PTB-XL Dataset, Classification Task, ResNet Models, and BiLSTM Model Implementation

[jtwells2]

Authors: Anurag Dixit - anuragd2@illinois.edu, Kent Spillner - kspillne@illinois.edu, John Wells - jtwells2@illinois.edu
Original Paper: https://proceedings.mlr.press/v149/nonaka21a.html
Dataset: https://www.kaggle.com/datasets/physionet/ptbxl-electrocardiography-database

Overview

This PR adds a full pipeline replication for In-depth Benchmarking of Deep Neural Network Architectures for ECG Diagnosis, with one dataset contribution, four model contributions, one multilabel classification task, a Jupyter Notebook file showing example usage and the team's ablation study.

Dataset

1. datasets/ptbxl.py - PyHealth dataset for PTB-XL data. Processes *.hea and *.mat files, assigns train/test/val splits, and loads requisite data into a dd.DataFrame.

Models

1. resnet.py - PyHealth model implementing ResNet-18
2. lambda_resnet.py - PyHealth model implementing Lambda ResNet-18
3. se_resnet.py - PyHealth model implementing SE-ResNet-50
4. resnet_ecg_base.py - file implementing building blocks used by the three ResNet models
5. bilstm_ecg.py - PyHealth model implementing a Bidirectional LSTM

Tasks

1. ptbxl_multilabel_classification.py - PyHealth task for PTB-XL multi-label ECG classification, can use "superdiagnostic" labels (5 classes) or "diagnostic" labels (27 classes) and use either 100 Hz or 500 Hz sampling of the PTB-XL data

Example Usage / Ablation Study

1. ptbxl_superdiagnostic_se_resnet.ipynb - Jupyter Notebook example file containing runs of the multilabel classification task with all four models on a subset of 7500 patients, along with an evaluation to match what was performed in the original paper.

Unit Tests

1. test_ptbxl.py - unit tests for the PTB-XL dataset, using synthetic data
2. test_resnet_ecg.py - unit tests for all three ResNet models (ResNet-18, Lambda-ResNet-18, SE-ResNet-50), using synthetic data
3. test_bilstm_ecg.py - unit tests for BiLSTM, using synthetic data