[Paper Review] Beat-ssl: Capturing Local ECG Morphology through Heartbeat-level Contrastive Learning with Soft Targets

Muhammad Ilham Rizqyawan, Peter Macfarlane+2 more|arXiv (Cornell University)|Jan 22, 2026

ECG Monitoring and Analysis0 citations

TL;DR

Beat-SSL introduces rhythm-level soft-target contrastive learning and heartbeat-level hard targets within a dual-context ECG pretraining framework, achieving strong segmentation performance and close-to-best multilabel rhythm classification with substantially less pretraining data.

ABSTRACT

Obtaining labelled ECG data for developing supervised models is challenging. Contrastive learning (CL) has emerged as a promising pretraining approach that enables effective transfer learning with limited labelled data. However, existing CL frameworks either focus solely on global context or fail to exploit ECG-specific characteristics. Furthermore, these methods rely on hard contrastive targets, which may not adequately capture the continuous nature of feature similarity in ECG signals. In this paper, we propose Beat-SSL, a contrastive learning framework that performs dual-context learning through both rhythm-level and heartbeat-level contrasting with soft targets. We evaluated our pretrained model on two downstream tasks: 1) multilabel classification for global rhythm assessment, and 2) ECG segmentation to assess its capacity to learn representations across both contexts. We conducted an ablation study and compared the best configuration with three other methods, including one ECG foundation model. Despite the foundation model's broader pretraining, Beat-SSL reached 93% of its performance in multilabel classification task and surpassed all other methods in the segmentation task by 4%.

Motivation & Objective

• Motivate learning robust ECG representations with limited labeled data using self-supervised contrastive learning.
• Leverage both global rhythm context and local heartbeat morphology for richer representations.
• Introduce soft-target contrast at rhythm level and hard targets at beat level to capture continuous similarity in ECGs.

Proposed method

• Transform 12-lead ECGs to the 3D vectorcardiography (VCG) domain and apply 3KG-style augmentations.
• Perform rhythm-level contrastive learning with a shared encoder using two augmented views and NT-Xent loss.
• Extract beat-level representations from the encoder output by ROI-like pooling around R-peaks and apply a beat projection head for contrastive learning.
• Use hard beat-level targets from a beat-classifier trained on Lead-II to guide beat-level Contrastive Learning.
• Introduce soft contrasting targets (soft_1 and soft_2) based on feature similarity or neighborhood weighting to provide continuous similarity signals.
• Optimize a generalized NT-Xent loss with weights that encode the soft targets between pairs.

Experimental results

Research questions

• RQ1 Can dual-context contrastive learning (rhythm-level and heartbeat-level) produce richer ECG representations than single-context approaches?
• RQ2 Do soft-target contrastive signals improve representation quality for global rhythm tasks and local beat morphology tasks compared to hard targets?
• RQ3 How does Beat-SSL perform on downstream tasks requiring global rhythm assessment and local waveform delineation (segmentation) relative to ECG foundation models and other CL methods.

Key findings

• Beat-SSL achieves strong downstream performance on two tasks: multilabel rhythm classification and ECG wave segmentation.
• The best configuration reaches 93% of the best F1-score achieved by a larger ECG foundation model while using significantly less pretraining data (about 31.8x less).
• Beat-SSL surpasses all other methods in the segmentation task by 4% (Dice/F1 metrics reported).
• Soft rhythm-level contrasting (soft_1) improves performance for the rhythm task, while the combination with beat-level hard targets yields strong segmentation performance.
• An ablation study shows exponentiating soft targets (power 50) consistently benefits performance across tasks and configurations.