tmehari / ptbxl_feature_benchmark

PTB-XL+, a comprehensive electrocardiographic feature dataset

This is the official code repository accompanying the new dataset PTB-XL+ hosted at physionet.com

Usage information

Preparation

1. Install dependencies from environment.yml by running conda env create -f environment.yml, activate the environment via conda activate ptbfeat

2. Follow the instructions in data_preprocessing.ipynb on how to download and preprocess PTB-XL. Download the features files 12sl_features.csv, ecgdeli_features.csv, ge12sl_glasgow_kit_ECGFeaturesMapToOMOP_draft1.xlsx unig_features.csv and put them in ./data/features. Further get the additional label informations PTB-XL+ contains about PTB-XL: 12sl_statements.csv and ptb_statements.csv and put them in ./data/statements. In the following, we assume for definiteness that the preprocessed PTB-XL can be found at ./data/ptb_xl_fs100, the features in ./data/features and the statements in ./data/statements.

Code Structure

The most functionality is provided in the files code/run_feature_benchmark.py and code/run_raw_benchmark.py.

Feature Experiments

code/run_feature_benchmark.py was used to perform the feature experiments, run

python code/run_feature_benchmark.py --dataset unig --modelname rf

to train a Random Forest on the Uni-G Features. Try a different dataset by changing the dataset parameter after --dataset. Available datasets are [unig, ecgdeli, 12sl].

The results of the runs will be written to ./output

Training a ResNet on PTB-XL labels

Train a xresnet1d50 on the normal labels of PTB-XL (at the most finegrained level) by running:

python code/run_raw_benchmark.py --dataset data/ptb_xl_fs100 --label_class label_all

Logs and trained models are saved in ./logs. Hence you can monitor training by tensorboard --logdir=./logs --port 6006. You can use a different label_set by varying --label_class Available label sets include:

Label set	Description
label_all	original PTB-XL label set
label_all_12sl	the 12SL label set of PTB-XL
label_all_12sl_ext_snomed	the 12SL label set of PTB-XL, mapped onto SNOMED labels
label_all_ptb_ext_snomed	the original label set of PTB-XL, mapped onto SNOMED labels
label_all_12sl_ext_snomed_union	the original label set of PTB-XL, mapped onto SNOMED labels, but only considering labels that also occur in the 12SL SNOMED label set
label_all_ptb_ext_snomed_union	the 12SL label set of PTB-XL, mapped onto SNOMED labels, but only considering labels that also occur in the orignal PTB-XL SNOMED label set

You can test a trained model by specifiying the location of the trained model (checkpoint) with the parameter --checkpoint_path

python code/run_raw_benchmark.py --dataset data/ptb_xl_fs100 --label_class label_all --checkpoint_path=path/to/checkpoint --test_only