Beat-Ssl Achieves 93% ECG Accuracy Via Novel Heartbeat-Level Contrastive Learning

Researchers are tackling the persistent problem of limited labelled data in electrocardiogram (ECG) analysis, a crucial barrier to developing effective diagnostic tools. Muhammad Ilham Rizqyawan, Peter Macfarlane, Stathis Hadjidemetriou, et al., from the University of Glasgow and the University of Limassol, present Beat-SSL, a novel contrastive learning framework designed to capture nuanced ECG morphology at the individual heartbeat level. This work distinguishes itself by combining rhythm-level and heartbeat-level contrasting alongside the innovative use of ‘soft’ targets, better reflecting the continuous similarities within ECG signals. By pretraining with Beat-SSL, the team achieved performance reaching 93% of fully supervised methods in rhythm assessment and significantly outperformed existing techniques in ECG segmentation by 4%, demonstrating a substantial advance in representation learning for cardiac health monitoring.

The research team achieved this by moving beyond traditional contrastive learning methods that often focus solely on global context or rely on simplistic hard contrastive targets, which fail to capture the nuanced similarities within ECG signals. Beat-SSL leverages dual-context learning, effectively integrating information from entire heart rhythms and individual heartbeats to create more robust and informative representations.

The study unveils a system that transforms 12-lead ECG signals into the vectorcardiography (VCG) domain, employing the established 3KG augmentation strategy to enhance data variability. Subsequently, contrastive learning is applied at both global and local levels, utilising innovative soft contrastive strategies that employ ECG feature similarity as a continuous target ranging from 0 to 1, a departure from conventional binary positive-negative pairings. This approach allows the model to discern subtle differences in beat morphologies, identifying key characteristics of waves like the P-wave, QRS-complex, and T-wave, which are fundamental units of cardiac activity. Experiments demonstrate that Beat-SSL effectively learns representations applicable across both rhythm and beat contexts, paving the way for improved diagnostic accuracy.
Researchers validated their pretrained model on two crucial downstream tasks: multilabel classification for comprehensive rhythm assessment and ECG segmentation to evaluate its capacity to learn representations at multiple scales. A rigorous ablation study was conducted to optimise the framework’s configuration, and the best performing setup was then compared against three other methods, including a prominent ECG foundation model. Despite the foundation model undergoing broader pretraining, Beat-SSL achieved an impressive 93% of its performance in the multilabel classification task, demonstrating its competitive edge. Notably, the team’s innovation surpassed all other methods in the ECG segmentation task by a significant 4%, highlighting its superior ability to capture local ECG morphology through heartbeat-level contrastive learning.

This achievement establishes Beat-SSL as a promising approach for semi-supervised learning in ECG analysis, potentially reducing the reliance on extensive labelled datasets and accelerating the development of advanced cardiac diagnostic tools. The. Experiments revealed that Beat-SSL reached 93% of the performance achieved by the best-performing method in a multilabel classification task, despite utilising 31.8times less pretraining data. Furthermore, the framework surpassed all other methods in an ECG segmentation task by a significant 4%, demonstrating its superior ability to learn representations across both rhythm and heartbeat contexts.
The team conducted an extensive ablation study, testing 14 combinations of rhythm-level contrasting, beat-level contrasting, and exponentiation to optimise performance on both downstream tasks. Results from this study showed that a combination of soft contrasting at the rhythm level, hard contrasting at the beat level, and exponentiation with a power of 50 achieved the best overall performance. Specifically, this configuration yielded an AUROC of 0.794 and F1-scores of 0.540 and 0.805 for tasks 1 and 2, respectively. Component-wise analysis confirmed that utilising an exponent of 50 consistently improved performance across all tasks and configurations, while incorporating soft contrasting at the rhythm level enhanced results for both downstream tasks.

Measurements confirm that Beat-SSL outperformed established frameworks, including TS2Vec, Domain-SSL, and ECG-FM, in the segmentation task. The model achieved a Dice-score of 0.911 and an F1-score of 0.881, demonstrating a clear advantage over the second-best method, TS2Vec. Statistical analysis using the Wilcoxon signed-rank test with Bonferroni correction revealed strong evidence (p 0.01) supporting the superior performance of Beat-SSL in both tasks. The team restricted metrics to samples between 500 and 4500 to avoid false positives arising from missing boundary labels in the LUDB dataset.

Tests prove that by leveraging soft contrast at the rhythm level and hard contrasting at the beat level, Beat-SSL effectively captures both local and global contextual information within the ECG signal. This capability is particularly beneficial for classifying conditions requiring analysis of both rhythm and morphological features, providing a strong local representation that significantly improves classification accuracy.