luyang-ai4med / popdx

Python implementation of POPDx - Predictions for unseen, rare, and common labels.

POPDx: An Automated Framework for Patient Phenotyping across 392,246 Individuals in the UK Biobank Study

POPDx (Population-based Objective Phenotyping by Deep Extrapolation) is a bilinear machine learning framework for simultaneous multi-phenotype recognition. For additional information, please refer to our manuscript, available at https://academic.oup.com/jamia/advance-article/doi/10.1093/jamia/ocac226/6873915.

To cite:
Yang, Lu, Sheng Wang, and Russ B. Altman. "POPDx: an automated framework for patient phenotyping across 392 246 individuals in the UK Biobank study." Journal of the American Medical Informatics Association 30.2 (2023): 245-255.

Tools for UK Biobank

Please stay tuned.

Installation

Please clone our github repository as follows:

git clone https://github.com/luyang-ai4med/POPDx.git

Dependencies

POPDx is developed in Python 3. We provide the conda environment containing the necessary dependencies. For your experiments, we suggest using a single GPU (e.g. NVIDIA Tesla V100 SXM2 16 GB).

conda env create -f popdx.yml
conda activate popdx

Label embeddings

Please refer to the sample notebook for generating the ICD-10/Phecode embeddings.

POPDx/code/create_label_embeddings.ipynb

Lines 1 to 6 in 2035055

	{
	"cells": [
	{
	"cell_type": "code",
	"execution_count": 1,
	"id": "d58208cb",

POPDx training

POPDx can be explored and run through the command lines as follows:

python code/POPDx_train.py -h
python code/POPDx_train.py -d './save/POPDx_train' 

Additional parameters can be defined by the user.

The script to train POPDx. 
Please specify the train/val datasets path in the python script.

optional arguments:
  -h, --help            show this help message and exit
  -d SAVE_DIR, --save_dir SAVE_DIR
                        The folder to save the trained POPDx model e.g.
                        "./save/POPDx_train"
  -s HIDDEN_SIZE, --hidden_size HIDDEN_SIZE
                        Default hidden size is 150.
  --use_gpu USE_GPU     Default setup is to use GPU.
  -lr LEARNING_RATE, --learning_rate LEARNING_RATE
                        Default learning rate is 0.0001
  -wd WEIGHT_DECAY, --weight_decay WEIGHT_DECAY
                        Default weight decay is 0

POPDx testing

POPDx can be tested through the command lines as follows:

python code/POPDx_test.py -h 
python code/POPDx_test.py -m "./save/POPDx_train/best_classifier.pth.tar" -o "./save/POPDx_train/test/"

Additional parameters can be defined by the user.

usage: POPDx_test.py [-h] -m MODEL_PATH -o OUTPUT_PATH [-s HIDDEN_SIZE]
                     [-b BATCH_SIZE] [--use_gpu USE_GPU]

The script to test POPDx. 
Please specify the path to the test datasets in the python script.

optional arguments:
  -h, --help            show this help message and exit
  -m MODEL_PATH, --model_path MODEL_PATH
                        The path to POPDx model e.g.
                        "./save/POPDx_train/best_classifier.pth.tar"
  -o OUTPUT_PATH, --output_path OUTPUT_PATH
                        The output directory e.g. "./save/POPDx_train/test/"
  -s HIDDEN_SIZE, --hidden_size HIDDEN_SIZE
                        Default hidden size is 150. Consistent with training.
  -b BATCH_SIZE, --batch_size BATCH_SIZE
                        Default batch size is 512.
  --use_gpu USE_GPU     Default setup is to not use GPU for test.