Add Shy Model for personalized diagnosis prediction

[yli302]

Contributor

• Name: Yujia Li
• NetID: yujia9
• Email: yujia9@illinois.edu

Type of Contribution

• Option 2: Model

Link to Original Paper

• https://arxiv.org/abs/2502.10689
• Yu, L., Cai, Y., Zhang, M., & Hu, X. (2025). Self-Explaining Hypergraph Neural Networks for Diagnosis Prediction. arXiv preprint arXiv:2502.10689.

Description of Implementation

This Pull Request implements SHy (Self-Explaining Hypergraph Neural Networks), an inherently interpretable model for diagnosis prediction that represents longitudinal Electronic Health Records (EHR) as patient-specific hypergraphs. The implementation is modularized into specialized classes and utility functions to align with PyHealth’s architecture.

It uses a GRU and self-attention mechanism to process the extracted phenotypes into a final diagnosis prediction. This structure allows clinicians to directly intervene by editing the generated phenotypes to refine the model's output.

File List

The following files should be reviewed for this contribution.

Core Implementation Files	Description
pyhealth/models/__init__.py	Exposes the SHy class within the pyhealth.models module to allow for standardized importing across the library
pyhealth/models/shy.py	The main implementation file containing the core logic: SHy class, Inherits from BaseModel; HypergraphMessagePassing class, Implements the UniGIN mechanism; PhenotypeMiner class, Implements the Gumbel-Softmax to extract discrete

Documentation RST files	Description
docs/api/models.rst	Exposes the SHy class within the pyhealth.models module to allow for standardized importing across the library
docs/api/models/pyhealth.models.SHy.rst	The RST documentation file for the SHy

Tests file	Description
tests/core/test_shy.py	Unit tests that uses minimal synthetic tensors to verify model instantiation, forward pass logic, output shapes, and gradient computation