Source code for the implementation of the paper: "Leveraging patient similarities via graph neural networks to predict phenotypes from temporal data" https://ieeexplore.ieee.org/document/10302556

Replicate benchmark

1. This replication of this source code requires MIMIC-III dataset. You will need to pass the following examination CITI to access the Physionet data repository:
2. For replication of the benchmark you can follow the instructions at:

• https://github.com/YerevaNN/mimic3-benchmarks

Benchmark replication

conda create -n "mimic3" python=3.7.13
conda activate mimic3
pip install -r requirements.txt
wget -r -N -c -np https://physionet.org/files/mimiciii-demo/1.4/ # this will create a physionet folder with the Database csvs
python -m mimic3benchmark.scripts.extract_subjects physionet.org/files/mimiciii-demo/1.4/ data/root
python -m mimic3benchmark.scripts.validate_events data/root/
python -m mimic3benchmark.scripts.extract_episodes_from_subjects data/root/
python -m mimic3benchmark.scripts.split_train_and_test data/root/



python -m mimic3benchmark.scripts.create_in_hospital_mortality data/root/ data/in-hospital-mortality/
python -m mimic3benchmark.scripts.create_decompensation data/root/ data/decompensation/
python -m mimic3benchmark.scripts.create_length_of_stay data/root/ data/length-of-stay/
python -m mimic3benchmark.scripts.create_phenotyping data/root/ data/phenotyping/
python -m mimic3benchmark.scripts.create_multitask data/root/ data/multitask/

python -m mimic3models.split_train_val data/phenotyping

python -um mimic3models.phenotyping.logistic.main --output_dir mimic3models/phenotyping/logistic
# ls data/phenotyping/statistical_features 
# test_X          test_ts         train_X         train_ts        val_X           val_ts
# test_names      test_y          train_names     train_y         val_names       val_y

# The libraries used for these LSTM are highly problematic and require specific version -> torch based equivalent of SCEHR has been used insted
# python -um mimic3models.phenotyping.main --network mimic3models/keras_models/lstm.py --dim 256 --timestep 1.0 --depth 1 --dropout 0.3 --mode train --batch_size 8 --output_dir mimic3models/phenotyping
python -m mimic3benchmark.evaluation.evaluate_phenotyping data/phenotyping/train data/phenotyping/train predictions/phenotyping/logistic train


# We used an alternative pytorch training of LSTM training and eval using SCEHR implementation based on SCEHR
python -m mimic3models.train_lstm --network mimic3models/lstm.py --data data/phenotyping/ --save

GNN training

environment setup

The libraries and python version are more recent than those used for the benchmark. Please raise an issue if you find difficulty in any of the steps below.

create graphs

conda env create -f environment.yaml
python create_homogeneous_graphs.py --edge_strategy trivial --node_embeddings_type stat --folder_name graphs
python create_homogeneous_graphs.py --edge_strategy random --node_embeddings_type stat --folder_name graphs
python create_homogeneous_graphs.py --edge_strategy expert_exact --node_embeddings_type stat --folder_name graphs
python create_homogeneous_graphs.py --edge_strategy knn_graph --node_embeddings_type stat --folder_name graphs
python create_homogeneous_graphs.py --edge_strategy expert_medium --node_embeddings_type stat --folder_name graphs
python create_homogeneous_graphs.py --edge_strategy expert_lenient --node_embeddings_type stat --folder_name graphs
'random', 'expert_exact', 'expert_medium', 'expert_lenient', 'knn_graph', 'trivial' 

expert rules connectivity strategies

python -m gnn__models.connectivity_strategies.expert_graph_m1_exact
python -m gnn__models.connectivity_strategies.expert_graph_m2_inter_category
python -m gnn__models.connectivity_strategies.expert_graph_m3_intracategory

Example gnn training

python train_gnn.py --model SAGEConv --data_folder graphs/data_trivial_stat/processed/ --epochs 1 --WD 0.001 --lr 0.0001 --hidden 8192 --batch_size 512 --model_name SAGEConv_nf_stat_es_knn_2011_05_19_13_55_26 --mode_training transductive --model_folder graph_model --experiment_name exp_v1_trivial_stat