
        ...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
        ...     ),
        ...     numeric_value_mask=torch.tensor([
        ...         [[True, False, True], [False, False, False]],
        ...         [[False, True, False], [True, True, True]]
        ...     ]),
        ... )
        >>> print(batch)
        MEDSTorchBatch:
        │ Mode: Subject-Event-Measurement (SEM)
        │ Static data? ✗
        │ Labels? ✗

        │ Shape:
        │ │ Batch size: 2
        │ │ Sequence length: 2
        │ │ Event length: 3
        │ │
        │ │ Per-event data: (2, 2)
        │ │ Per-measurement data: (2, 2, 3)

        │ Data:
        │ │ Event-level:
        │ │ │ time_delta_days (torch.float32):
        │ │ │ │ [[1.00, 2.10],
        │ │ │ │  [4.00, 0.20]]
        │ │ │ event_mask (torch.bool):
        │ │ │ │ [[ True,  True],
        │ │ │ │  [ True, False]]
        │ │
        │ │ Measurement-level:
        │ │ │ code (torch.int64):
        │ │ │ │ [[[1, 2, 3],
        │ │ │ │   [3, 0, 0]],
        │ │ │ │  [[5, 6, 0],
        │ │ │ │   [0, 0, 0]]]
        │ │ │ numeric_value (torch.float32):
        │ │ │ │ [[[ 1.,  0., -3.],
        │ │ │ │   [ 0.,  0.,  0.]],
        │ │ │ │  [[ 0.,  0.,  0.],
        │ │ │ │   [ 0.,  0.,  0.]]]
        │ │ │ numeric_value_mask (torch.bool):
        │ │ │ │ [[[ True, False,  True],
        │ │ │ │   [False, False, False]],
        │ │ │ │  [[False,  True, False],
        │ │ │ │   [ True,  True,  True]]]
        >>> print(batch.mode)
        SEM
        >>> print(batch.static_inclusion_mode)
        omit
        >>> print(batch.has_labels)
        False
        >>> print(batch.batch_size)
        2
        >>> print(batch.max_events_per_subject)
        2
        >>> print(batch.max_measurements_per_event)
        3
        >>> print(batch.max_measurements_per_subject)
        None
        >>> print(batch.max_static_measurements_per_subject)
        None

    #### Static Data `INCLUDE`/`"include"` Mode
    In this mode, static data is included as separate keys (the presence of such keys is the indicator that
    the batch is in this static data inclusion mode).

        >>> batch = MEDSTorchBatch(
        ...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.2]]),
        ...     event_mask=torch.tensor([[True, True], [True, False]]),
        ...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
        ...     numeric_value=torch.tensor(
        ...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
        ...     ),
        ...     numeric_value_mask=torch.tensor([
        ...         [[True, False, True], [False, False, False]],
        ...         [[False, True, False], [True, True, True]]
        ...     ]),
        ...     static_code=torch.tensor([[10], [9]]),
        ...     static_numeric_value=torch.tensor([[0.], [0.]]),
        ...     static_numeric_value_mask=torch.tensor([[False], [False]]),
        ... )
        >>> print(batch)
        MEDSTorchBatch:
        │ Mode: Subject-Event-Measurement (SEM)
        │ Static data? ✓
        │ Labels? ✗

        │ Shape:
        │ │ Batch size: 2
        │ │ Sequence length: 2
        │ │ Event length: 3
        │ │
        │ │ Per-event data: (2, 2)
        │ │ Per-measurement data: (2, 2, 3)
        │ │ Static data: (2, 1)

        │ Data:
        │ │ Event-level:
        │ │ │ time_delta_days (torch.float32):
        │ │ │ │ [[1.00, 2.10],
        │ │ │ │  [4.00, 0.20]]
        │ │ │ event_mask (torch.bool):
        │ │ │ │ [[ True,  True],
        │ │ │ │  [ True, False]]
        │ │
        │ │ Measurement-level:
        │ │ │ code (torch.int64):
        │ │ │ │ [[[1, 2, 3],
        │ │ │ │   [3, 0, 0]],
        │ │ │ │  [[5, 6, 0],
        │ │ │ │   [0, 0, 0]]]
        │ │ │ numeric_value (torch.float32):
        │ │ │ │ [[[ 1.,  0., -3.],
        │ │ │ │   [ 0.,  0.,  0.]],
        │ │ │ │  [[ 0.,  0.,  0.],
        │ │ │ │   [ 0.,  0.,  0.]]]
        │ │ │ numeric_value_mask (torch.bool):
        │ │ │ │ [[[ True, False,  True],
        │ │ │ │   [False, False, False]],
        │ │ │ │  [[False,  True, False],
        │ │ │ │   [ True,  True,  True]]]
        │ │
        │ │ Static:
        │ │ │ static_code (torch.int64):
        │ │ │ │ [[10],
        │ │ │ │  [ 9]]
        │ │ │ static_numeric_value (torch.float32):
        │ │ │ │ [[0.],
        │ │ │ │  [0.]]
        │ │ │ static_numeric_value_mask (torch.bool):
        │ │ │ │ [[False],
        │ │ │ │  [False]]
        >>> print(batch.mode)
        SEM
        >>> print(batch.static_inclusion_mode)
        include
        >>> print(batch.has_labels)
        False
        >>> print(batch.batch_size)
        2
        >>> print(batch.max_events_per_subject)
        2
        >>> print(batch.max_measurements_per_event)
        3
        >>> print(batch.max_measurements_per_subject)
        None
        >>> print(batch.max_static_measurements_per_subject)
        1

    #### Static Data `PREPEND`/`"prepend"` Mode
    In this mode, static data is prepended to the beginning of the sequence of dynamic data. They will not be
    separated out into their own keys, and some static data specific properties will raise errors, as
    determining their values are not currently supported in these modes (please raise an issue if you need
    this functionality). This mode is indicated by the presence of the `static_mask` tensor in the batch.
    Time-deltas for static events will be 0, and the `event_mask` will be `True`.

        >>> batch = MEDSTorchBatch(
        ...     time_delta_days=torch.tensor([[0.0, 1.0, 2.1], [0.0, 4.0, 0.2]]),
        ...     event_mask=torch.tensor([[True, True, True], [True, True, False]]),
        ...     static_mask=torch.tensor([[True, False, False], [True, False, False]]),
        ...     code=torch.tensor([[[10, 0, 0], [1, 2, 3], [3, 0, 0]], [[9, 0, 0], [5, 6, 0], [0, 0, 0]]]),
        ...     numeric_value=torch.tensor(
        ...         [[[0., 0., 0.], [1., 0., -3.], [0., 0., 0.]], [[0., 0., 0.], [0., 0., 0.], [0., 0., 0.]]]
        ...     ),
        ...     numeric_value_mask=torch.tensor([
        ...         [[False, True, False], [True, False, True], [False, False, False]],
        ...         [[False, True, False], [False, True, False], [True, True, True]]
        ...     ]),
        ... )
        >>> print(batch)
        MEDSTorchBatch:
        │ Mode: Subject-Event-Measurement (SEM)
        │ Static data? ✓ (prepended)
        │ Labels? ✗

        │ Shape:
        │ │ Batch size: 2
        │ │ Sequence length (static + dynamic): 3
        │ │ Event length: 3
        │ │
        │ │ Per-event data: (2, 3)
        │ │ Per-measurement data: (2, 3, 3)

        │ Data:
        │ │ Event-level:
        │ │ │ time_delta_days (torch.float32):
        │ │ │ │ [[0.00, 1.00, 2.10],
        │ │ │ │  [0.00, 4.00, 0.20]]
        │ │ │ event_mask (torch.bool):
        │ │ │ │ [[ True,  True,  True],
        │ │ │ │  [ True,  True, False]]
        │ │ │ static_mask (torch.bool):
        │ │ │ │ [[ True, False, False],
        │ │ │ │  [ True, False, False]]
        │ │
        │ │ Measurement-level:
        │ │ │ code (torch.int64):
        │ │ │ │ [[[10,  0,  0],
        │ │ │ │   [ 1,  2,  3],
        │ │ │ │   [ 3,  0,  0]],
        │ │ │ │  [[ 9,  0,  0],
        │ │ │ │   [ 5,  6,  0],
        │ │ │ │   [ 0,  0,  0]]]
        │ │ │ numeric_value (torch.float32):
        │ │ │ │ [[[ 0.,  0.,  0.],
        │ │ │ │   [ 1.,  0., -3.],
        │ │ │ │   [ 0.,  0.,  0.]],
        │ │ │ │  [[ 0.,  0.,  0.],
        │ │ │ │   [ 0.,  0.,  0.],
        │ │ │ │   [ 0.,  0.,  0.]]]
        │ │ │ numeric_value_mask (torch.bool):
        │ │ │ │ [[[False,  True, False],
        │ │ │ │   [ True, False,  True],
        │ │ │ │   [False, False, False]],
        │ │ │ │  [[False,  True, False],
        │ │ │ │   [False,  True, False],
        │ │ │ │   [ True,  True,  True]]]
        >>> print(batch.mode)
        SEM
        >>> print(batch.static_inclusion_mode)
        prepend
        >>> print(batch.has_labels)
        False
        >>> print(batch.batch_size)
        2
        >>> print(batch.max_events_per_subject)
        3
        >>> print(batch.max_measurements_per_event)
        3
        >>> print(batch.max_measurements_per_subject)
        None
        >>> batch.max_static_measurements_per_subject
        Traceback (most recent call last):
            ...
        ValueError: This is not supported in PREPEND mode as it requires a computation

    ### Subject-Measurement (SM) Mode
    In SM mode, the batch is organized as a tensor of measurements per subject, indicated by a 2D
    structure of the batch's main data elements (codes and numeric values).

    #### Static Data `OMIT`/`"omit"` Mode
    In this mode, no static data is included.

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[1, 2, 3, 3], [5, 6, 0, 0]]),
        ...     numeric_value=torch.tensor([[1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0]]),
        ...     numeric_value_mask=torch.tensor([[True, False, True, False], [False, True, False, True]]),
        ...     time_delta_days=torch.tensor([[1.0, 0.0, 0.0, 2.0], [4.0, 0.0, 0.0, 0.0]]),
        ... )
        >>> print(batch)
        MEDSTorchBatch:
        │ Mode: Subject-Measurement (SM)
        │ Static data? ✗
        │ Labels? ✗

        │ Shape:
        │ │ Batch size: 2
        │ │ Sequence length: 4
        │ │
        │ │ All dynamic data: (2, 4)

        │ Data:
        │ │ Dynamic:
        │ │ │ time_delta_days (torch.float32):
        │ │ │ │ [[1., 0., 0., 2.],
        │ │ │ │  [4., 0., 0., 0.]]
        │ │ │ code (torch.int64):
        │ │ │ │ [[1, 2, 3, 3],
        │ │ │ │  [5, 6, 0, 0]]
        │ │ │ numeric_value (torch.float32):
        │ │ │ │ [[ 1.,  0., -3.,  0.],
        │ │ │ │  [ 0.,  0.,  0.,  0.]]
        │ │ │ numeric_value_mask (torch.bool):
        │ │ │ │ [[ True, False,  True, False],
        │ │ │ │  [False,  True, False,  True]]
        >>> print(batch.mode)
        SM
        >>> print(batch.static_inclusion_mode)
        omit
        >>> print(batch.has_labels)
        False
        >>> print(batch.batch_size)
        2
        >>> print(batch.max_events_per_subject)
        None
        >>> print(batch.max_measurements_per_event)
        None
        >>> print(batch.max_measurements_per_subject)
        4
        >>> print(batch.max_static_measurements_per_subject)
        None

    #### Static Data `INCLUDE`/`"include"` Mode
    In this mode, static data is included as separate keys (the presence of such keys is the indicator that
    the batch is in this static data inclusion mode).

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[1, 2, 3, 3], [5, 6, 0, 0]]),
        ...     numeric_value=torch.tensor([[1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0]]),
        ...     numeric_value_mask=torch.tensor([[True, False, True, False], [False, True, False, True]]),
        ...     time_delta_days=torch.tensor([[1.0, 0.0, 0.0, 2.0], [4.0, 0.0, 0.0, 0.0]]),
        ...     static_code=torch.tensor([[10], [9]]),
        ...     static_numeric_value=torch.tensor([[0.], [0.]]),
        ...     static_numeric_value_mask=torch.tensor([[False], [False]]),
        ... )
        >>> print(batch)
        MEDSTorchBatch:
        │ Mode: Subject-Measurement (SM)
        │ Static data? ✓
        │ Labels? ✗

        │ Shape:
        │ │ Batch size: 2
        │ │ Sequence length: 4
        │ │
        │ │ All dynamic data: (2, 4)
        │ │ Static data: (2, 1)

        │ Data:
        │ │ Dynamic:
        │ │ │ time_delta_days (torch.float32):
        │ │ │ │ [[1., 0., 0., 2.],
        │ │ │ │  [4., 0., 0., 0.]]
        │ │ │ code (torch.int64):
        │ │ │ │ [[1, 2, 3, 3],
        │ │ │ │  [5, 6, 0, 0]]
        │ │ │ numeric_value (torch.float32):
        │ │ │ │ [[ 1.,  0., -3.,  0.],
        │ │ │ │  [ 0.,  0.,  0.,  0.]]
        │ │ │ numeric_value_mask (torch.bool):
        │ │ │ │ [[ True, False,  True, False],
        │ │ │ │  [False,  True, False,  True]]
        │ │
        │ │ Static:
        │ │ │ static_code (torch.int64):
        │ │ │ │ [[10],
        │ │ │ │  [ 9]]
        │ │ │ static_numeric_value (torch.float32):
        │ │ │ │ [[0.],
        │ │ │ │  [0.]]
        │ │ │ static_numeric_value_mask (torch.bool):
        │ │ │ │ [[False],
        │ │ │ │  [False]]
        >>> print(batch.mode)
        SM
        >>> print(batch.static_inclusion_mode)
        include
        >>> print(batch.has_labels)
        False
        >>> print(batch.batch_size)
        2
        >>> print(batch.max_events_per_subject)
        None
        >>> print(batch.max_measurements_per_event)
        None
        >>> print(batch.max_measurements_per_subject)
        4
        >>> print(batch.max_static_measurements_per_subject)
        1

    #### Static Data `PREPEND`/`"prepend"` Mode
    In this mode, static data is prepended to the beginning of the sequence of dynamic data. They will not be
    separated out into their own keys, and some static data specific properties will raise errors, as
    determining their values are not currently supported in these modes (please raise an issue if you need
    this functionality). This mode is indicated by the presence of the `static_mask` tensor in the batch.
    Time-deltas for static events will be 0, and the `event_mask` will be `True`.

