[Bounty] ConCare PyHealth 2.0 ver

[joshuasteier]

Pull Request: ConCare Model Update (1.0 → 2.0)

Contributor Information

• Name: Joshua Steier

Contribution Type

• Model Update
• New Model
• Dataset
• Task
• Bug Fix
• Documentation
• Other

High-Level Description

This PR updates the ConCare model from PyHealth 1.0 API to PyHealth 2.0 API.

Paper Reference

• Title: Concare: Personalized clinical feature embedding via capturing the healthcare context
• Authors: Liantao Ma et al.
• Venue: AAAI 2020
• Link: https://ojs.aaai.org/index.php/AAAI/article/view/5428

Changes Made

1. API Migration (1.0 → 2.0):

   • Replaced SampleEHRDataset with SampleDataset
   • Integrated EmbeddingModel for unified embedding handling
   • Removed explicit feature_keys, label_key, mode, use_embedding parameters
   • Simplified constructor to derive feature information from dataset schemas

2. Code Improvements:

   • Added comprehensive Google-style docstrings for all classes and methods
   • Added proper type hints throughout
   • Fixed bug in SingleAttention (removed undefined self.Wd initialization)
   • Improved code documentation with input/output descriptions
   • Added file header with paper reference and description

3. Testing:

   • Added comprehensive unit tests covering:
     • Model initialization (with/without static features)
     • Forward pass validation
     • Backward pass (gradient flow)
     • Embedding extraction
     • Custom hyperparameters
     • Multiclass classification
     • Single feature input

4. Examples:

   • Added example notebook for MIMIC-IV in-hospital mortality prediction
   • Added standalone Python script example

Files to Review

File	Description
pyhealth/models/concare.py	Main model implementation (updated)
tests/models/test_concare.py	Unit tests for ConCare model
examples/concare_mimic4_example.ipynb	Example notebook for MIMIC-IV

How to Test

1. Run Unit Tests

python -m pytest tests/models/test_concare.py -v

2. Run Quick Example (main block)

python pyhealth/models/concare.py

3. Expected Output

{
    'loss': tensor(..., grad_fn=<AddBackward0>),
    'y_prob': tensor([[...], [...]], grad_fn=<SigmoidBackward0>),
    'y_true': tensor([[...], [...]]),
    'logit': tensor([[...], [...]], grad_fn=<AddmmBackward0>)
}

Checklist

• Code follows PEP8 style (88 character line length)
• Code follows Google-style docstrings
• All functions have type hints
• All functions have input/output documentation
• File header includes author, paper title, link, and description
• Unit tests pass
• Example code runs successfully
• Backward pass (loss.backward()) works
• Code is rebased with main branch

Additional Notes

The ConCare model includes:

• Channel-wise GRUs: Separate GRU for each input feature dimension
• Time-aware attention: Captures temporal decay in healthcare context
• Multi-head self-attention: Captures feature interactions
• DeCov loss: Regularization to reduce feature redundancy
• Static feature support: Optional demographic/static features

This update maintains full backward compatibility with the original ConCare functionality while adopting the cleaner 2.0 API patterns.

[jhnwu3]

Will review more of the codebase later (as I find time tomorrow sorry)

For some reason this notebook doesn't show up in GitHub? "The Notebook Does Not Appear to Be Valid JSON"

[joshuasteier]

Thank you, no pressure.
I believe I've fixed this error now in the latest commit.
Much appreciated for your guidance.