sunlabuiuc
Add eICU LLM synthetic mortality example
Who I am: Mohsin Shah (NetID: mohsins2)
Contribution type: New example / use case of PyHealth.
High-level description: This PR adds an example script examples/eicu_llm_synthetic_mortality.py which demonstrates how to use PyHealth with small tabular EHR-style data and LLM-generated synthetic cohorts from my CS598 project. The example uses 10 hand-crafted ICU features and compares three training regimes for ICU mortality prediction:
Real train → Real test
GPT baseline synthetic train → Real test
GPT privacy-aware synthetic train → Real test
How to run / what files to look at
Example script: examples/eicu_llm_synthetic_mortality.py
The script directly loads the CSVs from my project repo via raw GitHub URLs: https://github.com/mohsinposts/CS598-DLH-LLM-eICU (real_icu_10feat.csv, synthetic_baseline_10feat_clean.csv, synthetic_privacy_10feat_clean.csv)
From the PyHealth repo root, run:
python examples/eicu_llm_synthetic_mortality.py
This prints ROC-AUC, PR-AUC, accuracy, F1, and loss.