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[10, 1, 2, 3, 3], [9, 5, 6, 0, 0]]),
        ...     numeric_value=torch.tensor([[0., 1., 0., -3., 0.], [0., 0., 0., 0., 0.]]),
        ...     numeric_value_mask=torch.tensor(
        ...         [[False, True, False, True, False], [False, False, True, False, True]]
        ...     ),
        ...     time_delta_days=torch.tensor([[0., 1., 0., 0., 2.], [0., 4., 0., 0., 0.]]),
        ...     static_mask=torch.tensor(
        ...         [[True, False, False, False, False], [True, False, False, False, False]]
        ...     ),
        ... )
        >>> print(batch)
        MEDSTorchBatch:
        │ Mode: Subject-Measurement (SM)
        │ Static data? ✓ (prepended)
        │ Labels? ✗

        │ Shape:
        │ │ Batch size: 2
        │ │ Sequence length (static + dynamic): 5
        │ │
        │ │ All [static; dynamic] data: (2, 5)

        │ Data:
        │ │ [Static; Dynamic]:
        │ │ │ time_delta_days (torch.float32):
        │ │ │ │ [[0., 1.,  ..., 0., 2.],
        │ │ │ │  [0., 4.,  ..., 0., 0.]]
        │ │ │ code (torch.int64):
        │ │ │ │ [[10,  1,  ...,  3,  3],
        │ │ │ │  [ 9,  5,  ...,  0,  0]]
        │ │ │ numeric_value (torch.float32):
        │ │ │ │ [[ 0.,  1.,  ..., -3.,  0.],
        │ │ │ │  [ 0.,  0.,  ...,  0.,  0.]]
        │ │ │ numeric_value_mask (torch.bool):
        │ │ │ │ [[False,  True,  ...,  True, False],
        │ │ │ │  [False, False,  ..., False,  True]]
        │ │ │ static_mask (torch.bool):
        │ │ │ │ [[ True, False,  ..., False, False],
        │ │ │ │  [ True, False,  ..., False, False]]
        >>> print(batch.mode)
        SM
        >>> print(batch.static_inclusion_mode)
        prepend
        >>> print(batch.has_labels)
        False
        >>> print(batch.batch_size)
        2
        >>> print(batch.max_events_per_subject)
        None
        >>> print(batch.max_measurements_per_event)
        None
        >>> print(batch.max_measurements_per_subject)
        5
        >>> batch.max_static_measurements_per_subject
        Traceback (most recent call last):
            ...
        ValueError: This is not supported in PREPEND mode as it requires a computation


    Note that labels can also be included

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[1, 2, 3, 3], [5, 6, 0, 0]]),
        ...     numeric_value=torch.tensor([[1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0]]),
        ...     numeric_value_mask=torch.tensor([[True, False, True, False], [False, True, False, True]]),
        ...     time_delta_days=torch.tensor([[1.0, 0.0, 0.0, 2.0], [4.0, 0.0, 0.0, 0.0]]),
        ...     static_code=torch.tensor([[1], [5]]),
        ...     static_numeric_value=torch.tensor([[1.0], [0.0]]),
        ...     static_numeric_value_mask=torch.tensor([[True], [True]]),
        ...     boolean_value=torch.tensor([True, False]),
        ... )
        >>> print(batch.has_labels)
        True
        >>> print(batch["boolean_value"])
        tensor([ True, False])
        >>> print(batch)
        MEDSTorchBatch:
        │ Mode: Subject-Measurement (SM)
        │ Static data? ✓
        │ Labels? ✓

        │ Shape:
        │ │ Batch size: 2
        │ │ Sequence length: 4
        │ │
        │ │ All dynamic data: (2, 4)
        │ │ Static data: (2, 1)
        │ │ Labels: torch.Size([2])

        │ Data:
        │ │ Dynamic:
        │ │ │ time_delta_days (torch.float32):
        │ │ │ │ [[1., 0., 0., 2.],
        │ │ │ │  [4., 0., 0., 0.]]
        │ │ │ code (torch.int64):
        │ │ │ │ [[1, 2, 3, 3],
        │ │ │ │  [5, 6, 0, 0]]
        │ │ │ numeric_value (torch.float32):
        │ │ │ │ [[ 1.,  0., -3.,  0.],
        │ │ │ │  [ 0.,  0.,  0.,  0.]]
        │ │ │ numeric_value_mask (torch.bool):
        │ │ │ │ [[ True, False,  True, False],
        │ │ │ │  [False,  True, False,  True]]
        │ │
        │ │ Static:
        │ │ │ static_code (torch.int64):
        │ │ │ │ [[1],
        │ │ │ │  [5]]
        │ │ │ static_numeric_value (torch.float32):
        │ │ │ │ [[1.],
        │ │ │ │  [0.]]
        │ │ │ static_numeric_value_mask (torch.bool):
        │ │ │ │ [[True],
        │ │ │ │  [True]]
        │ │
        │ │ Labels:
        │ │ │ boolean_value (torch.bool):
        │ │ │ │ [ True, False]

    The batch will automatically validate tensor shapes, types, and presence vs. omission. `code` is the
    only structurally required tensor (the batch's mode and shape are derived from it); every other dynamic
    field — `numeric_value`, `numeric_value_mask`, `time_delta_days`, `event_mask` — is optional and
    gated by `MEDSTorchDataConfig.include_numeric_value`, `include_time_delta`, and the batch mode (see
    issues #46 and #47):

        >>> batch = MEDSTorchBatch()
        Traceback (most recent call last):
            ...
        ValueError: Required tensor code is missing!
        >>> batch = MEDSTorchBatch(code="foobar")
        Traceback (most recent call last):
            ...
        TypeError: Field 'code' expected type <class 'torch.LongTensor'>, got type <class 'str'>.
        >>> batch = MEDSTorchBatch(code=torch.tensor([1.]))
        Traceback (most recent call last):
            ...
        TypeError: Field 'code' expected type <class 'torch.LongTensor'>, got type <class 'torch.Tensor'>.

    `numeric_value` and `numeric_value_mask` are treated as a pair — both must be present, or
    both must be omitted. Providing one without the other is rejected before any shape check:

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[1, 2, 3]]),
        ...     numeric_value=torch.zeros((1, 1, 3), dtype=torch.float32),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: numeric_value and numeric_value_mask must both be provided or both be
        None, but got numeric_value=present and numeric_value_mask=None.
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[1, 2, 3]]),
        ...     numeric_value_mask=torch.ones((1, 1, 3), dtype=torch.bool),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: numeric_value and numeric_value_mask must both be provided or both be
        None, but got numeric_value=None and numeric_value_mask=present.

    In addition, the shapes of the tensors must be consistent. To begin with, the code tensor's shape must
    correctly align with one of the allowed modes (SEM or SM):

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([1]),
        ...     numeric_value=torch.tensor([1.]),
        ...     numeric_value_mask=torch.tensor([True]),
        ...     time_delta_days=torch.tensor([1.]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Code shape must have length either 2 (SM mode) or 3 (SEM mode); got shape torch.Size([1])!

    If the code shape is in SM mode, the remaining tensors must have the correct shapes for that mode:

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[1]]),
        ...     numeric_value=torch.tensor([1.]),
        ...     numeric_value_mask=torch.tensor([True]),
        ...     time_delta_days=torch.tensor([1.]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Expected shape (1, 1) for time_delta_days, but got torch.Size([1])!
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[1]]),
        ...     numeric_value=torch.tensor([1.]),
        ...     numeric_value_mask=torch.tensor([True]),
        ...     time_delta_days=torch.tensor([[1.]]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Expected shape (1, 1) for numeric_value, but got torch.Size([1])!
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[1]]),
        ...     numeric_value=torch.tensor([[1.]]),
        ...     numeric_value_mask=torch.tensor([True]),
        ...     time_delta_days=torch.tensor([[1.]]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Expected shape (1, 1) for numeric_value_mask, but got torch.Size([1])!
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[1]]),
        ...     numeric_value=torch.tensor([[1.]]),
        ...     numeric_value_mask=torch.tensor([[True]]),
        ...     time_delta_days=torch.tensor([[1.]]),
        ... )

    You also can't provide an event mask in SM mode:

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[1]]),
        ...     numeric_value=torch.tensor([[1.]]),
        ...     numeric_value_mask=torch.tensor([[True]]),
        ...     time_delta_days=torch.tensor([[1.]]),
        ...     event_mask=torch.tensor([[True]]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Event mask should not be provided in SM mode!

    If the code shape is in SEM mode, the remaining tensors must similarly have the correct shapes for
    that mode, and you _must_ provide an event mask:

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([1.]),
        ...     numeric_value_mask=torch.tensor([True]),
        ...     time_delta_days=torch.tensor([1.]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Event mask must be provided in SEM mode!
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([1.]),
        ...     numeric_value_mask=torch.tensor([True]),
        ...     time_delta_days=torch.tensor([1.]),
        ...     event_mask=torch.tensor([True]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Expected shape (1, 2) for time_delta_days, but got torch.Size([1])!
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([1.]),
        ...     numeric_value_mask=torch.tensor([True]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([True]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Expected shape (1, 2) for event_mask, but got torch.Size([1])!
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([1.]),
        ...     numeric_value_mask=torch.tensor([True]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([[True, True]]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Expected shape (1, 2, 2) for numeric_value, but got torch.Size([1])!
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
        ...     numeric_value_mask=torch.tensor([True]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([[True, True]]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Expected shape (1, 2, 2) for numeric_value_mask, but got torch.Size([1])!
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
        ...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([[True, True]]),
        ... )

    If you provide static data explicitly, you must provide both the static code and numeric value tensors:

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
        ...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([[True, True]]),
        ...     static_code=torch.tensor([1, 2]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Static numeric value and mask must both be provided if static codes are!

    You can't provide static numeric values without static codes:

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
        ...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([[True, True]]),
        ...     static_numeric_value=torch.tensor([1., 2.]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Static numeric value and mask should not be provided without codes!

    You can't provide both static codes/values (for include mode) and static masks (for prepend mode):

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
        ...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([[True, True]]),
        ...     static_mask=torch.tensor([[True, True]]),
        ...     static_code=torch.tensor([1, 2]),
        ...     static_numeric_value=torch.tensor([1., 2.]),
        ...     static_numeric_value_mask=torch.tensor([True, True]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Static mask should not be provided if static codes are!
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
        ...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([[True, True]]),
        ...     static_mask=torch.tensor([[True, True]]),
        ...     static_numeric_value=torch.tensor([1., 2.]),
        ...     static_numeric_value_mask=torch.tensor([True, True]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Static numeric value and mask should not be provided with static mask!

    Static data tensors must also be provided with consistent shapes, both internally and with respect to
    the other tensors in that the batch size must be conserved.

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
        ...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([[True, True]]),
        ...     static_code=torch.tensor([1, 2]),
        ...     static_numeric_value=torch.tensor([1.]),
        ...     static_numeric_value_mask=torch.tensor([True, False, True]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Expected 2D static data tensors with a matching batch size (1), but got static_code shape
                    torch.Size([2])!
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
        ...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([[True, True]]),
        ...     static_code=torch.tensor([[1, 2]]),
        ...     static_numeric_value=torch.tensor([1.]),
        ...     static_numeric_value_mask=torch.tensor([True, False, True]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Expected shape (1, 2) for static_numeric_value, but got torch.Size([1])!
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
        ...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([[True, True]]),
        ...     static_code=torch.tensor([[1, 2]]),
        ...     static_numeric_value=torch.tensor([[1., 0.]]),
        ...     static_numeric_value_mask=torch.tensor([True, False, True]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Expected shape (1, 2) for static_numeric_value_mask, but got torch.Size([3])!
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
        ...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([[True, True]]),
        ...     static_code=torch.tensor([[1, 2]]),
        ...     static_numeric_value=torch.tensor([[1., 0.]]),
        ...     static_numeric_value_mask=torch.tensor([[True, False]]),
        ... )

    Similarly to static data, if labels are provided, they must be of shape (batch_size,):

        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
        ...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([[True, True]]),
        ...     boolean_value=torch.tensor([[True, False], [True, False]]),
        ... )
        Traceback (most recent call last):
            ...
        ValueError: Expected shape (1,) for boolean_value, but got torch.Size([2, 2])!
        >>> batch = MEDSTorchBatch(
        ...     code=torch.tensor([[[1, 2], [3, 0]]]),
        ...     numeric_value=torch.tensor([[[1., 0.], [0., 0.]]]),
        ...     numeric_value_mask=torch.tensor([[[True, False], [False, False]]]),
        ...     time_delta_days=torch.tensor([[1., 2.]]),
        ...     event_mask=torch.tensor([[True, True]]),
        ...     boolean_value=torch.tensor([True]),
        ... )
    """

    PAD_INDEX: ClassVar[int] = 0
    # Only `code` is structurally required — every batch must have at least one code
    # tensor to define the shape and mode of the batch. The other dynamic fields are
    # optional and gated by `MEDSTorchDataConfig.include_numeric_value` (controlling
    # `numeric_value` + `numeric_value_mask`) and `include_time_delta` (controlling
    # `time_delta_days`); callers may also omit `event_mask` in SM mode. See issues
    # 46 and 47.
    _REQ_TENSORS: ClassVar[list[str]] = ["code"]

    # Core dynamic data elements (measurement-level):
    code: torch.LongTensor | None = None
    numeric_value: torch.FloatTensor | None = None
    numeric_value_mask: torch.BoolTensor | None = None

    # Temporal information (event-level):
    time_delta_days: torch.FloatTensor | None = None
    event_mask: torch.BoolTensor | None = None

    # Static vs. dynamic differentiation (for prepending static data):
    static_mask: torch.BoolTensor | None = None

    # Static data elements (subject-level):
    static_code: torch.LongTensor | None = None
    static_numeric_value: torch.FloatTensor | None = None
    static_numeric_value_mask: torch.BoolTensor | None = None

    # Task label data elements (subject-level):
    boolean_value: torch.BoolTensor | None = None

    # Optional oversampling weight for step-through sampling: for each sample, the number of
    # dataset elements the originating subject expands into. Intended for per-sample loss
    # reweighting (e.g. `loss_weight = 1 / n_subject_windows`). Shape: `[batch_size]`.
    n_subject_windows: torch.LongTensor | None = None

    STATIC_TENSOR_NAMES: ClassVar[tuple[str]] = (
        "static_code",
        "static_numeric_value",
        "static_numeric_value_mask",
    )
    SE_TENSOR_NAMES: ClassVar[tuple[str]] = ("time_delta_days", "event_mask", "static_mask")
    SM_TENSOR_NAMES: ClassVar[tuple[str]] = (
        "time_delta_days",
        "code",
        "numeric_value",
        "numeric_value_mask",
        "static_mask",
    )
    SEM_TENSOR_NAMES: ClassVar[tuple[str]] = ("code", "numeric_value", "numeric_value_mask")
    LABEL_TENSOR_NAMES: ClassVar[tuple[str]] = ("boolean_value",)

    def __check_shape(self, name: str, shape: tuple[int, ...]) -> None:
        """Check that the shape of a tensor matches the expected shape, or raise an appropriate error."""
        got_shape = getattr(self, name).shape
        if got_shape != shape:
            raise ValueError(f"Expected shape {shape} for {name}, but got {got_shape}!")

    def __post_init__(self):
        """Check that the batch is well-formed, raising an error if it is not."""
        for field in fields(self):
            tensor_type = get_args(field.type)[0]
            match value := getattr(self, field.name):
                case None:
                    if field.name in self._REQ_TENSORS:
                        raise ValueError(f"Required tensor {field.name} is missing!")
                    else:
                        pass
                case tensor_type():
                    pass
                case _:
                    raise TypeError(
                        f"Field '{field.name}' expected type {tensor_type}, got type {type(value)}."
                    )

        # numeric_value and numeric_value_mask must be provided (or omitted) as a pair.
        if (self.numeric_value is None) != (self.numeric_value_mask is None):
            raise ValueError(
                "numeric_value and numeric_value_mask must both be provided or both be None, "
                f"but got numeric_value={'present' if self.numeric_value is not None else 'None'} "
                f"and numeric_value_mask={'present' if self.numeric_value_mask is not None else 'None'}."
            )

        # Dynamic-field shape checks. `time_delta_days`, `numeric_value`, and
        # `numeric_value_mask` are all optional now (see `_REQ_TENSORS` — gated by
        # `MEDSTorchDataConfig.include_time_delta` and `include_numeric_value`), so we only
        # check the shape of the ones the caller actually provided. `code` is always present
        # and `event_mask` is required in SEM mode because it defines the per-event shape.
        match self.mode:
            case BatchMode.SEM:
                if self.event_mask is None:
                    raise ValueError(f"Event mask must be provided in {self.mode} mode!")
                if self.time_delta_days is not None:
                    self.__check_shape("time_delta_days", self._SE_shape)
                self.__check_shape("event_mask", self._SE_shape)
                if self.numeric_value is not None:
                    self.__check_shape("numeric_value", self._SEM_shape)
                if self.numeric_value_mask is not None:
                    self.__check_shape("numeric_value_mask", self._SEM_shape)
            case BatchMode.SM:
                if self.event_mask is not None:
                    raise ValueError(f"Event mask should not be provided in {self.mode} mode!")
                if self.time_delta_days is not None:
                    self.__check_shape("time_delta_days", self._SM_shape)
                if self.numeric_value is not None:
                    self.__check_shape("numeric_value", self._SM_shape)
                if self.numeric_value_mask is not None:
                    self.__check_shape("numeric_value_mask", self._SM_shape)
            case _:  # pragma: no cover
                raise ValueError(f"Invalid mode {self.mode}!")

        match self.static_inclusion_mode:
            case StaticInclusionMode.INCLUDE:
                if self.static_mask is not None:
                    raise ValueError("Static mask should not be provided if static codes are!")
                if self.static_numeric_value is None or self.static_numeric_value_mask is None:
                    raise ValueError(
                        "Static numeric value and mask must both be provided if static codes are!"
                    )
                if len(self.static_code.shape) != 2 or self.static_code.shape[0] != self.batch_size:
                    raise ValueError(
                        f"Expected 2D static data tensors with a matching batch size ({self.batch_size}), "
                        f"but got static_code shape {self.static_code.shape}!"
                    )
                self.__check_shape("static_numeric_value", self._static_shape)
                self.__check_shape("static_numeric_value_mask", self._static_shape)
            case StaticInclusionMode.OMIT:
                if self.static_numeric_value is not None or self.static_numeric_value_mask is not None:
                    raise ValueError("Static numeric value and mask should not be provided without codes!")
            case StaticInclusionMode.PREPEND:
                if self.static_numeric_value is not None or self.static_numeric_value_mask is not None:
                    raise ValueError("Static numeric value and mask should not be provided with static mask!")
                if self.mode == BatchMode.SEM:
                    self.__check_shape("static_mask", self._SE_shape)
                elif self.mode == BatchMode.SM:
                    self.__check_shape("static_mask", self._SM_shape)
            case _:  # pragma: no cover
                raise ValueError(f"Invalid static inclusion mode {self.static_inclusion_mode}!")

        if self.has_labels:
            self.__check_shape("boolean_value", (self.batch_size,))

        if self.n_subject_windows is not None:
            self.__check_shape("n_subject_windows", (self.batch_size,))

    # Here we define some operators to make this behave like a dictionary:
    def __getitem__(self, key: str) -> torch.Tensor:
        """Get a tensor from the batch by key."""
        return getattr(self, key)

    def __setitem__(self, key: str, value: torch.Tensor) -> None:
        """Set a tensor in the batch by key.

