Continuity Contrastive Representations of ECG for Heart Block Detection from Only Lead-I

Teya Bergamaschi, Collin Stultz, Ridwan Alam
Proceedings of the 4th Machine Learning for Health Symposium, PMLR 259:130-142, 2025.

Abstract

Early detection of heart block can prevent life-threatening outcomes in patients with cardiac conduction disorders. While 12-lead ECG interpretation is the clinical standard apparatus, this work investigates detecting heart block from the lead-I ECG signals, the lead available on commercial smartwatches. We evaluate two state-of-the-art architectures: residual neural network and transformer encoder, both trained in a self-supervised contrastive learning manner with a novel signal-continuity-based ECG view definition on a dataset of 3.6 million ECGs from Massachusetts General Hospital. These models learn efficient ECG representations, which are used for heart block detection via linear probing on the PTB-XL dataset, a public ECG resource. To provide performance benchmarks, we compare our self-supervised models to supervised adaptations of both models trained on 10.6 thousand single-lead PTB-XL ECGs. Our analysis evaluates the performance of each model using the area under the receiver-operating curve (AUC), sensitivity, and specificity. We observe improved performance from the self-supervised pretraining. Additionally, we demonstrate the robust generalizability of these models in scarce-data scenarios, maintaining consistent performance with a reduced number of labeled training examples. This study highlights the potential of self-supervised learning in lead-I ECG diagnostics, offering promising implications for clinical applications where labeled data is scarce.

Cite this Paper


BibTeX
@InProceedings{pmlr-v259-bergamaschi25a, title = {Continuity Contrastive Representations of ECG for Heart Block Detection from Only Lead-I}, author = {Bergamaschi, Teya and Stultz, Collin and Alam, Ridwan}, booktitle = {Proceedings of the 4th Machine Learning for Health Symposium}, pages = {130--142}, year = {2025}, editor = {Hegselmann, Stefan and Zhou, Helen and Healey, Elizabeth and Chang, Trenton and Ellington, Caleb and Mhasawade, Vishwali and Tonekaboni, Sana and Argaw, Peniel and Zhang, Haoran}, volume = {259}, series = {Proceedings of Machine Learning Research}, month = {15--16 Dec}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v259/main/assets/bergamaschi25a/bergamaschi25a.pdf}, url = {https://proceedings.mlr.press/v259/bergamaschi25a.html}, abstract = {Early detection of heart block can prevent life-threatening outcomes in patients with cardiac conduction disorders. While 12-lead ECG interpretation is the clinical standard apparatus, this work investigates detecting heart block from the lead-I ECG signals, the lead available on commercial smartwatches. We evaluate two state-of-the-art architectures: residual neural network and transformer encoder, both trained in a self-supervised contrastive learning manner with a novel signal-continuity-based ECG view definition on a dataset of 3.6 million ECGs from Massachusetts General Hospital. These models learn efficient ECG representations, which are used for heart block detection via linear probing on the PTB-XL dataset, a public ECG resource. To provide performance benchmarks, we compare our self-supervised models to supervised adaptations of both models trained on 10.6 thousand single-lead PTB-XL ECGs. Our analysis evaluates the performance of each model using the area under the receiver-operating curve (AUC), sensitivity, and specificity. We observe improved performance from the self-supervised pretraining. Additionally, we demonstrate the robust generalizability of these models in scarce-data scenarios, maintaining consistent performance with a reduced number of labeled training examples. This study highlights the potential of self-supervised learning in lead-I ECG diagnostics, offering promising implications for clinical applications where labeled data is scarce.} }
Endnote
%0 Conference Paper %T Continuity Contrastive Representations of ECG for Heart Block Detection from Only Lead-I %A Teya Bergamaschi %A Collin Stultz %A Ridwan Alam %B Proceedings of the 4th Machine Learning for Health Symposium %C Proceedings of Machine Learning Research %D 2025 %E Stefan Hegselmann %E Helen Zhou %E Elizabeth Healey %E Trenton Chang %E Caleb Ellington %E Vishwali Mhasawade %E Sana Tonekaboni %E Peniel Argaw %E Haoran Zhang %F pmlr-v259-bergamaschi25a %I PMLR %P 130--142 %U https://proceedings.mlr.press/v259/bergamaschi25a.html %V 259 %X Early detection of heart block can prevent life-threatening outcomes in patients with cardiac conduction disorders. While 12-lead ECG interpretation is the clinical standard apparatus, this work investigates detecting heart block from the lead-I ECG signals, the lead available on commercial smartwatches. We evaluate two state-of-the-art architectures: residual neural network and transformer encoder, both trained in a self-supervised contrastive learning manner with a novel signal-continuity-based ECG view definition on a dataset of 3.6 million ECGs from Massachusetts General Hospital. These models learn efficient ECG representations, which are used for heart block detection via linear probing on the PTB-XL dataset, a public ECG resource. To provide performance benchmarks, we compare our self-supervised models to supervised adaptations of both models trained on 10.6 thousand single-lead PTB-XL ECGs. Our analysis evaluates the performance of each model using the area under the receiver-operating curve (AUC), sensitivity, and specificity. We observe improved performance from the self-supervised pretraining. Additionally, we demonstrate the robust generalizability of these models in scarce-data scenarios, maintaining consistent performance with a reduced number of labeled training examples. This study highlights the potential of self-supervised learning in lead-I ECG diagnostics, offering promising implications for clinical applications where labeled data is scarce.
APA
Bergamaschi, T., Stultz, C. & Alam, R.. (2025). Continuity Contrastive Representations of ECG for Heart Block Detection from Only Lead-I. Proceedings of the 4th Machine Learning for Health Symposium, in Proceedings of Machine Learning Research 259:130-142 Available from https://proceedings.mlr.press/v259/bergamaschi25a.html.

Related Material