PPH-Extract

Overview

This repository contains the code for the paper "Zero-shot Interpretable Phenotyping of Postpartum Hemorrhage Using Large Language Models". While we unfortunately cannot release the data used in the paper because it contains protected health information, we release the code base, which includes all prompts we use in our experiments.

Installation

Download the repo

First, download our github repository via the following:

git clone https://github.com/EmilyAlsentzer/PPH-Extract
cd PPH-Extract

Environment Setup

To create a conda environment containing all of the required packages, ensure that conda is installed and run the following:

conda env create -f environment.yml
conda activate pph-extract

Install the Repo

After the conda environment is created and activated, install the github repo with the following:

pip install -e .

Set up the config file

Go to config.py and set the project directory (PROJECT_DIR) to be the path where you will store the data and model outputs for the project.

Usage

Perform zero-shot prediction

You can generate predictions for notes using the script at model/zero_shot_predictions.py.

a) If you do not have annotations for your notes, cd into the model folder and run as follows:

 accelerate launch \
  --main_process_port PORT_NUMBER \
  --config_file bf_16.yaml  \
  zero_shot_predictions.py  \
  --label LABEL \
  --batch_size BATCH_SIZE \
  --max_length MAX_LEN \
  --prompt_desc PROMPT \
  --unlabelled \
  --unlab_filename FILENAME_OF_UNLABELED_DATA

where PORT_NUMBER, BATCH_SIZE, and MAX_LEN are parameters, LABEL is one of the labels in config.binary_labels or config.ie_labels, PROMPT is the type of prompt used (all prompts can be found in the get_prompt function), and FILENAME_OF_UNLABELED_DATA is the filename of the csv containing the unlabelled clinical notes found in the config.UNLABELLED_DATA_DIR directory. This uses Huggingface's Accelerate to fit more data into memory at once.

b) if you do have annotations for your notes, cd into the model folder and run as follows:

 accelerate launch \
  --main_process_port PORT_NUMBER \
  --config_file bf_16.yaml  \
  zero_shot_predictions.py  \
  --label LABEL \
  --batch_size BATCH_SIZE \
  --max_length MAX_LEN \
  --prompt_desc PROMPT \
  --annotations_filename FILENAME_OF_ANNOTATIONS_CSV \
  --all_notes_filename FILENAME_OF_ALL_NOTES_CSV \
  --labeled FILENAME_OF_LABELED_DATA

where PORT_NUMBER, LABEL, BATCH_SIZE, MAX_LEN, and PROMPT are as described above, FILENAME_OF_ANNOTATIONS_CSV is the csv containing the annotations for all notes (processed using process_annotator_output.py), FILENAME_OF_ALL_NOTES_CSV is the csv containing all notes, and FILENAME_OF_LABELED_DATA is the csv containing the labelled notes we're trying to evaluate.