TM-HGNN

[ACL 2023] Clinical Note Owns its Hierarchy: Multi-Level Hypergraph Neural Networks for Patient-Level Representation Learning

Requirements

• CUDA=11.3
• cuDNN=8.2.0
• python=3.9.12
• pandas=1.4.2
• torch=1.11.0
• torch_geometric=2.1.0

Usage

We follow MIMIC-III Benchmark (Harutyunyan et al.) for preprocess clinical notes. The preprocessed NOTEEVENTS data for in-hospital-mortality should be in data/DATA_RAW/in-hospital-mortality, divided into two folders (train_note and test_note).

Prepare Notes

python graph_construction/prepare_notes/extract_cleaned_notes.py
python graph_construction/prepare_notes/create_hyper_df.py

extract_cleaned_notes.py cleans clinical notes in data/DATA_RAW/in-hospital-mortality, which results in column "Fixed TEXT" in each csv file. Word2vec token embeddings with 100 dimensions are created and saved in data/DATA_RAW/root/word2vec_100.

create_hyper_df.py creates dataframe from data/DATA_RAW/in-hospital-mortality where each row represents each word. The results are stored in data/DATA_PRE/in-hospital-mortality, divided into two folders (train_hyper and test_hyper).

Construct Multi-level Hypergraphs

python graph_construction/prepare_notes/PygNotesGraphDataset.py --split train
python graph_construction/prepare_notes/PygNotesGraphDataset.py --split test

PygNotesGraphDataset.py creates multi-level hypergraphs with cutoff in data/IMDB_HCUT/in-hospital-mortality.

Model Train

python tmhgnn/train.py