How nyumedml/gnn_for_ehr works
This repository implements graph-based EHR representation learning for predicting disease outcomes from ICU records. Researchers preprocess MIMIC or eICU encounter data into labeled feature matrices, then train and evaluate a variational graph neural network and save model checkpoints. The README also describes interpretation through singular-value analysis and representation clustering; that capability is documented, but its implementation is not identifiable from the sampled source, so no runtime wiring is asserted for it.
Main entry points: train.py Training and evaluation loop · preprocess_mimic.py MIMIC preprocessing · preprocess_eicu.py eICU preprocessing · data Processed splits
Components 12
EHR preparation
- MIMIC preprocessing
data/mimic/preprocess_mimic.py - eICU preprocessing
data/eicu/preprocess_eicu.py
Model learning
- Training and evaluation loop
train.py - Batch loading
utils.py - Variational graph model
model.py - Singular-value analysis
plots
Data and results
- Processed splits
data - Validation AUPRC
utils.py - Model checkpoints
Other components
- Researcher
- Raw EHR records
- Disease-outcome predictions
Connections 15
- Researcher → MIMIC preprocessing: runs preprocessing
- Researcher → eICU preprocessing: runs preprocessing
- Raw EHR records → MIMIC preprocessing: reads records
- Raw EHR records → eICU preprocessing: reads records
- MIMIC preprocessing → Processed splits: writes splits
- eICU preprocessing → Processed splits: writes splits
- Researcher → Training and evaluation loop: starts training
- Processed splits → Training and evaluation loop: loads matrices
- Training and evaluation loop → Batch loading: loads batches
- Training and evaluation loop → Variational graph model: trains and evaluates
- Batch loading → Training and evaluation loop: returns batches
- Training and evaluation loop → Validation AUPRC: requests evaluation
- Validation AUPRC → Training and evaluation loop: returns AUPRC
- Training and evaluation loop → Model checkpoints: saves weights
- Variational graph model → Disease-outcome predictions: computes logits