        Only valid if the key is a valid field.
        """
        raise ValueError("MEDSTorchBatch is immutable!")

    def keys(self) -> Generator[str, None, None]:
        """Get the keys of the batch."""
        for field in fields(self):
            if getattr(self, field.name) is not None:
                yield field.name

    def values(self) -> Generator[torch.Tensor, None, None]:
        """Get the values of the batch."""
        for key in self.keys():
            yield self[key]

    def items(self) -> Generator[tuple[str, torch.Tensor], None, None]:
        """Get the items of the batch."""
        yield from zip(self.keys(), self.values(), strict=True)

    @property
    def mode(self) -> BatchMode:
        """The mode of the batch, reflecting the internal organization of subject measurements."""
        match len(self.code.shape):
            case 2:
                return BatchMode.SM
            case 3:
                return BatchMode.SEM
            case _:
                raise ValueError(
                    "Code shape must have length either 2 (SM mode) or 3 (SEM mode); "
                    f"got shape {self.code.shape}!"
                )

    @property
    def static_inclusion_mode(self) -> StaticInclusionMode:
        if self.static_code is not None:
            return StaticInclusionMode.INCLUDE
        elif self.static_mask is not None:
            return StaticInclusionMode.PREPEND
        else:
            return StaticInclusionMode.OMIT

    @property
    def has_labels(self) -> bool:
        """Whether the batch has labels."""
        return self.boolean_value is not None

    @property
    def batch_size(self) -> int:
        """The number of subjects in the batch."""
        return self.code.shape[0]

    @property
    def max_events_per_subject(self) -> int | None:
        """The maximum number of events for any subject in the batch.

        Only valid in SEM mode.
        """
        return self.code.shape[1] if self.mode is BatchMode.SEM else None

    @property
    def max_measurements_per_event(self) -> int | None:
        """The maximum number of measurements for any event in the batch.

        Only valid in SEM mode.
        """
        return self.code.shape[2] if self.mode is BatchMode.SEM else None

    @property
    def max_measurements_per_subject(self) -> int | None:
        """The maximum number of measurements for any subject in the batch.

        Only valid in SM mode.
        """
        return self.code.shape[1] if self.mode is BatchMode.SM else None

    @property
    def max_static_measurements_per_subject(self) -> int | None:
        """The maximum number of static measurements for any subject in the batch."""
        match self.static_inclusion_mode:
            case StaticInclusionMode.INCLUDE:
                return self.static_code.shape[1]
            case StaticInclusionMode.PREPEND:
                raise ValueError("This is not supported in PREPEND mode as it requires a computation")
            case StaticInclusionMode.OMIT:
                return None

    @property
    def _SE_shape(self) -> tuple[int, int]:
        """Returns the subject-event shape of the batch. Only valid in SEM mode.

        Examples:
            >>> batch = MEDSTorchBatch(
            ...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.0]]),
            ...     event_mask=torch.tensor([[True, True], [True, False]]),
            ...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
            ...     numeric_value=torch.tensor(
            ...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
            ...     ),
            ...     numeric_value_mask=torch.tensor([
            ...         [[True, False, True], [False, False, False]],
            ...         [[False, True, False], [True, True, True]]
            ...     ]), # Note the padding values may be  True or False
            ... )
            >>> print(batch._SE_shape)
            (2, 2)
        """
        return (self.batch_size, self.max_events_per_subject)

    @property
    def _SEM_shape(self) -> tuple[int, int, int]:
        """Returns the subject-event-measurement shape of the batch. Only valid in SEM mode.

        Examples:
            >>> batch = MEDSTorchBatch(
            ...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.0]]),
            ...     event_mask=torch.tensor([[True, True], [True, False]]),
            ...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
            ...     numeric_value=torch.tensor(
            ...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
            ...     ),
            ...     numeric_value_mask=torch.tensor([
            ...         [[True, False, True], [False, False, False]],
            ...         [[False, True, False], [True, True, True]]
            ...     ]), # Note the padding values may be  True or False
            ... )
            >>> print(batch._SEM_shape)
            (2, 2, 3)
        """
        return (
            self.batch_size,
            self.max_events_per_subject,
            self.max_measurements_per_event,
        )

    @property
    def _SM_shape(self) -> tuple[int, int]:
        """Returns the subject-measurement shape of the batch. Only valid in SM mode.

        Examples:
            >>> batch = MEDSTorchBatch(
            ...     time_delta_days=torch.tensor([[1.0, 0.0, 0.0, 2.1], [4.0, 0.0, 0.0, 0.0]]),
            ...     code=torch.tensor([[1, 2, 3, 3], [5, 6, 0, 0]]),
            ...     numeric_value=torch.tensor([[1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0]]),
            ...     numeric_value_mask=torch.tensor([[True, False, True, False], [False, True, False, True]]),
            ... )
            >>> print(batch._SM_shape)
            (2, 4)
        """
        return (self.batch_size, self.max_measurements_per_subject)

    @property
    def _static_shape(self) -> tuple[int, int]:
        """Returns the static data shape of the batch. Only valid if the batch has static data.

        Examples:
            >>> batch = MEDSTorchBatch(
            ...     time_delta_days=torch.tensor([[1.0, 0.0, 0.0, 2.1], [4.0, 0.0, 0.0, 0.0]]),
            ...     code=torch.tensor([[1, 2, 3, 3], [5, 6, 0, 0]]),
            ...     numeric_value=torch.tensor([[1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0]]),
            ...     numeric_value_mask=torch.tensor([[True, False, True, False], [False, True, False, True]]),
            ...     static_code=torch.tensor([[1], [5]]),
            ...     static_numeric_value=torch.tensor([[1.0], [0.0]]),
            ...     static_numeric_value_mask=torch.tensor([[True], [True]]),
            ... )
            >>> print(batch._static_shape)
            (2, 1)
        """
        return (self.batch_size, self.max_static_measurements_per_subject)

    def __shape_str_lines(self) -> list[str]:
        """Gets the lines in the string representation corresponding to the shape block."""
        shape_lines = ["Shape:"]

        shape_lines.append(f"{BRANCH}Batch size: {self.batch_size}")

        seq_len_n = "Sequence length"
        if self.static_inclusion_mode == StaticInclusionMode.PREPEND:
            seq_len_n = f"{seq_len_n} (static + dynamic)"

        match self.mode:
            case BatchMode.SM:
                shape_lines.append(f"{BRANCH}{seq_len_n}: {self.max_measurements_per_subject}")
            case BatchMode.SEM:
                shape_lines.append(f"{BRANCH}{seq_len_n}: {self.max_events_per_subject}")
                shape_lines.append(f"{BRANCH}Event length: {self.max_measurements_per_event}")

        shape_lines.append(BRANCH)

        match self.mode:
            case BatchMode.SM:
                if self.static_inclusion_mode == StaticInclusionMode.PREPEND:
                    dynamic_str = "All [static; dynamic] data"
                else:
                    dynamic_str = "All dynamic data"
                shape_lines.append(f"{BRANCH}{dynamic_str}: {self._SM_shape}")
            case BatchMode.SEM:
                shape_lines.append(f"{BRANCH}Per-event data: {self._SE_shape}")
                shape_lines.append(f"{BRANCH}Per-measurement data: {self._SEM_shape}")

        if self.static_inclusion_mode == StaticInclusionMode.INCLUDE:
            shape_lines.append(f"{BRANCH}Static data: {self._static_shape}")

        if self.has_labels:
            shape_lines.append(f"{BRANCH}Labels: {self.boolean_value.shape}")
        return shape_lines

    def __mode_str_lines(self) -> list[str]:
        """Gets the lines in the string representation corresponding to the mode block."""
        mode_lines = []
        match self.mode:
            case BatchMode.SM:
                mode_lines.append(f"Mode: Subject-Measurement ({self.mode})")
            case BatchMode.SEM:
                mode_lines.append(f"Mode: Subject-Event-Measurement ({self.mode})")

        match self.static_inclusion_mode:
            case StaticInclusionMode.INCLUDE:
                mode_lines.append("Static data? ✓")
            case StaticInclusionMode.PREPEND:
                mode_lines.append("Static data? ✓ (prepended)")
            case StaticInclusionMode.OMIT:
                mode_lines.append("Static data? ✗")

        labels_symbol = "✓" if self.has_labels else "✗"
        mode_lines.append(f"Labels? {labels_symbol}")

        return mode_lines

    @staticmethod
    def __str_tensor_val(tensor: torch.Tensor) -> str:
        """Strips the `tensor(` prefix, `)` suffix, leading/trailing , and newlines."""

        tensor_str = str(tensor).replace("tensor(", "       ").replace(")", "")
        tensor_str = "\n".join([x for x in tensor_str.splitlines() if x.strip()])
        tensor_str = textwrap.dedent(tensor_str).strip()
        return tensor_str

    def __str_tensor_list(self, header: str, tensors: list[str]) -> list[str]:
        """Gets string representation lines for the requested tensors."""
        out = [f"{header}:"]
        for tensor_n in tensors:
            tensor = getattr(self, tensor_n)
            if tensor is None:
                continue

            out.append(f"{BRANCH}{tensor_n} ({tensor.dtype}):")
            tensor_str = self.__str_tensor_val(tensor)
            out.extend(textwrap.indent(tensor_str, BRANCH + BRANCH).splitlines())

        return out

    def __SM_str_lines(self) -> list[str]:
        """Gets the lines in the string representation corresponding to the SM data tensors."""
        n = "[Static; Dynamic]" if self.static_inclusion_mode == StaticInclusionMode.PREPEND else "Dynamic"
        return self.__str_tensor_list(n, self.SM_TENSOR_NAMES)

    def __SE_str_lines(self) -> list[str]:
        """Gets the lines in the string representation corresponding to the SE (event-level) data tensors."""
        return self.__str_tensor_list("Event-level", self.SE_TENSOR_NAMES)

    def __SEM_str_lines(self) -> list[str]:
        """Gets the lines in the string representation for the SEM (measurement-level) tensors."""
        return self.__str_tensor_list("Measurement-level", self.SEM_TENSOR_NAMES)

    def __static_str_lines(self) -> list[str]:
        """Gets the lines in the string representation corresponding to the static data tensors."""
        return self.__str_tensor_list("Static", self.STATIC_TENSOR_NAMES)

    def __labels_str_lines(self) -> list[str]:
        """Gets the lines in the string representation corresponding to the labels."""
        return self.__str_tensor_list("Labels", self.LABEL_TENSOR_NAMES)

    def __data_str_lines(self) -> list[str]:
        """Gets the lines in the string representation corresponding to the data block."""

        data_lines = ["Data:"]

        match self.mode:
            case BatchMode.SM:
                data_lines.extend([f"{BRANCH}{line}" for line in self.__SM_str_lines()])
            case BatchMode.SEM:
                data_lines.extend([f"{BRANCH}{line}" for line in self.__SE_str_lines()])
                data_lines.append(BRANCH)
                data_lines.extend([f"{BRANCH}{line}" for line in self.__SEM_str_lines()])

        if self.static_inclusion_mode == StaticInclusionMode.INCLUDE:
            data_lines.append(BRANCH)
            data_lines.extend([f"{BRANCH}{line}" for line in self.__static_str_lines()])

        if self.has_labels:
            data_lines.append(BRANCH)
            data_lines.extend([f"{BRANCH}{line}" for line in self.__labels_str_lines()])

        return data_lines

    def __str__(self) -> str:
        """A human-readable string representation of the batch.

        This is mostly designed for printing in doctests, and so avoids totally blank newlines (as those
        generate ugly <BLANKLINE> tags in the output).


        Examples:
            >>> print(MEDSTorchBatch(
            ...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.0]]),
            ...     event_mask=torch.tensor([[True, True], [True, False]]),
            ...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
            ...     numeric_value=torch.tensor(
            ...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
            ...     ),
            ...     numeric_value_mask=torch.tensor([
            ...         [[True, False, True], [False, False, False]],
            ...         [[False, True, False], [True, True, True]]
            ...     ]),
            ... ))
            MEDSTorchBatch:
            │ Mode: Subject-Event-Measurement (SEM)
            │ Static data? ✗
            │ Labels? ✗

            │ Shape:
            │ │ Batch size: 2
            │ │ Sequence length: 2
            │ │ Event length: 3
            │ │
            │ │ Per-event data: (2, 2)
            │ │ Per-measurement data: (2, 2, 3)

            │ Data:
            │ │ Event-level:
            │ │ │ time_delta_days (torch.float32):
            │ │ │ │ [[1.00, 2.10],
            │ │ │ │  [4.00, 0.00]]
            │ │ │ event_mask (torch.bool):
            │ │ │ │ [[ True,  True],
            │ │ │ │  [ True, False]]
            │ │
            │ │ Measurement-level:
            │ │ │ code (torch.int64):
            │ │ │ │ [[[1, 2, 3],
            │ │ │ │   [3, 0, 0]],
            │ │ │ │  [[5, 6, 0],
            │ │ │ │   [0, 0, 0]]]
            │ │ │ numeric_value (torch.float32):
            │ │ │ │ [[[ 1.,  0., -3.],
            │ │ │ │   [ 0.,  0.,  0.]],
            │ │ │ │  [[ 0.,  0.,  0.],
            │ │ │ │   [ 0.,  0.,  0.]]]
            │ │ │ numeric_value_mask (torch.bool):
            │ │ │ │ [[[ True, False,  True],
            │ │ │ │   [False, False, False]],
            │ │ │ │  [[False,  True, False],
            │ │ │ │   [ True,  True,  True]]]
            >>> print(MEDSTorchBatch(
            ...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.0]]),
            ...     event_mask=torch.tensor([[True, True], [True, False]]),
            ...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
            ...     numeric_value=torch.tensor(
            ...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
            ...     ),
            ...     numeric_value_mask=torch.tensor([
            ...         [[True, False, True], [False, False, False]],
            ...         [[False, True, False], [True, True, True]]
            ...     ]),
            ...     static_code=torch.tensor([[1], [5]]),
            ...     static_numeric_value=torch.tensor([[1.0], [0.0]]),
            ...     static_numeric_value_mask=torch.tensor([[True], [True]]),
            ... ))
            MEDSTorchBatch:
            │ Mode: Subject-Event-Measurement (SEM)
            │ Static data? ✓
            │ Labels? ✗

            │ Shape:
            │ │ Batch size: 2
            │ │ Sequence length: 2
            │ │ Event length: 3
            │ │
            │ │ Per-event data: (2, 2)
            │ │ Per-measurement data: (2, 2, 3)
            │ │ Static data: (2, 1)

