Implement Temporal Pointwise Convolutional (TPC) Model in Length of Stay prediction task

[kchendv]

Contributors

• Kevin Chen (kevinc17@illinois.edu)
• Vismayak Mohanarajan (mohanar2@illinois.edu)
• Ashley Wang (azw6@illinois.edu)

Type of Contribution

• Model
• Task
• Processor

Link to Original Paper

Temporal Pointwise Convolutional Networks for Length of Stay Prediction in the Intensive Care Unit

High-level Description of Implementation

This pull request implements the paper "Temporal Pointwise Convolutional Networks for Length of Stay
Prediction in the Intensive Care Unit". Specifically, it implements

1. Processors for timeseries and pointwise data preprocessing from the MIMIC-IV dataset

• Time-series processor (TPCTimeseriesProcessor)
  • Hourly binning + forward-fill with decay
  • Data scaling based on 5th/95th percentiles
• Static processor (TPCStaticProcessor)
  • One-hot encodes categorical static features
  • Data scaling based on 5th/95th percentiles
• Regression-sequence label processor (RegressionSequenceProcessor)
  • Converts variable-length label sequences to tensors for padding/collation

2. Implementation of the remaining LoS predication as a task

• Generates hourly remaining ICU length-of-stay regression targets
• Starts predictions at 5 hours, truncates sequences to 14 days
• Uses ICU stays (icustays) and filters to adult patients
• Produces a raw payload designed for TPC preprocessing (prefill window + irregular observations from chartevents/labevents)

3. Implementation of the Temporal Pointwise Convolution (TPC) model

• Stacked TPC blocks combining:
  • Per-feature causal temporal convolutions
  • Pointwise mixing across features per time step
  • Dense skip connections
• Trains with masked MSLE over valid (non-padding) timesteps

Files to review

Examples

• Training: examples/mimic4_remaining_los_tpc.py
• Ablation: examples/mimic4_remaining_los_tpc_ablation.py

Core Implementation

• pyhealth/models/tpc.py
• pyhealth/processors/regression_sequence_processor.py
• pyhealth/processors/tpc_static_processor.py
• pyhealth/processors/tpc_timeseries_processor.py
• pyhealth/tasks/remaining_length_of_stay_tpc_mimic4.py

Tests

• tests/core/test_tpc.py

Documentation

• docs/api/models.rst
• docs/api/models/pyhealth.models.TPC.rst
• docs/api/processors.rst
• docs/api/processors/pyhealth.processors.RegressionSequenceProcessor.rst
• docs/api/processors/pyhealth.processors.TPCStaticProcessor.rst
• docs/api/processors/pyhealth.processors.TPCTimeseriesProcessor.rst
• docs/api/tasks.rst
• docs/api/tasks/pyhealth.tasks.RemainingLengthOfStayTPC_MIMIC4.rst

Misc / Config

• pyhealth/datasets/configs/mimic4_ehr.yaml
• pyhealth/__init__.py
• pyhealth/models/__init__.py
• pyhealth/processors/__init__.py
• pyhealth/tasks/__init__.py