paul-grundmann / protopatient

Code for the AACL 2022 Paper "This Patient Looks Like That Patient: Prototypical Networks for Interpretable Diagnosis Prediction from Clinical Text"

ProtoPatient

This repository contains source code for the experiments conducted in the paper This Patient Looks Like That Patient: Prototypical Networks for Interpretable Diagnosis Prediction from Clinical Text.

A demo application can be found here: ProtoPatient Demo.

Usage

Install requirements: pip install -r requirements.txt

For training the ProtoPatient model, which achieved the best results in our experiments, run:

python training.py 
            --model_type PROTO
            --train_file {TRAIN.csv}
            --val_file {VAL.csv}
            --test_file {TEST.csv}
            --num_warmup_steps 5000
            --num_training_steps 5000
            --lr_features 0.000005
            --lr_prototypes 0.001
            --lr_others 0.001
            --use_attention True
            --reduce_hidden_size 256
            --all_labels_path {ALL_LABELS.txt}

To train the PubMedBERT baseline with the best hyperparameters, use the following command:

python training.py 
            --model_type BERT
            --train_file {TRAIN.csv}
            --val_file {VAL.csv}
            --test_file {TEST.csv}
            --num_warmup_steps 1000
            --num_training_steps 5000
            --lr_features 0.00005
            --all_labels_path {ALL_LABELS.txt}