            │ Data:
            │ │ Event-level:
            │ │ │ time_delta_days (torch.float32):
            │ │ │ │ [[1.00, 2.10],
            │ │ │ │  [4.00, 0.00]]
            │ │ │ event_mask (torch.bool):
            │ │ │ │ [[ True,  True],
            │ │ │ │  [ True, False]]
            │ │
            │ │ Measurement-level:
            │ │ │ code (torch.int64):
            │ │ │ │ [[[1, 2, 3],
            │ │ │ │   [3, 0, 0]],
            │ │ │ │  [[5, 6, 0],
            │ │ │ │   [0, 0, 0]]]
            │ │ │ numeric_value (torch.float32):
            │ │ │ │ [[[ 1.,  0., -3.],
            │ │ │ │   [ 0.,  0.,  0.]],
            │ │ │ │  [[ 0.,  0.,  0.],
            │ │ │ │   [ 0.,  0.,  0.]]]
            │ │ │ numeric_value_mask (torch.bool):
            │ │ │ │ [[[ True, False,  True],
            │ │ │ │   [False, False, False]],
            │ │ │ │  [[False,  True, False],
            │ │ │ │   [ True,  True,  True]]]
            │ │
            │ │ Static:
            │ │ │ static_code (torch.int64):
            │ │ │ │ [[1],
            │ │ │ │  [5]]
            │ │ │ static_numeric_value (torch.float32):
            │ │ │ │ [[1.],
            │ │ │ │  [0.]]
            │ │ │ static_numeric_value_mask (torch.bool):
            │ │ │ │ [[True],
            │ │ │ │  [True]]
            >>> print(MEDSTorchBatch(
            ...     time_delta_days=torch.tensor([[0.0, 1.0, 2.1], [0.0, 4.0, 0.0]]),
            ...     event_mask=torch.tensor([[True, True, True], [True, True, False]]),
            ...     static_mask=torch.tensor([[True, False, False], [True, False, False]]),
            ...     code=torch.tensor([[[1, 0, 0], [1, 2, 3], [3, 0, 0]], [[5, 0, 0], [5, 6, 0], [0, 0, 0]]]),
            ...     numeric_value=torch.tensor(
            ...         [[[1.0, 0.0, 0.0], [1.0, 0.0, -3.0], [0.0, 0.0, 0.0]],
            ...          [[0.0, 0.0, 0.0], [0.0, 0.0,  0.0], [0.0, 0.0, 0.0]]]
            ...     ),
            ...     numeric_value_mask=torch.tensor([
            ...         [[True, False, False], [True, False, True], [False, False, False]],
            ...         [[True, False, False], [False, True, False], [True, True, True]]
            ...     ]),
            ... ))
            MEDSTorchBatch:
            │ Mode: Subject-Event-Measurement (SEM)
            │ Static data? ✓ (prepended)
            │ Labels? ✗

            │ Shape:
            │ │ Batch size: 2
            │ │ Sequence length (static + dynamic): 3
            │ │ Event length: 3
            │ │
            │ │ Per-event data: (2, 3)
            │ │ Per-measurement data: (2, 3, 3)

            │ Data:
            │ │ Event-level:
            │ │ │ time_delta_days (torch.float32):
            │ │ │ │ [[0.00, 1.00, 2.10],
            │ │ │ │  [0.00, 4.00, 0.00]]
            │ │ │ event_mask (torch.bool):
            │ │ │ │ [[ True,  True,  True],
            │ │ │ │  [ True,  True, False]]
            │ │ │ static_mask (torch.bool):
            │ │ │ │ [[ True, False, False],
            │ │ │ │  [ True, False, False]]
            │ │
            │ │ Measurement-level:
            │ │ │ code (torch.int64):
            │ │ │ │ [[[1, 0, 0],
            │ │ │ │   [1, 2, 3],
            │ │ │ │   [3, 0, 0]],
            │ │ │ │  [[5, 0, 0],
            │ │ │ │   [5, 6, 0],
            │ │ │ │   [0, 0, 0]]]
            │ │ │ numeric_value (torch.float32):
            │ │ │ │ [[[ 1.,  0.,  0.],
            │ │ │ │   [ 1.,  0., -3.],
            │ │ │ │   [ 0.,  0.,  0.]],
            │ │ │ │  [[ 0.,  0.,  0.],
            │ │ │ │   [ 0.,  0.,  0.],
            │ │ │ │   [ 0.,  0.,  0.]]]
            │ │ │ numeric_value_mask (torch.bool):
            │ │ │ │ [[[ True, False, False],
            │ │ │ │   [ True, False,  True],
            │ │ │ │   [False, False, False]],
            │ │ │ │  [[ True, False, False],
            │ │ │ │   [False,  True, False],
            │ │ │ │   [ True,  True,  True]]]
            >>> print(MEDSTorchBatch(
            ...     time_delta_days=torch.tensor([[1.0, 0.0, 0.0, 2.1], [4.0, 0.0, 0.0, 0.0]]),
            ...     code=torch.tensor([[1, 2, 3, 3], [5, 6, 0, 0]]),
            ...     numeric_value=torch.tensor([[1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0]]),
            ...     numeric_value_mask=torch.tensor([[True, False, True, False], [False, True, False, True]]),
            ...     static_code=torch.tensor([[1], [5]]),
            ...     static_numeric_value=torch.tensor([[1.0], [0.0]]),
            ...     static_numeric_value_mask=torch.tensor([[True], [True]]),
            ...     boolean_value=torch.tensor([True, False]),
            ... ))
            MEDSTorchBatch:
            │ Mode: Subject-Measurement (SM)
            │ Static data? ✓
            │ Labels? ✓

            │ Shape:
            │ │ Batch size: 2
            │ │ Sequence length: 4
            │ │
            │ │ All dynamic data: (2, 4)
            │ │ Static data: (2, 1)
            │ │ Labels: torch.Size([2])

            │ Data:
            │ │ Dynamic:
            │ │ │ time_delta_days (torch.float32):
            │ │ │ │ [[1.00, 0.00, 0.00, 2.10],
            │ │ │ │  [4.00, 0.00, 0.00, 0.00]]
            │ │ │ code (torch.int64):
            │ │ │ │ [[1, 2, 3, 3],
            │ │ │ │  [5, 6, 0, 0]]
            │ │ │ numeric_value (torch.float32):
            │ │ │ │ [[ 1.,  0., -3.,  0.],
            │ │ │ │  [ 0.,  0.,  0.,  0.]]
            │ │ │ numeric_value_mask (torch.bool):
            │ │ │ │ [[ True, False,  True, False],
            │ │ │ │  [False,  True, False,  True]]
            │ │
            │ │ Static:
            │ │ │ static_code (torch.int64):
            │ │ │ │ [[1],
            │ │ │ │  [5]]
            │ │ │ static_numeric_value (torch.float32):
            │ │ │ │ [[1.],
            │ │ │ │  [0.]]
            │ │ │ static_numeric_value_mask (torch.bool):
            │ │ │ │ [[True],
            │ │ │ │  [True]]
            │ │
            │ │ Labels:
            │ │ │ boolean_value (torch.bool):
            │ │ │ │ [ True, False]
            >>> print(MEDSTorchBatch(
            ...     time_delta_days=torch.tensor([[0.0, 1.0, 0.0, 0.0, 2.1], [0.0, 4.0, 0.0, 0.0, 0.0]]),
            ...     code=torch.tensor([[1, 1, 2, 3, 3], [5, 5, 6, 0, 0]]),
            ...     numeric_value=torch.tensor([[1.0, 1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0]]),
            ...     numeric_value_mask=torch.tensor(
            ...         [[True, True, False, True, False], [True, False, True, False, True]]
            ...     ),
            ...     static_mask=torch.tensor(
            ...         [[True, False, False, False, False], [True, False, False, False, False]]
            ...     ),
            ...     boolean_value=torch.tensor([True, False]),
            ... ))
            MEDSTorchBatch:
            │ Mode: Subject-Measurement (SM)
            │ Static data? ✓ (prepended)
            │ Labels? ✓

            │ Shape:
            │ │ Batch size: 2
            │ │ Sequence length (static + dynamic): 5
            │ │
            │ │ All [static; dynamic] data: (2, 5)
            │ │ Labels: torch.Size([2])

            │ Data:
            │ │ [Static; Dynamic]:
            │ │ │ time_delta_days (torch.float32):
            │ │ │ │ [[0.00, 1.00,  ..., 0.00, 2.10],
            │ │ │ │  [0.00, 4.00,  ..., 0.00, 0.00]]
            │ │ │ code (torch.int64):
            │ │ │ │ [[1, 1, ..., 3, 3],
            │ │ │ │  [5, 5, ..., 0, 0]]
            │ │ │ numeric_value (torch.float32):
            │ │ │ │ [[ 1., 1.,  ..., -3.,  0.],
            │ │ │ │  [ 0., 0.,  ...,  0.,  0.]]
            │ │ │ numeric_value_mask (torch.bool):
            │ │ │ │ [[ True,  True, ...,  True, False],
            │ │ │ │  [ True, False, ..., False,  True]]
            │ │ │ static_mask (torch.bool):
            │ │ │ │ [[ True, False, ..., False, False],
            │ │ │ │  [ True, False, ..., False, False]]
            │ │
            │ │ Labels:
            │ │ │ boolean_value (torch.bool):
            │ │ │ │ [ True, False]
        """

        lines = [f"{self.__class__.__name__}:"]

        torch.set_printoptions(precision=2, threshold=5, edgeitems=2)

        lines.extend([f"{BRANCH}{line}" for line in self.__mode_str_lines()])
        lines.append(BRANCH)
        lines.extend([f"{BRANCH}{line}" for line in self.__shape_str_lines()])
        lines.append(BRANCH)
        lines.extend([f"{BRANCH}{line}" for line in self.__data_str_lines()])

        torch.set_printoptions(profile="default")

        lines = [line.rstrip() for line in lines]

        return "\n".join(lines)

_SEM_shape property

Returns the subject-event-measurement shape of the batch. Only valid in SEM mode.

Examples:

>>> batch = MEDSTorchBatch(
...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.0]]),
...     event_mask=torch.tensor([[True, True], [True, False]]),
...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
...     numeric_value=torch.tensor(
...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
...     ),
...     numeric_value_mask=torch.tensor([
...         [[True, False, True], [False, False, False]],
...         [[False, True, False], [True, True, True]]
...     ]), # Note the padding values may be  True or False
... )
>>> print(batch._SEM_shape)
(2, 2, 3)

_SE_shape property

Returns the subject-event shape of the batch. Only valid in SEM mode.

Examples:

>>> batch = MEDSTorchBatch(
...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.0]]),
...     event_mask=torch.tensor([[True, True], [True, False]]),
...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
...     numeric_value=torch.tensor(
...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
...     ),
...     numeric_value_mask=torch.tensor([
...         [[True, False, True], [False, False, False]],
...         [[False, True, False], [True, True, True]]
...     ]), # Note the padding values may be  True or False
... )
>>> print(batch._SE_shape)
(2, 2)

_SM_shape property

Returns the subject-measurement shape of the batch. Only valid in SM mode.

Examples:

>>> batch = MEDSTorchBatch(
...     time_delta_days=torch.tensor([[1.0, 0.0, 0.0, 2.1], [4.0, 0.0, 0.0, 0.0]]),
...     code=torch.tensor([[1, 2, 3, 3], [5, 6, 0, 0]]),
...     numeric_value=torch.tensor([[1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0]]),
...     numeric_value_mask=torch.tensor([[True, False, True, False], [False, True, False, True]]),
... )
>>> print(batch._SM_shape)
(2, 4)

_static_shape property

Returns the static data shape of the batch. Only valid if the batch has static data.

Examples:

>>> batch = MEDSTorchBatch(
...     time_delta_days=torch.tensor([[1.0, 0.0, 0.0, 2.1], [4.0, 0.0, 0.0, 0.0]]),
...     code=torch.tensor([[1, 2, 3, 3], [5, 6, 0, 0]]),
...     numeric_value=torch.tensor([[1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0]]),
...     numeric_value_mask=torch.tensor([[True, False, True, False], [False, True, False, True]]),
...     static_code=torch.tensor([[1], [5]]),
...     static_numeric_value=torch.tensor([[1.0], [0.0]]),
...     static_numeric_value_mask=torch.tensor([[True], [True]]),
... )
>>> print(batch._static_shape)
(2, 1)

batch_size property

The number of subjects in the batch.

has_labels property

Whether the batch has labels.

max_events_per_subject property

The maximum number of events for any subject in the batch.

Only valid in SEM mode.

max_measurements_per_event property

The maximum number of measurements for any event in the batch.

Only valid in SEM mode.

max_measurements_per_subject property

The maximum number of measurements for any subject in the batch.

Only valid in SM mode.

max_static_measurements_per_subject property

The maximum number of static measurements for any subject in the batch.

mode property

The mode of the batch, reflecting the internal organization of subject measurements.

__SEM_str_lines()

Gets the lines in the string representation for the SEM (measurement-level) tensors.

Source code in meds_torchdata/types.py
def __SEM_str_lines(self) -> list[str]:
    """Gets the lines in the string representation for the SEM (measurement-level) tensors."""
    return self.__str_tensor_list("Measurement-level", self.SEM_TENSOR_NAMES)

__SE_str_lines()

Gets the lines in the string representation corresponding to the SE (event-level) data tensors.

Source code in meds_torchdata/types.py
def __SE_str_lines(self) -> list[str]:
    """Gets the lines in the string representation corresponding to the SE (event-level) data tensors."""
    return self.__str_tensor_list("Event-level", self.SE_TENSOR_NAMES)

__SM_str_lines()

Gets the lines in the string representation corresponding to the SM data tensors.

Source code in meds_torchdata/types.py
def __SM_str_lines(self) -> list[str]:
    """Gets the lines in the string representation corresponding to the SM data tensors."""
    n = "[Static; Dynamic]" if self.static_inclusion_mode == StaticInclusionMode.PREPEND else "Dynamic"
    return self.__str_tensor_list(n, self.SM_TENSOR_NAMES)

__check_shape(name, shape)

Check that the shape of a tensor matches the expected shape, or raise an appropriate error.

Source code in meds_torchdata/types.py
def __check_shape(self, name: str, shape: tuple[int, ...]) -> None:
    """Check that the shape of a tensor matches the expected shape, or raise an appropriate error."""
    got_shape = getattr(self, name).shape
    if got_shape != shape:
        raise ValueError(f"Expected shape {shape} for {name}, but got {got_shape}!")

__data_str_lines()

Gets the lines in the string representation corresponding to the data block.

Source code in meds_torchdata/types.py
def __data_str_lines(self) -> list[str]:
    """Gets the lines in the string representation corresponding to the data block."""

    data_lines = ["Data:"]

    match self.mode:
        case BatchMode.SM:
            data_lines.extend([f"{BRANCH}{line}" for line in self.__SM_str_lines()])
        case BatchMode.SEM:
            data_lines.extend([f"{BRANCH}{line}" for line in self.__SE_str_lines()])
            data_lines.append(BRANCH)
            data_lines.extend([f"{BRANCH}{line}" for line in self.__SEM_str_lines()])

    if self.static_inclusion_mode == StaticInclusionMode.INCLUDE:
        data_lines.append(BRANCH)
        data_lines.extend([f"{BRANCH}{line}" for line in self.__static_str_lines()])

    if self.has_labels:
        data_lines.append(BRANCH)
        data_lines.extend([f"{BRANCH}{line}" for line in self.__labels_str_lines()])

    return data_lines

__getitem__(key)

Get a tensor from the batch by key.

Source code in meds_torchdata/types.py
def __getitem__(self, key: str) -> torch.Tensor:
    """Get a tensor from the batch by key."""
    return getattr(self, key)

__labels_str_lines()

Gets the lines in the string representation corresponding to the labels.

Source code in meds_torchdata/types.py
def __labels_str_lines(self) -> list[str]:
    """Gets the lines in the string representation corresponding to the labels."""
    return self.__str_tensor_list("Labels", self.LABEL_TENSOR_NAMES)

__mode_str_lines()

Gets the lines in the string representation corresponding to the mode block.

Source code in meds_torchdata/types.py
def __mode_str_lines(self) -> list[str]:
    """Gets the lines in the string representation corresponding to the mode block."""
    mode_lines = []
    match self.mode:
        case BatchMode.SM:
            mode_lines.append(f"Mode: Subject-Measurement ({self.mode})")
        case BatchMode.SEM:
            mode_lines.append(f"Mode: Subject-Event-Measurement ({self.mode})")

    match self.static_inclusion_mode:
        case StaticInclusionMode.INCLUDE:
            mode_lines.append("Static data? ✓")
        case StaticInclusionMode.PREPEND:
            mode_lines.append("Static data? ✓ (prepended)")
        case StaticInclusionMode.OMIT:
            mode_lines.append("Static data? ✗")

    labels_symbol = "✓" if self.has_labels else "✗"
    mode_lines.append(f"Labels? {labels_symbol}")

    return mode_lines

__post_init__()

Check that the batch is well-formed, raising an error if it is not.

