Multilabel 12-Lead Electrocardiogram Classification Using Beat To Sequence Autoencoders

Alexander William Wong, Amir Salimi, Abram Hindle, Sunil Vasu Kalmady, Padma Kaul

2021/01/29

Multilabel 12-Lead Electrocardiogram Classification Using Beat To Sequence Autoencoders

Authors

Alexander William Wong, Amir Salimi, Abram Hindle, Sunil Vasu Kalmady, Padma Kaul

Venue

Abstract

The 12-lead electrocardiogram (ECG) measures the electrical activity of the heart for physicians to use in diagnosing cardiac disorders. This paper investigates the multi-label, multi-class classification of ECG records into one or more of 27 possible medical diagnoses. Our multi-step approach uses conventional physiological algorithms for segmentation of heartbeats from the baseline signals. We stack a heartbeat autoencoder over heartbeat windows to make embeddings, then we encode this sequence of embeddings to make an ECG embedding which we then classify on. We utilize the public dataset of 43,101 available ECG records provided by the PhysioNet/CinC 2020 challenge, performing repeated random subsampling and splitting the available records into 80% training, 10% validation, and 10% test splits, 20 times. We attain a mean test split challenge score of 0.248 with an overall macro F 1 score of 0.260 across the 27 labels.

Bibtex

@inproceedings{wong2021ICASSP-ecg-autoencoder,
 abstract = {The 12-lead electrocardiogram (ECG) measures the electrical activity of the heart for physicians to use in diagnosing cardiac disorders. This paper investigates the multi-label, multi-class classification of ECG records into one or more of 27 possible medical diagnoses. Our multi-step approach uses conventional physiological algorithms for segmentation of heartbeats from the baseline signals. We stack a heartbeat autoencoder over heartbeat windows to make embeddings, then we encode this sequence of embeddings to make an ECG embedding which we then classify on. We utilize the public dataset of 43,101 available ECG records provided by the PhysioNet/CinC 2020 challenge, performing repeated random subsampling and splitting the available records into 80% training, 10% validation, and 10% test splits, 20 times. We attain a mean test split challenge score of 0.248 with an overall macro F 1 score of 0.260 across the 27 labels.},
 accepted = {2021-01-29},
 author = {Alexander William Wong and Amir Salimi and Abram Hindle and Sunil Vasu Kalmady and Padma Kaul},
 authors = {Alexander William Wong, Amir Salimi, Abram Hindle, Sunil Vasu Kalmady, Padma Kaul},
 booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
 code = {wong2021ICASSP-ecg-autoencoder},
 date = {2021-06-10},
 funding = {NSERC Discovery},
 location = {Toronto, Canada},
 pagerange = {1--4},
 pages = {1--4},
 rate = {},
 role = {Co-Author},
 title = {Multilabel 12-Lead Electrocardiogram Classification Using Beat To Sequence Autoencoders},
 type = {inproceedings},
 url = {http://softwareprocess.ca/pubs/wong2021ICASSP-ecg-autoencoder.pdf},
 venue = {IEEE International Conference on Acoustics, Speech and Signal Processing},
 year = {2021}
}