pyhealth.processors.StageNetTensorProcessor

Processor for StageNet numeric inputs with coupled value/time data and forward-fill imputation.

class pyhealth.processors.StageNetTensorProcessor

Bases: TemporalFeatureProcessor

Feature processor for StageNet NUMERIC inputs with coupled value/time data.

This processor handles numeric feature sequences (flat or nested) and applies forward-fill imputation to handle missing values (NaN/None). For categorical codes, use StageNetProcessor instead.

Format: {

“value”: [[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]], # nested numerics “time”: [0.0, 1.5] or None

}

The processor automatically detects: - List of numbers -> flat numeric sequences - List of lists of numbers -> nested numeric sequences (feature vectors)

Imputation Strategy: - Forward-fill: Missing values (NaN/None) are filled with the last observed

value for that feature dimension. If no prior value exists, 0.0 is used.

• Applied per feature dimension independently

Returns:

Tuple of (time_tensor, value_tensor) where time_tensor can be None

Examples

>>> # Case 1: Feature vectors with missing values
>>> processor = StageNetTensorProcessor()
>>> data = {
...     "value": [[1.0, None, 3.0], [None, 5.0, 6.0], [7.0, 8.0, None]],
...     "time": [0.0, 1.5, 3.0]
... }
>>> time, values = processor.process(data)
>>> values  # [[1.0, 0.0, 3.0], [1.0, 5.0, 6.0], [7.0, 8.0, 6.0]]
>>> values.dtype  # torch.float32
>>> time.shape    # (3,)

fit(samples, field)

Determine input structure.

Parameters:

• samples (Iterable[Dict[str, Any]]) – List of sample dictionaries

• key – The key in samples that contains tuple (time, values)

Return type:

None

process(value)

Process tuple format numeric data into tensors.

Applies forward-fill imputation to handle NaN/None values. For each feature dimension, missing values are filled with the last observed value (or 0.0 if no prior value exists).

Parameters:

value (Tuple[Optional[List], List]) – Tuple of (time, values) where values are numerics

Return type:

Tuple[Optional[Tensor], Tensor]

Returns:

Tuple of (time_tensor, value_tensor), time can be None

size()

Return feature dimension.

is_token()

Numeric values are continuous, not discrete tokens.

Return type:

bool

schema()

Output is a tuple of (time_tensor, value_tensor).

Return type:

tuple[str, ...]

dim()

Number of dimensions for each output tensor.

Time tensor is 1D. Value tensor is 1D (flat) or 2D (nested). Must be called after fit().

Return type:

tuple[int, ...]

Returns:

(1, 1) for flat values or (1, 2) for nested values.

spatial()

Whether each dimension of the value tensor is spatial.

Return type:

tuple[bool, ...]

modality()

Continuous lab/vital measurements → NUMERIC modality.

Return type:

ModalityType

value_dim()

Number of numeric features per time-step (used with nn.Linear). Must be called after fit().

Return type:

int

process_temporal(value)

Return dict output for UnifiedMultimodalEmbeddingModel.

Returns:

FloatTensor (T, F), “time”: FloatTensor (T,) or None}

Return type:

{“value”

load(path)

Optional: Load processor state from disk.

Parameters:

path (str) – File path to load processor state from.

Return type:

None

save(path)

Optional: Save processor state to disk.

Parameters:

path (str) – File path to save processor state.

Return type:

None