Source code in meds_torchdata/types.py
def __post_init__(self):
    """Check that the batch is well-formed, raising an error if it is not."""
    for field in fields(self):
        tensor_type = get_args(field.type)[0]
        match value := getattr(self, field.name):
            case None:
                if field.name in self._REQ_TENSORS:
                    raise ValueError(f"Required tensor {field.name} is missing!")
                else:
                    pass
            case tensor_type():
                pass
            case _:
                raise TypeError(
                    f"Field '{field.name}' expected type {tensor_type}, got type {type(value)}."
                )

    # numeric_value and numeric_value_mask must be provided (or omitted) as a pair.
    if (self.numeric_value is None) != (self.numeric_value_mask is None):
        raise ValueError(
            "numeric_value and numeric_value_mask must both be provided or both be None, "
            f"but got numeric_value={'present' if self.numeric_value is not None else 'None'} "
            f"and numeric_value_mask={'present' if self.numeric_value_mask is not None else 'None'}."
        )

    # Dynamic-field shape checks. `time_delta_days`, `numeric_value`, and
    # `numeric_value_mask` are all optional now (see `_REQ_TENSORS` — gated by
    # `MEDSTorchDataConfig.include_time_delta` and `include_numeric_value`), so we only
    # check the shape of the ones the caller actually provided. `code` is always present
    # and `event_mask` is required in SEM mode because it defines the per-event shape.
    match self.mode:
        case BatchMode.SEM:
            if self.event_mask is None:
                raise ValueError(f"Event mask must be provided in {self.mode} mode!")
            if self.time_delta_days is not None:
                self.__check_shape("time_delta_days", self._SE_shape)
            self.__check_shape("event_mask", self._SE_shape)
            if self.numeric_value is not None:
                self.__check_shape("numeric_value", self._SEM_shape)
            if self.numeric_value_mask is not None:
                self.__check_shape("numeric_value_mask", self._SEM_shape)
        case BatchMode.SM:
            if self.event_mask is not None:
                raise ValueError(f"Event mask should not be provided in {self.mode} mode!")
            if self.time_delta_days is not None:
                self.__check_shape("time_delta_days", self._SM_shape)
            if self.numeric_value is not None:
                self.__check_shape("numeric_value", self._SM_shape)
            if self.numeric_value_mask is not None:
                self.__check_shape("numeric_value_mask", self._SM_shape)
        case _:  # pragma: no cover
            raise ValueError(f"Invalid mode {self.mode}!")

    match self.static_inclusion_mode:
        case StaticInclusionMode.INCLUDE:
            if self.static_mask is not None:
                raise ValueError("Static mask should not be provided if static codes are!")
            if self.static_numeric_value is None or self.static_numeric_value_mask is None:
                raise ValueError(
                    "Static numeric value and mask must both be provided if static codes are!"
                )
            if len(self.static_code.shape) != 2 or self.static_code.shape[0] != self.batch_size:
                raise ValueError(
                    f"Expected 2D static data tensors with a matching batch size ({self.batch_size}), "
                    f"but got static_code shape {self.static_code.shape}!"
                )
            self.__check_shape("static_numeric_value", self._static_shape)
            self.__check_shape("static_numeric_value_mask", self._static_shape)
        case StaticInclusionMode.OMIT:
            if self.static_numeric_value is not None or self.static_numeric_value_mask is not None:
                raise ValueError("Static numeric value and mask should not be provided without codes!")
        case StaticInclusionMode.PREPEND:
            if self.static_numeric_value is not None or self.static_numeric_value_mask is not None:
                raise ValueError("Static numeric value and mask should not be provided with static mask!")
            if self.mode == BatchMode.SEM:
                self.__check_shape("static_mask", self._SE_shape)
            elif self.mode == BatchMode.SM:
                self.__check_shape("static_mask", self._SM_shape)
        case _:  # pragma: no cover
            raise ValueError(f"Invalid static inclusion mode {self.static_inclusion_mode}!")

    if self.has_labels:
        self.__check_shape("boolean_value", (self.batch_size,))

    if self.n_subject_windows is not None:
        self.__check_shape("n_subject_windows", (self.batch_size,))

__setitem__(key, value)

Set a tensor in the batch by key.

Only valid if the key is a valid field.

Source code in meds_torchdata/types.py
def __setitem__(self, key: str, value: torch.Tensor) -> None:
    """Set a tensor in the batch by key.

    Only valid if the key is a valid field.
    """
    raise ValueError("MEDSTorchBatch is immutable!")

__shape_str_lines()

Gets the lines in the string representation corresponding to the shape block.

Source code in meds_torchdata/types.py
def __shape_str_lines(self) -> list[str]:
    """Gets the lines in the string representation corresponding to the shape block."""
    shape_lines = ["Shape:"]

    shape_lines.append(f"{BRANCH}Batch size: {self.batch_size}")

    seq_len_n = "Sequence length"
    if self.static_inclusion_mode == StaticInclusionMode.PREPEND:
        seq_len_n = f"{seq_len_n} (static + dynamic)"

    match self.mode:
        case BatchMode.SM:
            shape_lines.append(f"{BRANCH}{seq_len_n}: {self.max_measurements_per_subject}")
        case BatchMode.SEM:
            shape_lines.append(f"{BRANCH}{seq_len_n}: {self.max_events_per_subject}")
            shape_lines.append(f"{BRANCH}Event length: {self.max_measurements_per_event}")

    shape_lines.append(BRANCH)

    match self.mode:
        case BatchMode.SM:
            if self.static_inclusion_mode == StaticInclusionMode.PREPEND:
                dynamic_str = "All [static; dynamic] data"
            else:
                dynamic_str = "All dynamic data"
            shape_lines.append(f"{BRANCH}{dynamic_str}: {self._SM_shape}")
        case BatchMode.SEM:
            shape_lines.append(f"{BRANCH}Per-event data: {self._SE_shape}")
            shape_lines.append(f"{BRANCH}Per-measurement data: {self._SEM_shape}")

    if self.static_inclusion_mode == StaticInclusionMode.INCLUDE:
        shape_lines.append(f"{BRANCH}Static data: {self._static_shape}")

    if self.has_labels:
        shape_lines.append(f"{BRANCH}Labels: {self.boolean_value.shape}")
    return shape_lines

__static_str_lines()

Gets the lines in the string representation corresponding to the static data tensors.

Source code in meds_torchdata/types.py
def __static_str_lines(self) -> list[str]:
    """Gets the lines in the string representation corresponding to the static data tensors."""
    return self.__str_tensor_list("Static", self.STATIC_TENSOR_NAMES)

__str__()

A human-readable string representation of the batch.

This is mostly designed for printing in doctests, and so avoids totally blank newlines (as those generate ugly tags in the output).

Examples:

>>> print(MEDSTorchBatch(
...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.0]]),
...     event_mask=torch.tensor([[True, True], [True, False]]),
...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
...     numeric_value=torch.tensor(
...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
...     ),
...     numeric_value_mask=torch.tensor([
...         [[True, False, True], [False, False, False]],
...         [[False, True, False], [True, True, True]]
...     ]),
... ))
MEDSTorchBatch:
│ Mode: Subject-Event-Measurement (SEM)
│ Static data? ✗
│ Labels? ✗

│ Shape:
│ │ Batch size: 2
│ │ Sequence length: 2
│ │ Event length: 3
│ │
│ │ Per-event data: (2, 2)
│ │ Per-measurement data: (2, 2, 3)

│ Data:
│ │ Event-level:
│ │ │ time_delta_days (torch.float32):
│ │ │ │ [[1.00, 2.10],
│ │ │ │  [4.00, 0.00]]
│ │ │ event_mask (torch.bool):
│ │ │ │ [[ True,  True],
│ │ │ │  [ True, False]]
│ │
│ │ Measurement-level:
│ │ │ code (torch.int64):
│ │ │ │ [[[1, 2, 3],
│ │ │ │   [3, 0, 0]],
│ │ │ │  [[5, 6, 0],
│ │ │ │   [0, 0, 0]]]
│ │ │ numeric_value (torch.float32):
│ │ │ │ [[[ 1.,  0., -3.],
│ │ │ │   [ 0.,  0.,  0.]],
│ │ │ │  [[ 0.,  0.,  0.],
│ │ │ │   [ 0.,  0.,  0.]]]
│ │ │ numeric_value_mask (torch.bool):
│ │ │ │ [[[ True, False,  True],
│ │ │ │   [False, False, False]],
│ │ │ │  [[False,  True, False],
│ │ │ │   [ True,  True,  True]]]
>>> print(MEDSTorchBatch(
...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.0]]),
...     event_mask=torch.tensor([[True, True], [True, False]]),
...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
...     numeric_value=torch.tensor(
...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
...     ),
...     numeric_value_mask=torch.tensor([
...         [[True, False, True], [False, False, False]],
...         [[False, True, False], [True, True, True]]
...     ]),
...     static_code=torch.tensor([[1], [5]]),
...     static_numeric_value=torch.tensor([[1.0], [0.0]]),
...     static_numeric_value_mask=torch.tensor([[True], [True]]),
... ))
MEDSTorchBatch:
│ Mode: Subject-Event-Measurement (SEM)
│ Static data? ✓
│ Labels? ✗

│ Shape:
│ │ Batch size: 2
│ │ Sequence length: 2
│ │ Event length: 3
│ │
│ │ Per-event data: (2, 2)
│ │ Per-measurement data: (2, 2, 3)
│ │ Static data: (2, 1)

│ Data:
│ │ Event-level:
│ │ │ time_delta_days (torch.float32):
│ │ │ │ [[1.00, 2.10],
│ │ │ │  [4.00, 0.00]]
│ │ │ event_mask (torch.bool):
│ │ │ │ [[ True,  True],
│ │ │ │  [ True, False]]
│ │
│ │ Measurement-level:
│ │ │ code (torch.int64):
│ │ │ │ [[[1, 2, 3],
│ │ │ │   [3, 0, 0]],
│ │ │ │  [[5, 6, 0],
│ │ │ │   [0, 0, 0]]]
│ │ │ numeric_value (torch.float32):
│ │ │ │ [[[ 1.,  0., -3.],
│ │ │ │   [ 0.,  0.,  0.]],
│ │ │ │  [[ 0.,  0.,  0.],
│ │ │ │   [ 0.,  0.,  0.]]]
│ │ │ numeric_value_mask (torch.bool):
│ │ │ │ [[[ True, False,  True],
│ │ │ │   [False, False, False]],
│ │ │ │  [[False,  True, False],
│ │ │ │   [ True,  True,  True]]]
│ │
│ │ Static:
│ │ │ static_code (torch.int64):
│ │ │ │ [[1],
│ │ │ │  [5]]
│ │ │ static_numeric_value (torch.float32):
│ │ │ │ [[1.],
│ │ │ │  [0.]]
│ │ │ static_numeric_value_mask (torch.bool):
│ │ │ │ [[True],
│ │ │ │  [True]]
>>> print(MEDSTorchBatch(
...     time_delta_days=torch.tensor([[0.0, 1.0, 2.1], [0.0, 4.0, 0.0]]),
...     event_mask=torch.tensor([[True, True, True], [True, True, False]]),
...     static_mask=torch.tensor([[True, False, False], [True, False, False]]),
...     code=torch.tensor([[[1, 0, 0], [1, 2, 3], [3, 0, 0]], [[5, 0, 0], [5, 6, 0], [0, 0, 0]]]),
...     numeric_value=torch.tensor(
...         [[[1.0, 0.0, 0.0], [1.0, 0.0, -3.0], [0.0, 0.0, 0.0]],
...          [[0.0, 0.0, 0.0], [0.0, 0.0,  0.0], [0.0, 0.0, 0.0]]]
...     ),
...     numeric_value_mask=torch.tensor([
...         [[True, False, False], [True, False, True], [False, False, False]],
...         [[True, False, False], [False, True, False], [True, True, True]]
...     ]),
... ))
MEDSTorchBatch:
│ Mode: Subject-Event-Measurement (SEM)
│ Static data? ✓ (prepended)
│ Labels? ✗

│ Shape:
│ │ Batch size: 2
│ │ Sequence length (static + dynamic): 3
│ │ Event length: 3
│ │
│ │ Per-event data: (2, 3)
│ │ Per-measurement data: (2, 3, 3)

│ Data:
│ │ Event-level:
│ │ │ time_delta_days (torch.float32):
│ │ │ │ [[0.00, 1.00, 2.10],
│ │ │ │  [0.00, 4.00, 0.00]]
│ │ │ event_mask (torch.bool):
│ │ │ │ [[ True,  True,  True],
│ │ │ │  [ True,  True, False]]
│ │ │ static_mask (torch.bool):
│ │ │ │ [[ True, False, False],
│ │ │ │  [ True, False, False]]
│ │
│ │ Measurement-level:
│ │ │ code (torch.int64):
│ │ │ │ [[[1, 0, 0],
│ │ │ │   [1, 2, 3],
│ │ │ │   [3, 0, 0]],
│ │ │ │  [[5, 0, 0],
│ │ │ │   [5, 6, 0],
│ │ │ │   [0, 0, 0]]]
│ │ │ numeric_value (torch.float32):
│ │ │ │ [[[ 1.,  0.,  0.],
│ │ │ │   [ 1.,  0., -3.],
│ │ │ │   [ 0.,  0.,  0.]],
│ │ │ │  [[ 0.,  0.,  0.],
│ │ │ │   [ 0.,  0.,  0.],
│ │ │ │   [ 0.,  0.,  0.]]]
│ │ │ numeric_value_mask (torch.bool):
│ │ │ │ [[[ True, False, False],
│ │ │ │   [ True, False,  True],
│ │ │ │   [False, False, False]],
│ │ │ │  [[ True, False, False],
│ │ │ │   [False,  True, False],
│ │ │ │   [ True,  True,  True]]]
>>> print(MEDSTorchBatch(
...     time_delta_days=torch.tensor([[1.0, 0.0, 0.0, 2.1], [4.0, 0.0, 0.0, 0.0]]),
...     code=torch.tensor([[1, 2, 3, 3], [5, 6, 0, 0]]),
...     numeric_value=torch.tensor([[1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0]]),
...     numeric_value_mask=torch.tensor([[True, False, True, False], [False, True, False, True]]),
...     static_code=torch.tensor([[1], [5]]),
...     static_numeric_value=torch.tensor([[1.0], [0.0]]),
...     static_numeric_value_mask=torch.tensor([[True], [True]]),
...     boolean_value=torch.tensor([True, False]),
... ))
MEDSTorchBatch:
│ Mode: Subject-Measurement (SM)
│ Static data? ✓
│ Labels? ✓

│ Shape:
│ │ Batch size: 2
│ │ Sequence length: 4
│ │
│ │ All dynamic data: (2, 4)
│ │ Static data: (2, 1)
│ │ Labels: torch.Size([2])

│ Data:
│ │ Dynamic:
│ │ │ time_delta_days (torch.float32):
│ │ │ │ [[1.00, 0.00, 0.00, 2.10],
│ │ │ │  [4.00, 0.00, 0.00, 0.00]]
│ │ │ code (torch.int64):
│ │ │ │ [[1, 2, 3, 3],
│ │ │ │  [5, 6, 0, 0]]
│ │ │ numeric_value (torch.float32):
│ │ │ │ [[ 1.,  0., -3.,  0.],
│ │ │ │  [ 0.,  0.,  0.,  0.]]
│ │ │ numeric_value_mask (torch.bool):
│ │ │ │ [[ True, False,  True, False],
│ │ │ │  [False,  True, False,  True]]
│ │
│ │ Static:
│ │ │ static_code (torch.int64):
│ │ │ │ [[1],
│ │ │ │  [5]]
│ │ │ static_numeric_value (torch.float32):
│ │ │ │ [[1.],
│ │ │ │  [0.]]
│ │ │ static_numeric_value_mask (torch.bool):
│ │ │ │ [[True],
│ │ │ │  [True]]
│ │
│ │ Labels:
│ │ │ boolean_value (torch.bool):
│ │ │ │ [ True, False]
>>> print(MEDSTorchBatch(
...     time_delta_days=torch.tensor([[0.0, 1.0, 0.0, 0.0, 2.1], [0.0, 4.0, 0.0, 0.0, 0.0]]),
...     code=torch.tensor([[1, 1, 2, 3, 3], [5, 5, 6, 0, 0]]),
...     numeric_value=torch.tensor([[1.0, 1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0]]),
...     numeric_value_mask=torch.tensor(
...         [[True, True, False, True, False], [True, False, True, False, True]]
...     ),
...     static_mask=torch.tensor(
...         [[True, False, False, False, False], [True, False, False, False, False]]
...     ),
...     boolean_value=torch.tensor([True, False]),
... ))
MEDSTorchBatch:
│ Mode: Subject-Measurement (SM)
│ Static data? ✓ (prepended)
│ Labels? ✓

│ Shape:
│ │ Batch size: 2
│ │ Sequence length (static + dynamic): 5
│ │
│ │ All [static; dynamic] data: (2, 5)
│ │ Labels: torch.Size([2])

│ Data:
│ │ [Static; Dynamic]:
│ │ │ time_delta_days (torch.float32):
│ │ │ │ [[0.00, 1.00,  ..., 0.00, 2.10],
│ │ │ │  [0.00, 4.00,  ..., 0.00, 0.00]]
│ │ │ code (torch.int64):
│ │ │ │ [[1, 1, ..., 3, 3],
│ │ │ │  [5, 5, ..., 0, 0]]
│ │ │ numeric_value (torch.float32):
│ │ │ │ [[ 1., 1.,  ..., -3.,  0.],
│ │ │ │  [ 0., 0.,  ...,  0.,  0.]]
│ │ │ numeric_value_mask (torch.bool):
│ │ │ │ [[ True,  True, ...,  True, False],
│ │ │ │  [ True, False, ..., False,  True]]
│ │ │ static_mask (torch.bool):
│ │ │ │ [[ True, False, ..., False, False],
│ │ │ │  [ True, False, ..., False, False]]
│ │
│ │ Labels:
│ │ │ boolean_value (torch.bool):
│ │ │ │ [ True, False]
Source code in meds_torchdata/types.py
def __str__(self) -> str:
    """A human-readable string representation of the batch.

