[Logiquo]

This is the second PR in the stack toward unified sequential features, building on #1343. It adds SequenceEmbeddingModel and three prediction baselines for packed SequenceProcessorV2 records: Unified_RNN, Unified_MLP, and Unified_LogisticRegression. It also migrates YAIB tasks and per-step labels to V2 and provides MIMIC-IV YAIB training examples.

Sequence embeddings and models

• SequenceEmbeddingModel embeds multiple token namespaces and projects numeric values at each level, preserving offsets and timestamps. It supports spatial value layouts, optional missing-indicator concatenation, and pretrained token embeddings. Empty numeric batches now reshape correctly, including when spatial axes are retained.
• All three models preserve hierarchical groups and encode bottom-up. Each parent combines its own embedding with its child summary. Empty groups receive zero summaries; a stay without lab orders still participates at the stay level.
• Unified_RNN mean-pools set levels and runs an RNN over sequence/series groups, retaining the last output. Optional time-gap features capture irregular timing. There are no hierarchy or set_pooling selectors.
• Unified_MLP applies independent MLPs to records at every hierarchy level before mean pooling. Static tensors are flattened per sample and passed through their own MLPs. Hidden widths, layer counts, and dropout can be configured per field.
• Unified_LogisticRegression mean-pools record embeddings bottom-up and concatenates raw flattened static vectors before a linear head. It remains a linear head over learned embeddings, rather than classical regression on raw features.
• The MLP and logistic regression models flatten/project spatial values within each record and do not model spatial locality. Their means weight parents equally and do not model temporal order. Existing model classes remain available.

Per-step labels and YAIB integration

• TimeseriesBinaryLabelProcessor and TimeseriesRegressionLabelProcessor provide packed values, times, and batch offsets through SequenceProcessorV2. BaseModel and Trainer support packed per-step losses and inference.
• At label time t, Unified_RNN uses the last available top-level state. Unified_MLP and Unified_LogisticRegression use the cumulative mean of top-level record representations with timestamps <= t. Before the first record, field summaries are zero. Static representations accompany each label record.
• Per-step inputs require a series top level; resampled record i is available at i * step. A top-level timestamp must indicate availability of the complete record, including children. Bidirectional RNNs are rejected for per-step prediction.
• Migrates YAIB AKI, kidney function, length of stay, mortality, and sepsis tasks to sequence_v2 features and per-step labels.
• examples/yaib_mimic4_multimodal_rnn2.py trains Unified_RNN across the five tasks.
• examples/yaib_mimic4_unified_mlp.py follows the same split, preflight, training, checkpoint, and reporting flow. Select --model mlp or --model logistic_regression.

Validation

• Passed 64 tests across the new baselines, sequence embeddings, Unified_RNN, per-step labels, time-series modes, and legacy MLP/logistic regression tests. Coverage includes nested means, empty groups, static-only inputs, missing indicators, pretrained embeddings, CPU/CUDA comparisons, and Trainer integration.
• Passed an additional focused regression test for all-empty numeric embeddings with flattened and retained spatial axes.
• Existing YAIB integration validation passed 65 tests across the RNN, time-series, per-step-label, and YAIB test modules.
• Pyright passed for the new example. CLI help and CPU --check runs passed for both model selections. Full MIMIC-IV training was not run for the new example.
• git diff --check passed.

Following PRs

Update Transformer, CNN/TCN, and EHRMamba to consume sequence_v2 fields, and add a packed_to_padded adapter for dense sequence layers.

[github-actions]

Test Coverage

Total: 74.75% (21241 / 28416 lines)

Modified file	Coverage	Lines
pyhealth/models/__init__.py	100.00%	56 / 56
pyhealth/models/base_model.py	92.81%	129 / 139
pyhealth/models/logistic_regression.py	84.00%	126 / 150
pyhealth/models/mlp.py	76.67%	115 / 150
pyhealth/models/rnn.py	92.41%	341 / 369
pyhealth/models/sequence_embedding.py	98.65%	146 / 148
pyhealth/processors/timeseries_processor.py	87.38%	90 / 103
pyhealth/tasks/yaib_mimic4.py	93.06%	523 / 562
pyhealth/trainer.py	72.69%	197 / 271

Workflow run · 7e4e75a