Generative Data by β-Variational Autoencoders Help Build Stronger Classifiers: ECG Use Case

Yousef Nademi and Sunil V Kalmady and Weijie Sun and Amir Salimi and Abram Hindle and Padma Kaul and Russell Greiner

2023/09/24

Generative Data by β-Variational Autoencoders Help Build Stronger Classifiers: ECG Use Case

Authors

Yousef Nademi and Sunil V Kalmady and Weijie Sun and Amir Salimi and Abram Hindle and Padma Kaul and Russell Greiner

Venue

Abstract

We explore the challenge of learning models that use electrocardiogram (ECG) data to diagnose various cardiovascular diseases. Here, we explore whether classifiers trained on a dataset of real labeled ECGs, augmented with synthetic ECGs, can perform better than ones trained on unaugmented datasets. We first used a dataset of ECGs, each labelled with one or more of 15 diagnoses, from 244,077 patients to train an unsupervised $eta$-VAE model, that could generate time series of 12-lead ECG signals for each of the diagnoses. We then used this generative model to generate ECGs with the ST-segment Elevated (STE) abnormality, which we added to the public dataset of ECG abnormalities (n = 6877, over normal (Sinus Rhythm) and 8 different abnormalities) of China Physiological Signal Challenge 2018, and found a learner trained on this extended dataset performed better than one trained on only the original data on the targeted STE label but also enhanced its performance for the classification of 4 other labels.

Bibtex

@inproceedings{nademi2023SIPAIM-autoencoder-ECG,
 abstract = {We explore the challenge of learning models that use electrocardiogram (ECG) data to diagnose various cardiovascular diseases. Here, we explore whether classifiers trained on a dataset of real labeled ECGs, augmented with synthetic ECGs, can perform better than ones trained on unaugmented datasets. We first used a dataset of ECGs, each labelled with one or more of 15 diagnoses, from 244,077 patients to train an unsupervised $eta$-VAE model, that could generate time series of 12-lead ECG signals for each of the diagnoses. We then used this generative model to generate ECGs with the ST-segment Elevated (STE) abnormality, which we added to the public dataset of ECG abnormalities (n = 6877, over normal (Sinus Rhythm) and 8 different abnormalities) of China Physiological Signal Challenge 2018, and found a learner trained on this extended dataset performed better than one trained on only the original data on the targeted STE label but also enhanced its performance for the classification of 4 other labels.},
 accepted = {2023-09-24},
 author = {Yousef Nademi and Sunil V Kalmady and Weijie Sun and Amir Salimi and Abram Hindle  and Padma Kaul and Russell Greiner},
 authors = {Yousef Nademi and Sunil V Kalmady and Weijie Sun and Amir Salimi and Abram Hindle  and Padma Kaul and Russell Greiner},
 booktitle = {2023 19th International Symposium on Medical Information Processing and Analysis (SIPAIM)},
 code = {nademi2023SIPAIM-autoencoder-ECG},
 date = {2023-11-15},
 doi = {10.1109/sipaim56729.2023.10373478},
 funding = {NSERC Discovery},
 location = {Mexico City, Mexico},
 pages = {1--7},
 rate = {},
 role = {Co-Author},
 title = {Generative Data by β-Variational Autoencoders Help Build Stronger Classifiers: ECG Use Case},
 type = {inproceedings},
 url = {http://softwareprocess.ca/pubs/nademi2023SIPAIM-autoencoder-ECG.pdf},
 venue = {2023 19th International Symposium on Medical Information Processing and Analysis (SIPAIM)},
 year = {2024}
}