    This is mostly designed for printing in doctests, and so avoids totally blank newlines (as those
    generate ugly <BLANKLINE> tags in the output).


    Examples:
        >>> print(MEDSTorchBatch(
        ...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.0]]),
        ...     event_mask=torch.tensor([[True, True], [True, False]]),
        ...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
        ...     numeric_value=torch.tensor(
        ...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
        ...     ),
        ...     numeric_value_mask=torch.tensor([
        ...         [[True, False, True], [False, False, False]],
        ...         [[False, True, False], [True, True, True]]
        ...     ]),
        ... ))
        MEDSTorchBatch:
        │ Mode: Subject-Event-Measurement (SEM)
        │ Static data? ✗
        │ Labels? ✗

        │ Shape:
        │ │ Batch size: 2
        │ │ Sequence length: 2
        │ │ Event length: 3
        │ │
        │ │ Per-event data: (2, 2)
        │ │ Per-measurement data: (2, 2, 3)

        │ Data:
        │ │ Event-level:
        │ │ │ time_delta_days (torch.float32):
        │ │ │ │ [[1.00, 2.10],
        │ │ │ │  [4.00, 0.00]]
        │ │ │ event_mask (torch.bool):
        │ │ │ │ [[ True,  True],
        │ │ │ │  [ True, False]]
        │ │
        │ │ Measurement-level:
        │ │ │ code (torch.int64):
        │ │ │ │ [[[1, 2, 3],
        │ │ │ │   [3, 0, 0]],
        │ │ │ │  [[5, 6, 0],
        │ │ │ │   [0, 0, 0]]]
        │ │ │ numeric_value (torch.float32):
        │ │ │ │ [[[ 1.,  0., -3.],
        │ │ │ │   [ 0.,  0.,  0.]],
        │ │ │ │  [[ 0.,  0.,  0.],
        │ │ │ │   [ 0.,  0.,  0.]]]
        │ │ │ numeric_value_mask (torch.bool):
        │ │ │ │ [[[ True, False,  True],
        │ │ │ │   [False, False, False]],
        │ │ │ │  [[False,  True, False],
        │ │ │ │   [ True,  True,  True]]]
        >>> print(MEDSTorchBatch(
        ...     time_delta_days=torch.tensor([[1.0, 2.1], [4.0, 0.0]]),
        ...     event_mask=torch.tensor([[True, True], [True, False]]),
        ...     code=torch.tensor([[[1, 2, 3], [3, 0, 0]], [[5, 6, 0], [0, 0, 0]]]),
        ...     numeric_value=torch.tensor(
        ...         [[[1.0, 0.0, -3.0], [0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]]
        ...     ),
        ...     numeric_value_mask=torch.tensor([
        ...         [[True, False, True], [False, False, False]],
        ...         [[False, True, False], [True, True, True]]
        ...     ]),
        ...     static_code=torch.tensor([[1], [5]]),
        ...     static_numeric_value=torch.tensor([[1.0], [0.0]]),
        ...     static_numeric_value_mask=torch.tensor([[True], [True]]),
        ... ))
        MEDSTorchBatch:
        │ Mode: Subject-Event-Measurement (SEM)
        │ Static data? ✓
        │ Labels? ✗

        │ Shape:
        │ │ Batch size: 2
        │ │ Sequence length: 2
        │ │ Event length: 3
        │ │
        │ │ Per-event data: (2, 2)
        │ │ Per-measurement data: (2, 2, 3)
        │ │ Static data: (2, 1)

        │ Data:
        │ │ Event-level:
        │ │ │ time_delta_days (torch.float32):
        │ │ │ │ [[1.00, 2.10],
        │ │ │ │  [4.00, 0.00]]
        │ │ │ event_mask (torch.bool):
        │ │ │ │ [[ True,  True],
        │ │ │ │  [ True, False]]
        │ │
        │ │ Measurement-level:
        │ │ │ code (torch.int64):
        │ │ │ │ [[[1, 2, 3],
        │ │ │ │   [3, 0, 0]],
        │ │ │ │  [[5, 6, 0],
        │ │ │ │   [0, 0, 0]]]
        │ │ │ numeric_value (torch.float32):
        │ │ │ │ [[[ 1.,  0., -3.],
        │ │ │ │   [ 0.,  0.,  0.]],
        │ │ │ │  [[ 0.,  0.,  0.],
        │ │ │ │   [ 0.,  0.,  0.]]]
        │ │ │ numeric_value_mask (torch.bool):
        │ │ │ │ [[[ True, False,  True],
        │ │ │ │   [False, False, False]],
        │ │ │ │  [[False,  True, False],
        │ │ │ │   [ True,  True,  True]]]
        │ │
        │ │ Static:
        │ │ │ static_code (torch.int64):
        │ │ │ │ [[1],
        │ │ │ │  [5]]
        │ │ │ static_numeric_value (torch.float32):
        │ │ │ │ [[1.],
        │ │ │ │  [0.]]
        │ │ │ static_numeric_value_mask (torch.bool):
        │ │ │ │ [[True],
        │ │ │ │  [True]]
        >>> print(MEDSTorchBatch(
        ...     time_delta_days=torch.tensor([[0.0, 1.0, 2.1], [0.0, 4.0, 0.0]]),
        ...     event_mask=torch.tensor([[True, True, True], [True, True, False]]),
        ...     static_mask=torch.tensor([[True, False, False], [True, False, False]]),
        ...     code=torch.tensor([[[1, 0, 0], [1, 2, 3], [3, 0, 0]], [[5, 0, 0], [5, 6, 0], [0, 0, 0]]]),
        ...     numeric_value=torch.tensor(
        ...         [[[1.0, 0.0, 0.0], [1.0, 0.0, -3.0], [0.0, 0.0, 0.0]],
        ...          [[0.0, 0.0, 0.0], [0.0, 0.0,  0.0], [0.0, 0.0, 0.0]]]
        ...     ),
        ...     numeric_value_mask=torch.tensor([
        ...         [[True, False, False], [True, False, True], [False, False, False]],
        ...         [[True, False, False], [False, True, False], [True, True, True]]
        ...     ]),
        ... ))
        MEDSTorchBatch:
        │ Mode: Subject-Event-Measurement (SEM)
        │ Static data? ✓ (prepended)
        │ Labels? ✗

        │ Shape:
        │ │ Batch size: 2
        │ │ Sequence length (static + dynamic): 3
        │ │ Event length: 3
        │ │
        │ │ Per-event data: (2, 3)
        │ │ Per-measurement data: (2, 3, 3)

        │ Data:
        │ │ Event-level:
        │ │ │ time_delta_days (torch.float32):
        │ │ │ │ [[0.00, 1.00, 2.10],
        │ │ │ │  [0.00, 4.00, 0.00]]
        │ │ │ event_mask (torch.bool):
        │ │ │ │ [[ True,  True,  True],
        │ │ │ │  [ True,  True, False]]
        │ │ │ static_mask (torch.bool):
        │ │ │ │ [[ True, False, False],
        │ │ │ │  [ True, False, False]]
        │ │
        │ │ Measurement-level:
        │ │ │ code (torch.int64):
        │ │ │ │ [[[1, 0, 0],
        │ │ │ │   [1, 2, 3],
        │ │ │ │   [3, 0, 0]],
        │ │ │ │  [[5, 0, 0],
        │ │ │ │   [5, 6, 0],
        │ │ │ │   [0, 0, 0]]]
        │ │ │ numeric_value (torch.float32):
        │ │ │ │ [[[ 1.,  0.,  0.],
        │ │ │ │   [ 1.,  0., -3.],
        │ │ │ │   [ 0.,  0.,  0.]],
        │ │ │ │  [[ 0.,  0.,  0.],
        │ │ │ │   [ 0.,  0.,  0.],
        │ │ │ │   [ 0.,  0.,  0.]]]
        │ │ │ numeric_value_mask (torch.bool):
        │ │ │ │ [[[ True, False, False],
        │ │ │ │   [ True, False,  True],
        │ │ │ │   [False, False, False]],
        │ │ │ │  [[ True, False, False],
        │ │ │ │   [False,  True, False],
        │ │ │ │   [ True,  True,  True]]]
        >>> print(MEDSTorchBatch(
        ...     time_delta_days=torch.tensor([[1.0, 0.0, 0.0, 2.1], [4.0, 0.0, 0.0, 0.0]]),
        ...     code=torch.tensor([[1, 2, 3, 3], [5, 6, 0, 0]]),
        ...     numeric_value=torch.tensor([[1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0]]),
        ...     numeric_value_mask=torch.tensor([[True, False, True, False], [False, True, False, True]]),
        ...     static_code=torch.tensor([[1], [5]]),
        ...     static_numeric_value=torch.tensor([[1.0], [0.0]]),
        ...     static_numeric_value_mask=torch.tensor([[True], [True]]),
        ...     boolean_value=torch.tensor([True, False]),
        ... ))
        MEDSTorchBatch:
        │ Mode: Subject-Measurement (SM)
        │ Static data? ✓
        │ Labels? ✓

        │ Shape:
        │ │ Batch size: 2
        │ │ Sequence length: 4
        │ │
        │ │ All dynamic data: (2, 4)
        │ │ Static data: (2, 1)
        │ │ Labels: torch.Size([2])

        │ Data:
        │ │ Dynamic:
        │ │ │ time_delta_days (torch.float32):
        │ │ │ │ [[1.00, 0.00, 0.00, 2.10],
        │ │ │ │  [4.00, 0.00, 0.00, 0.00]]
        │ │ │ code (torch.int64):
        │ │ │ │ [[1, 2, 3, 3],
        │ │ │ │  [5, 6, 0, 0]]
        │ │ │ numeric_value (torch.float32):
        │ │ │ │ [[ 1.,  0., -3.,  0.],
        │ │ │ │  [ 0.,  0.,  0.,  0.]]
        │ │ │ numeric_value_mask (torch.bool):
        │ │ │ │ [[ True, False,  True, False],
        │ │ │ │  [False,  True, False,  True]]
        │ │
        │ │ Static:
        │ │ │ static_code (torch.int64):
        │ │ │ │ [[1],
        │ │ │ │  [5]]
        │ │ │ static_numeric_value (torch.float32):
        │ │ │ │ [[1.],
        │ │ │ │  [0.]]
        │ │ │ static_numeric_value_mask (torch.bool):
        │ │ │ │ [[True],
        │ │ │ │  [True]]
        │ │
        │ │ Labels:
        │ │ │ boolean_value (torch.bool):
        │ │ │ │ [ True, False]
        >>> print(MEDSTorchBatch(
        ...     time_delta_days=torch.tensor([[0.0, 1.0, 0.0, 0.0, 2.1], [0.0, 4.0, 0.0, 0.0, 0.0]]),
        ...     code=torch.tensor([[1, 1, 2, 3, 3], [5, 5, 6, 0, 0]]),
        ...     numeric_value=torch.tensor([[1.0, 1.0, 0.0, -3.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0]]),
        ...     numeric_value_mask=torch.tensor(
        ...         [[True, True, False, True, False], [True, False, True, False, True]]
        ...     ),
        ...     static_mask=torch.tensor(
        ...         [[True, False, False, False, False], [True, False, False, False, False]]
        ...     ),
        ...     boolean_value=torch.tensor([True, False]),
        ... ))
        MEDSTorchBatch:
        │ Mode: Subject-Measurement (SM)
        │ Static data? ✓ (prepended)
        │ Labels? ✓

        │ Shape:
        │ │ Batch size: 2
        │ │ Sequence length (static + dynamic): 5
        │ │
        │ │ All [static; dynamic] data: (2, 5)
        │ │ Labels: torch.Size([2])

        │ Data:
        │ │ [Static; Dynamic]:
        │ │ │ time_delta_days (torch.float32):
        │ │ │ │ [[0.00, 1.00,  ..., 0.00, 2.10],
        │ │ │ │  [0.00, 4.00,  ..., 0.00, 0.00]]
        │ │ │ code (torch.int64):
        │ │ │ │ [[1, 1, ..., 3, 3],
        │ │ │ │  [5, 5, ..., 0, 0]]
        │ │ │ numeric_value (torch.float32):
        │ │ │ │ [[ 1., 1.,  ..., -3.,  0.],
        │ │ │ │  [ 0., 0.,  ...,  0.,  0.]]
        │ │ │ numeric_value_mask (torch.bool):
        │ │ │ │ [[ True,  True, ...,  True, False],
        │ │ │ │  [ True, False, ..., False,  True]]
        │ │ │ static_mask (torch.bool):
        │ │ │ │ [[ True, False, ..., False, False],
        │ │ │ │  [ True, False, ..., False, False]]
        │ │
        │ │ Labels:
        │ │ │ boolean_value (torch.bool):
        │ │ │ │ [ True, False]
    """

    lines = [f"{self.__class__.__name__}:"]

    torch.set_printoptions(precision=2, threshold=5, edgeitems=2)

    lines.extend([f"{BRANCH}{line}" for line in self.__mode_str_lines()])
    lines.append(BRANCH)
    lines.extend([f"{BRANCH}{line}" for line in self.__shape_str_lines()])
    lines.append(BRANCH)
    lines.extend([f"{BRANCH}{line}" for line in self.__data_str_lines()])

    torch.set_printoptions(profile="default")

    lines = [line.rstrip() for line in lines]

    return "\n".join(lines)

__str_tensor_list(header, tensors)

Gets string representation lines for the requested tensors.

Source code in meds_torchdata/types.py
def __str_tensor_list(self, header: str, tensors: list[str]) -> list[str]:
    """Gets string representation lines for the requested tensors."""
    out = [f"{header}:"]
    for tensor_n in tensors:
        tensor = getattr(self, tensor_n)
        if tensor is None:
            continue

        out.append(f"{BRANCH}{tensor_n} ({tensor.dtype}):")
        tensor_str = self.__str_tensor_val(tensor)
        out.extend(textwrap.indent(tensor_str, BRANCH + BRANCH).splitlines())

    return out

__str_tensor_val(tensor) staticmethod

Strips the tensor( prefix, ) suffix, leading/trailing , and newlines.

Source code in meds_torchdata/types.py
@staticmethod
def __str_tensor_val(tensor: torch.Tensor) -> str:
    """Strips the `tensor(` prefix, `)` suffix, leading/trailing , and newlines."""

    tensor_str = str(tensor).replace("tensor(", "       ").replace(")", "")
    tensor_str = "\n".join([x for x in tensor_str.splitlines() if x.strip()])
    tensor_str = textwrap.dedent(tensor_str).strip()
    return tensor_str

items()

Get the items of the batch.

Source code in meds_torchdata/types.py
def items(self) -> Generator[tuple[str, torch.Tensor], None, None]:
    """Get the items of the batch."""
    yield from zip(self.keys(), self.values(), strict=True)

keys()

Get the keys of the batch.

Source code in meds_torchdata/types.py
def keys(self) -> Generator[str, None, None]:
    """Get the keys of the batch."""
    for field in fields(self):
        if getattr(self, field.name) is not None:
            yield field.name

values()

Get the values of the batch.

Source code in meds_torchdata/types.py
def values(self) -> Generator[torch.Tensor, None, None]:
    """Get the values of the batch."""
    for key in self.keys():
        yield self[key]

PaddingSide

Bases: StrEnum

An enumeration of the possible padding sides for the dataset (either left or right).

Attributes:

Name Type Description
LEFT

Pad the sequence on the left side. This is useful for autoregressive generation.

RIGHT

Pad the sequence on the right side. This is more typical and used in general model training.

Source code in meds_torchdata/types.py
class PaddingSide(StrEnum):
    """An enumeration of the possible padding sides for the dataset (either left or right).

    Attributes:
        LEFT: Pad the sequence on the left side. This is useful for autoregressive generation.
        RIGHT: Pad the sequence on the right side. This is more typical and used in general model training.
    """

    LEFT = "left"
    RIGHT = "right"

StaticData

Bases: NamedTuple

Simple data structure to hold static data, capturing both codes and numeric values.

As a NamedTuple, can be accessed both by index (e.g. data[0]) and by attribute (e.g. data.code).

Attributes:

Name Type Description
code list[int]

List of integer codes.

numeric_value list[float | None]

List of float or None numeric values.

Source code in meds_torchdata/types.py
class StaticData(NamedTuple):
    """Simple data structure to hold static data, capturing both codes and numeric values.

    As a `NamedTuple`, can be accessed both by index (e.g. `data[0]`) and by attribute (e.g. `data.code`).

    Attributes:
        code: List of integer codes.
        numeric_value: List of float or None numeric values.
    """

    code: list[int]
    numeric_value: list[float | None]

    def to_JNRT(
        self,
        batch_mode: BatchMode,
        schema: dict | None = None,
        keys: set[str] | None = None,
    ) -> JointNestedRaggedTensorDict:
        """Converts the static data into a JointNestedRaggedTensorDict representation.

        Args:
            batch_mode: The batch mode to use for the conversion (either SEM or SM).
            schema: The schema to use for the conversion.
            keys: Optional filter restricting which top-level keys (`code`, `numeric_value`,
                `time_delta_days`) appear in the output. Used so the static JNRT keyset can
                match a dynamic JNRT loaded with `JointNestedRaggedTensorDict(..., keys=...)`
                before `concatenate`. When `None` (default) all keys are emitted.

        Returns:
            A JointNestedRaggedTensorDict representation of the static data, including the code, numeric
            value, and a time delta of NaN, at the appropriate dimensionality for the given batch mode.

        Raises:
            ValueError: If the batch mode is not SEM or SM.

        Examples:
            >>> from nested_ragged_tensors.ragged_numpy import pprint_dense
            >>> static_data = StaticData(code=[1, 2, 3], numeric_value=[1.0, 2.0, 3.0])
            >>> pprint_dense(static_data.to_JNRT(BatchMode.SEM).to_dense())
            time_delta_days
            [nan]
            .
            ---
            .
            dim1/mask
            [[ True  True  True]]
            .
            code
            [[1 2 3]]
            .
            numeric_value
            [[1. 2. 3.]]
            >>> pprint_dense(static_data.to_JNRT(BatchMode.SM).to_dense())
            code
            [1 2 3]
            .
            numeric_value
            [1. 2. 3.]
            .
            time_delta_days
            [nan nan nan]

        You can also pass a schema to control the types:

            >>> with_schema = static_data.to_JNRT(BatchMode.SM, {"code": float, "numeric_value": int})
            >>> pprint_dense(with_schema.to_dense())
            code
            [1. 2. 3.]
            .
            numeric_value
            [1 2 3]
            .
            time_delta_days
            [nan nan nan]

        `keys=` drops the unlisted top-level keys from the output — used to keep the
        static JNRT's keyset aligned with a dynamic JNRT that was loaded via NRT's
        `keys=` subset before `concatenate`:

            >>> pprint_dense(static_data.to_JNRT(BatchMode.SM, keys={"code"}).to_dense())
            code
            [1 2 3]
            >>> pprint_dense(static_data.to_JNRT(BatchMode.SM, keys={"code", "time_delta_days"}).to_dense())
            code
            [1 2 3]
            .
            time_delta_days
            [nan nan nan]

        Passing an invalid batch mode will raise an error:

            >>> pprint_dense(static_data.to_JNRT("foobar").to_dense())
            Traceback (most recent call last):
                ...
            ValueError: Invalid batch mode foobar!
        """

        match batch_mode:
            case BatchMode.SEM:
                static_dict = {
                    "time_delta_days": [np.nan],
                    "code": [self.code],
                    "numeric_value": [self.numeric_value],
                }
            case BatchMode.SM:
                static_dict = {
                    "time_delta_days": [np.nan for _ in range(len(self.code))],
                    "code": self.code,
                    "numeric_value": self.numeric_value,
                }
            case _:
                raise ValueError(f"Invalid batch mode {batch_mode}!")

        if keys is not None:
            static_dict = {k: v for k, v in static_dict.items() if k in keys}

        return JointNestedRaggedTensorDict(static_dict, schema=schema)

to_JNRT(batch_mode, schema=None, keys=None)

Converts the static data into a JointNestedRaggedTensorDict representation.

Parameters:

Name Type Description Default
batch_mode BatchMode

The batch mode to use for the conversion (either SEM or SM).

required
schema dict | None

The schema to use for the conversion.

None
keys set[str] | None

Optional filter restricting which top-level keys (code, numeric_value, time_delta_days) appear in the output. Used so the static JNRT keyset can match a dynamic JNRT loaded with JointNestedRaggedTensorDict(..., keys=...) before concatenate. When None (default) all keys are emitted.

None

Returns:

Type Description
JointNestedRaggedTensorDict

A JointNestedRaggedTensorDict representation of the static data, including the code, numeric

JointNestedRaggedTensorDict

value, and a time delta of NaN, at the appropriate dimensionality for the given batch mode.

Raises:

Type Description
ValueError

If the batch mode is not SEM or SM.

Examples:

>>> from nested_ragged_tensors.ragged_numpy import pprint_dense
>>> static_data = StaticData(code=[1, 2, 3], numeric_value=[1.0, 2.0, 3.0])
>>> pprint_dense(static_data.to_JNRT(BatchMode.SEM).to_dense())
time_delta_days
[nan]
.
---
.
dim1/mask
[[ True  True  True]]
.
code
[[1 2 3]]
.
numeric_value
[[1. 2. 3.]]
>>> pprint_dense(static_data.to_JNRT(BatchMode.SM).to_dense())
code
[1 2 3]
.
numeric_value
[1. 2. 3.]
.
time_delta_days
[nan nan nan]

You can also pass a schema to control the types:

>>> with_schema = static_data.to_JNRT(BatchMode.SM, {"code": float, "numeric_value": int})
>>> pprint_dense(with_schema.to_dense())
code
[1. 2. 3.]
.
numeric_value
[1 2 3]
.
time_delta_days
[nan nan nan]

keys= drops the unlisted top-level keys from the output — used to keep the static JNRT’s keyset aligned with a dynamic JNRT that was loaded via NRT’s keys= subset before concatenate:

>>> pprint_dense(static_data.to_JNRT(BatchMode.SM, keys={"code"}).to_dense())
code
[1 2 3]
>>> pprint_dense(static_data.to_JNRT(BatchMode.SM, keys={"code", "time_delta_days"}).to_dense())
code
[1 2 3]
.
time_delta_days
[nan nan nan]

Passing an invalid batch mode will raise an error:

>>> pprint_dense(static_data.to_JNRT("foobar").to_dense())
Traceback (most recent call last):
    ...
ValueError: Invalid batch mode foobar!
Source code in meds_torchdata/types.py
def to_JNRT(
    self,
    batch_mode: BatchMode,
    schema: dict | None = None,
    keys: set[str] | None = None,
) -> JointNestedRaggedTensorDict:
    """Converts the static data into a JointNestedRaggedTensorDict representation.

    Args:
        batch_mode: The batch mode to use for the conversion (either SEM or SM).
        schema: The schema to use for the conversion.
        keys: Optional filter restricting which top-level keys (`code`, `numeric_value`,
            `time_delta_days`) appear in the output. Used so the static JNRT keyset can
            match a dynamic JNRT loaded with `JointNestedRaggedTensorDict(..., keys=...)`
            before `concatenate`. When `None` (default) all keys are emitted.

    Returns:
        A JointNestedRaggedTensorDict representation of the static data, including the code, numeric
        value, and a time delta of NaN, at the appropriate dimensionality for the given batch mode.

    Raises:
        ValueError: If the batch mode is not SEM or SM.

    Examples:
        >>> from nested_ragged_tensors.ragged_numpy import pprint_dense
        >>> static_data = StaticData(code=[1, 2, 3], numeric_value=[1.0, 2.0, 3.0])
        >>> pprint_dense(static_data.to_JNRT(BatchMode.SEM).to_dense())
        time_delta_days
        [nan]
        .
        ---
        .
        dim1/mask
        [[ True  True  True]]
        .
        code
        [[1 2 3]]
        .
        numeric_value
        [[1. 2. 3.]]
        >>> pprint_dense(static_data.to_JNRT(BatchMode.SM).to_dense())
        code
        [1 2 3]
        .
        numeric_value
        [1. 2. 3.]
        .
        time_delta_days
        [nan nan nan]

    You can also pass a schema to control the types:

        >>> with_schema = static_data.to_JNRT(BatchMode.SM, {"code": float, "numeric_value": int})
        >>> pprint_dense(with_schema.to_dense())
        code
        [1. 2. 3.]
        .
        numeric_value
        [1 2 3]
        .
        time_delta_days
        [nan nan nan]

    `keys=` drops the unlisted top-level keys from the output — used to keep the
    static JNRT's keyset aligned with a dynamic JNRT that was loaded via NRT's
    `keys=` subset before `concatenate`:

        >>> pprint_dense(static_data.to_JNRT(BatchMode.SM, keys={"code"}).to_dense())
        code
        [1 2 3]
        >>> pprint_dense(static_data.to_JNRT(BatchMode.SM, keys={"code", "time_delta_days"}).to_dense())
        code
        [1 2 3]
        .
        time_delta_days
        [nan nan nan]

    Passing an invalid batch mode will raise an error:

        >>> pprint_dense(static_data.to_JNRT("foobar").to_dense())
        Traceback (most recent call last):
            ...
        ValueError: Invalid batch mode foobar!
    """

    match batch_mode:
        case BatchMode.SEM:
            static_dict = {
                "time_delta_days": [np.nan],
                "code": [self.code],
                "numeric_value": [self.numeric_value],
            }
        case BatchMode.SM:
            static_dict = {
                "time_delta_days": [np.nan for _ in range(len(self.code))],
                "code": self.code,
                "numeric_value": self.numeric_value,
            }
        case _:
            raise ValueError(f"Invalid batch mode {batch_mode}!")

    if keys is not None:
        static_dict = {k: v for k, v in static_dict.items() if k in keys}

    return JointNestedRaggedTensorDict(static_dict, schema=schema)

StaticInclusionMode

Bases: StrEnum

An enumeration of the possible vehicles to include static measurements.

Attributes:

Name Type Description
PREPEND

Prepend the static measurements to the beginning of the sequence of dynamic data. They will be treated as a standalone event in the sequence with a time delta of 0 days.

INCLUDE

Include the static measurements as a separate output key in each batch.

OMIT

Omit the static measurements entirely.

Source code in meds_torchdata/types.py
class StaticInclusionMode(StrEnum):
    """An enumeration of the possible vehicles to include static measurements.

    Attributes:
        PREPEND: Prepend the static measurements to the beginning of the sequence of dynamic data. They will
                 be treated as a standalone event in the sequence with a time delta of 0 days.
        INCLUDE: Include the static measurements as a separate output key in each batch.
        OMIT: Omit the static measurements entirely.
    """

    PREPEND = "prepend"
    INCLUDE = "include"
    OMIT = "omit"

SubsequenceSamplingStrategy

Bases: StrEnum

An enumeration of the possible subsequence sampling strategies for the dataset.

Attributes:

Name Type Description
RANDOM

Randomly sample a subsequence from the full sequence. Start offsets are drawn uniformly over [0, seq_len - max_seq_len], which yields a trapezoidal per-event inclusion distribution: events near the middle of the sequence appear in roughly max_seq_len times as many windows as events at the boundaries. See issue #67.

BALANCED_RANDOM

Randomly sample a subsequence such that every event in the sequence has equal probability of being included in the sampled window. This is done by drawing a (possibly negative) start offset uniformly from {-(max_seq_len - 1), ..., seq_len - 1} and clipping the resulting window to the sequence. Windows near the boundaries are therefore shorter than max_seq_len; the collator handles padding downstream. The per-event inclusion probability is exactly max_seq_len / (seq_len + max_seq_len - 1) for every position, removing the structural boundary bias of RANDOM. See issue #67.

TO_END

Sample a subsequence from the end of the full sequence. Note this starts at the last element and moves back.

FROM_START

Sample a subsequence from the start of the full sequence.

STEP_THROUGH

Deterministically walk through every permitted subsequence of the full sequence in order, stepping by MEDSTorchDataConfig.step_through_stride (or equivalently step_through_overlap) elements. The stride / overlap is expressed in the same unit as max_seq_lenevents in BatchMode.SEM and measurements in BatchMode.SM — so windows line up with what the user specified regardless of mode. Unlike the other strategies, this expands one subject into multiple dataset elements (one per window), so len(dataset) grows for subjects with longer sequences. MEDSPytorchDataset.__init__ writes the per-window end-event into self.index directly, and in SM mode additionally records the measurement-level window end in a parallel self.step_through_meas_ends array. subsample_st_offset delegates to TO_END (take the last max_seq_len elements of the loaded prefix), so step-through is “modify the index, run TO_END per entry”. In SM mode, the measurement-level end is passed explicitly through MEDSTorchDataConfig.process_dynamic_data via its explicit_end parameter, which lets the window terminate mid-event and preserves the exact measurement-level coverage the user asked for. A warning with the observed expansion stats is logged on startup; see MEDSTorchDataConfig.include_subject_window_counts_in_batch for the corresponding loss-reweighting escape hatch.

Performance note for SM mode: each per-window load reads events [0, enclosing_event_end) from disk, where enclosing_event_end is the smallest event index whose cumulative measurement count covers the target window’s measurement end. If the dataset has very fat events (many measurements per timestamp) and max_seq_len is small relative to the per-event measurement count, consecutive windows can share the same enclosing event and re-read the same prefix — so SM-mode step-through I/O does not scale purely with max_seq_len in that regime. For cohorts with thin events (≤ a few measurements per timestamp) this is a non-issue. Switching to BatchMode.SEM avoids the re-read entirely because the window is already event-aligned.

Methods:

Name Description
subsample_st_offset

Subsample starting offset based on maximum sequence length and sampling strategy. This method can be used on instances (e.g., SubsequenceSamplingStrategy.RANDOM.subsample_st_offset) but is most often used as a static class level method for maximal clarity.

Source code in meds_torchdata/types.py
class SubsequenceSamplingStrategy(StrEnum):
    """An enumeration of the possible subsequence sampling strategies for the dataset.

    Attributes:
        RANDOM: Randomly sample a subsequence from the full sequence. Start offsets are drawn
            uniformly over `[0, seq_len - max_seq_len]`, which yields a trapezoidal per-event
            inclusion distribution: events near the middle of the sequence appear in roughly
            `max_seq_len` times as many windows as events at the boundaries. See issue #67.
        BALANCED_RANDOM: Randomly sample a subsequence such that every event in the sequence has
            equal probability of being included in the sampled window. This is done by drawing a
            (possibly negative) start offset uniformly from `{-(max_seq_len - 1), ..., seq_len - 1}`
            and clipping the resulting window to the sequence. Windows near the boundaries are
            therefore shorter than `max_seq_len`; the collator handles padding downstream. The
            per-event inclusion probability is exactly `max_seq_len / (seq_len + max_seq_len - 1)`
            for every position, removing the structural boundary bias of `RANDOM`. See issue #67.
        TO_END: Sample a subsequence from the end of the full sequence.
            Note this starts at the last element and moves back.
        FROM_START: Sample a subsequence from the start of the full sequence.
        STEP_THROUGH: Deterministically walk through every permitted subsequence of the full
            sequence in order, stepping by `MEDSTorchDataConfig.step_through_stride` (or
            equivalently `step_through_overlap`) elements. The stride / overlap is expressed
            in the same unit as `max_seq_len` — **events** in `BatchMode.SEM` and
            **measurements** in `BatchMode.SM` — so windows line up with what the user
            specified regardless of mode. Unlike the other strategies, this expands one
            subject into multiple dataset elements (one per window), so `len(dataset)` grows
            for subjects with longer sequences. `MEDSPytorchDataset.__init__` writes the
            per-window end-event into `self.index` directly, and in SM mode additionally
            records the measurement-level window end in a parallel
            `self.step_through_meas_ends` array. `subsample_st_offset` delegates to `TO_END`
            (take the last `max_seq_len` elements of the loaded prefix), so step-through is
            "modify the index, run `TO_END` per entry". In SM mode, the measurement-level end
            is passed explicitly through `MEDSTorchDataConfig.process_dynamic_data` via its
            `explicit_end` parameter, which lets the window terminate mid-event and preserves
            the exact measurement-level coverage the user asked for. A warning with the
            observed expansion stats is logged on startup; see
            `MEDSTorchDataConfig.include_subject_window_counts_in_batch` for the
            corresponding loss-reweighting escape hatch.

            Performance note for SM mode: each per-window load reads events
            `[0, enclosing_event_end)` from disk, where `enclosing_event_end` is the smallest
            event index whose cumulative measurement count covers the target window's
            measurement end. If the dataset has very fat events (many measurements per
            timestamp) and `max_seq_len` is small relative to the per-event measurement
            count, consecutive windows can share the same enclosing event and re-read the
            same prefix — so SM-mode step-through I/O does **not** scale purely with
            `max_seq_len` in that regime. For cohorts with thin events (≤ a few measurements
            per timestamp) this is a non-issue. Switching to `BatchMode.SEM` avoids the
            re-read entirely because the window is already event-aligned.

    Methods:
        subsample_st_offset: Subsample starting offset based on maximum sequence length and sampling strategy.
            This method can be used on instances
            (e.g., SubsequenceSamplingStrategy.RANDOM.subsample_st_offset) but is most often used as a static
            class level method for maximal clarity.
    """

    RANDOM = "random"
    BALANCED_RANDOM = "balanced_random"
    TO_END = "to_end"
    FROM_START = "from_start"
    STEP_THROUGH = "step_through"

    def subsample_st_offset(
        self,
        seq_len: int,
        max_seq_len: int,
        rng: SEED_OR_RNG = None,
    ) -> int:
        """Subsample starting offset based on maximum sequence length and sampling strategy.

        The method is an ordinary instance method on the enum; callers typically invoke it via the
        class-level sugar `SubsequenceSamplingStrategy.subsample_st_offset(strategy, ...)`, which
        binds `strategy` (one of `RANDOM`, `BALANCED_RANDOM`, `TO_END`, `FROM_START`,
        `STEP_THROUGH`, or the equivalent string value) to `self` and forwards the rest.

        Args:
            seq_len: Length of the sequence.
            max_seq_len: Maximum allowed sequence length.
            rng: Random number generator for random sampling. If None, a new generator is created. If an
                integer, a new generator is created with that seed.

        Returns:
            The (integral) start offset within the sequence based on the sampling strategy. Always an
            `int`: when `seq_len <= max_seq_len` (no sub-sampling needed) every strategy returns `0`,
            so callers can slice `data[st : st + max_seq_len]` unconditionally — the `min(seq_len, ...)`
            clamp in `MEDSTorchDataConfig.process_dynamic_data` handles the short-sequence case. See
            issue #71 for the history behind dropping the previous `None`-on-fit sentinel.

        Examples:
            >>> SubsequenceSamplingStrategy.subsample_st_offset("from_start", 10, 5)
            0
            >>> SubsequenceSamplingStrategy.subsample_st_offset(SubsequenceSamplingStrategy.TO_END, 10, 5)
            5
            >>> SubsequenceSamplingStrategy.subsample_st_offset("random", 10, 5, rng=1)
            2
            >>> SubsequenceSamplingStrategy.RANDOM.subsample_st_offset(10, 10)
            0

            `STEP_THROUGH` delegates to `TO_END`: each index entry already points at the
            "load me up to this event" endpoint, so the sampler just takes the last
            `max_seq_len` elements of the loaded prefix. `MEDSPytorchDataset.__init__`
            inserts the extra per-window index entries before this method is ever called.

            >>> SubsequenceSamplingStrategy.STEP_THROUGH.subsample_st_offset(10, 5)
            5
            >>> SubsequenceSamplingStrategy.STEP_THROUGH.subsample_st_offset(5, 10)
            0

            The random sampler must be able to place the window flush against the end of the
            sequence (i.e. sample `st = seq_len - max_seq_len`, so that the last event at index
            `seq_len - 1` is included). Prior to the fix for issue #67 this was off-by-one and
            the last event was never reachable:

            >>> possible_st = {
            ...     SubsequenceSamplingStrategy.subsample_st_offset("random", 10, 5, rng=s)
            ...     for s in range(1000)
            ... }
            >>> sorted(possible_st)
            [0, 1, 2, 3, 4, 5]

            `BALANCED_RANDOM` draws a signed start offset uniformly from
            `{-(max_seq_len - 1), ..., seq_len - 1}`, which is how we get a flat per-event
            inclusion distribution (see issue #67). The full support of `st` is reachable:

            >>> possible_st = {
            ...     SubsequenceSamplingStrategy.subsample_st_offset("balanced_random", 10, 5, rng=s)
            ...     for s in range(5000)
            ... }
            >>> sorted(possible_st)
            [-4, -3, -2, -1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9]

            `BALANCED_RANDOM` returns `0` (no subsampling) when the sequence already fits:

            >>> SubsequenceSamplingStrategy.BALANCED_RANDOM.subsample_st_offset(5, 10)
            0

            >>> SubsequenceSamplingStrategy.subsample_st_offset("foo", 10, 5)
            Traceback (most recent call last):
                ...
            ValueError: Invalid subsequence sampling strategy foo!
        """

        if seq_len <= max_seq_len:
            # No sub-sampling needed — caller will slice `data[0 : min(seq_len, max_seq_len)]`
            # and get the whole sequence. Return `0` so the signature stays monomorphic.
            return 0

        match self:
            case SubsequenceSamplingStrategy.RANDOM:
                # NOTE: `choice(n)` draws from `[0, n)`, so to let `st` reach `seq_len - max_seq_len`
                # (and thus let `data[st:st + max_seq_len]` include the final event at index
                # `seq_len - 1`) we must pass `seq_len - max_seq_len + 1`. See issue #67.
                return int(resolve_rng(rng).choice(seq_len - max_seq_len + 1))
            case SubsequenceSamplingStrategy.BALANCED_RANDOM:
                # Draw `st` uniformly from `{-(max_seq_len - 1), ..., seq_len - 1}` so that every
                # event at index `i` is contained in exactly `max_seq_len` of the `seq_len +
                # max_seq_len - 1` equally-likely windows, giving a uniform per-event inclusion
                # probability of `max_seq_len / (seq_len + max_seq_len - 1)`. Callers must handle
                # the negative/overhanging offset by clipping the slice to `[0, seq_len)`.
                return int(resolve_rng(rng).integers(-(max_seq_len - 1), seq_len))
            case SubsequenceSamplingStrategy.TO_END | SubsequenceSamplingStrategy.STEP_THROUGH:
                # STEP_THROUGH is just TO_END over a modified index: the dataset constructor
                # has already inserted one `(subject_id, window_end_event)` entry per window,
                # so by the time this runs the loaded prefix *is* the step-through window and
                # TO_END takes its last `max_seq_len` elements.
                return seq_len - max_seq_len
            case SubsequenceSamplingStrategy.FROM_START:
                return 0
            case _:
                raise ValueError(f"Invalid subsequence sampling strategy {self}!")

subsample_st_offset(seq_len, max_seq_len, rng=None)

Subsample starting offset based on maximum sequence length and sampling strategy.

The method is an ordinary instance method on the enum; callers typically invoke it via the class-level sugar SubsequenceSamplingStrategy.subsample_st_offset(strategy, ...), which binds strategy (one of RANDOM, BALANCED_RANDOM, TO_END, FROM_START, STEP_THROUGH, or the equivalent string value) to self and forwards the rest.

Parameters:

Name Type Description Default
seq_len int

Length of the sequence.

required
max_seq_len int

Maximum allowed sequence length.

required
rng SEED_OR_RNG

Random number generator for random sampling. If None, a new generator is created. If an integer, a new generator is created with that seed.

None

Returns:

Type Description
int

The (integral) start offset within the sequence based on the sampling strategy. Always an

int

int: when seq_len <= max_seq_len (no sub-sampling needed) every strategy returns 0,

int

so callers can slice data[st : st + max_seq_len] unconditionally — the min(seq_len, ...)

int

clamp in MEDSTorchDataConfig.process_dynamic_data handles the short-sequence case. See

int

issue #71 for the history behind dropping the previous None-on-fit sentinel.

Examples:

>>> SubsequenceSamplingStrategy.subsample_st_offset("from_start", 10, 5)
0
>>> SubsequenceSamplingStrategy.subsample_st_offset(SubsequenceSamplingStrategy.TO_END, 10, 5)
5
>>> SubsequenceSamplingStrategy.subsample_st_offset("random", 10, 5, rng=1)
2
>>> SubsequenceSamplingStrategy.RANDOM.subsample_st_offset(10, 10)
0

STEP_THROUGH delegates to TO_END: each index entry already points at the “load me up to this event” endpoint, so the sampler just takes the last max_seq_len elements of the loaded prefix. MEDSPytorchDataset.__init__ inserts the extra per-window index entries before this method is ever called.

>>> SubsequenceSamplingStrategy.STEP_THROUGH.subsample_st_offset(10, 5)
5
>>> SubsequenceSamplingStrategy.STEP_THROUGH.subsample_st_offset(5, 10)
0

The random sampler must be able to place the window flush against the end of the sequence (i.e. sample st = seq_len - max_seq_len, so that the last event at index seq_len - 1 is included). Prior to the fix for issue #67 this was off-by-one and the last event was never reachable:

>>> possible_st = {
...     SubsequenceSamplingStrategy.subsample_st_offset("random", 10, 5, rng=s)
...     for s in range(1000)
... }
>>> sorted(possible_st)
[0, 1, 2, 3, 4, 5]

BALANCED_RANDOM draws a signed start offset uniformly from {-(max_seq_len - 1), ..., seq_len - 1}, which is how we get a flat per-event inclusion distribution (see issue #67). The full support of st is reachable:

>>> possible_st = {
...     SubsequenceSamplingStrategy.subsample_st_offset("balanced_random", 10, 5, rng=s)
...     for s in range(5000)
... }
>>> sorted(possible_st)
[-4, -3, -2, -1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9]

BALANCED_RANDOM returns 0 (no subsampling) when the sequence already fits:

>>> SubsequenceSamplingStrategy.BALANCED_RANDOM.subsample_st_offset(5, 10)
0
>>> SubsequenceSamplingStrategy.subsample_st_offset("foo", 10, 5)
Traceback (most recent call last):
    ...
ValueError: Invalid subsequence sampling strategy foo!
Source code in meds_torchdata/types.py
def subsample_st_offset(
    self,
    seq_len: int,
    max_seq_len: int,
    rng: SEED_OR_RNG = None,
) -> int:
    """Subsample starting offset based on maximum sequence length and sampling strategy.

    The method is an ordinary instance method on the enum; callers typically invoke it via the
    class-level sugar `SubsequenceSamplingStrategy.subsample_st_offset(strategy, ...)`, which
    binds `strategy` (one of `RANDOM`, `BALANCED_RANDOM`, `TO_END`, `FROM_START`,
    `STEP_THROUGH`, or the equivalent string value) to `self` and forwards the rest.

    Args:
        seq_len: Length of the sequence.
        max_seq_len: Maximum allowed sequence length.
        rng: Random number generator for random sampling. If None, a new generator is created. If an
            integer, a new generator is created with that seed.

    Returns:
        The (integral) start offset within the sequence based on the sampling strategy. Always an
        `int`: when `seq_len <= max_seq_len` (no sub-sampling needed) every strategy returns `0`,
        so callers can slice `data[st : st + max_seq_len]` unconditionally — the `min(seq_len, ...)`
        clamp in `MEDSTorchDataConfig.process_dynamic_data` handles the short-sequence case. See
        issue #71 for the history behind dropping the previous `None`-on-fit sentinel.

    Examples:
        >>> SubsequenceSamplingStrategy.subsample_st_offset("from_start", 10, 5)
        0
        >>> SubsequenceSamplingStrategy.subsample_st_offset(SubsequenceSamplingStrategy.TO_END, 10, 5)
        5
        >>> SubsequenceSamplingStrategy.subsample_st_offset("random", 10, 5, rng=1)
        2
        >>> SubsequenceSamplingStrategy.RANDOM.subsample_st_offset(10, 10)
        0

        `STEP_THROUGH` delegates to `TO_END`: each index entry already points at the
        "load me up to this event" endpoint, so the sampler just takes the last
        `max_seq_len` elements of the loaded prefix. `MEDSPytorchDataset.__init__`
        inserts the extra per-window index entries before this method is ever called.

        >>> SubsequenceSamplingStrategy.STEP_THROUGH.subsample_st_offset(10, 5)
        5
        >>> SubsequenceSamplingStrategy.STEP_THROUGH.subsample_st_offset(5, 10)
        0

        The random sampler must be able to place the window flush against the end of the
        sequence (i.e. sample `st = seq_len - max_seq_len`, so that the last event at index
        `seq_len - 1` is included). Prior to the fix for issue #67 this was off-by-one and
        the last event was never reachable:

        >>> possible_st = {
        ...     SubsequenceSamplingStrategy.subsample_st_offset("random", 10, 5, rng=s)
        ...     for s in range(1000)
        ... }
        >>> sorted(possible_st)
        [0, 1, 2, 3, 4, 5]

        `BALANCED_RANDOM` draws a signed start offset uniformly from
        `{-(max_seq_len - 1), ..., seq_len - 1}`, which is how we get a flat per-event
        inclusion distribution (see issue #67). The full support of `st` is reachable:

        >>> possible_st = {
        ...     SubsequenceSamplingStrategy.subsample_st_offset("balanced_random", 10, 5, rng=s)
        ...     for s in range(5000)
        ... }
        >>> sorted(possible_st)
        [-4, -3, -2, -1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9]

        `BALANCED_RANDOM` returns `0` (no subsampling) when the sequence already fits:

        >>> SubsequenceSamplingStrategy.BALANCED_RANDOM.subsample_st_offset(5, 10)
        0

        >>> SubsequenceSamplingStrategy.subsample_st_offset("foo", 10, 5)
        Traceback (most recent call last):
            ...
        ValueError: Invalid subsequence sampling strategy foo!
    """

    if seq_len <= max_seq_len:
        # No sub-sampling needed — caller will slice `data[0 : min(seq_len, max_seq_len)]`
        # and get the whole sequence. Return `0` so the signature stays monomorphic.
        return 0

    match self:
        case SubsequenceSamplingStrategy.RANDOM:
            # NOTE: `choice(n)` draws from `[0, n)`, so to let `st` reach `seq_len - max_seq_len`
            # (and thus let `data[st:st + max_seq_len]` include the final event at index
            # `seq_len - 1`) we must pass `seq_len - max_seq_len + 1`. See issue #67.
            return int(resolve_rng(rng).choice(seq_len - max_seq_len + 1))
        case SubsequenceSamplingStrategy.BALANCED_RANDOM:
            # Draw `st` uniformly from `{-(max_seq_len - 1), ..., seq_len - 1}` so that every
            # event at index `i` is contained in exactly `max_seq_len` of the `seq_len +
            # max_seq_len - 1` equally-likely windows, giving a uniform per-event inclusion
            # probability of `max_seq_len / (seq_len + max_seq_len - 1)`. Callers must handle
            # the negative/overhanging offset by clipping the slice to `[0, seq_len)`.
            return int(resolve_rng(rng).integers(-(max_seq_len - 1), seq_len))
        case SubsequenceSamplingStrategy.TO_END | SubsequenceSamplingStrategy.STEP_THROUGH:
            # STEP_THROUGH is just TO_END over a modified index: the dataset constructor
            # has already inserted one `(subject_id, window_end_event)` entry per window,
            # so by the time this runs the loaded prefix *is* the step-through window and
            # TO_END takes its last `max_seq_len` elements.
            return seq_len - max_seq_len
        case SubsequenceSamplingStrategy.FROM_START:
            return 0
        case _:
            raise ValueError(f"Invalid subsequence sampling strategy {self}!")