.png){kind=icon}
.png){kind=icon}
Return
Self-Supervised Contrastive Learning Enables Robust ECG-Based Cardiac Classification
.png){kind=icon}
OA
AI
DOI：10.1016/j.hroo.2026.01.016.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
•Self-supervised learning enables accurate ECG-based cardiac classification even when labeled clinical data are scarce. Models pretrained on unlabeled ECGs achieved substantially higher performance than standard supervised models when trained with limited labeled data, supporting use in real-world settings where expert annotations are expensive or unavailable. •ECG representations learned from large, unlabeled cohorts generalize across clinically distinct tasks. Pretraining on one cardiac phenotype (e.g., left ventricular dysfunction) improved performance on a different biochemical outcome (serum potassium abnormality), suggesting that contrastive learning captures broadly relevant cardiac signal features rather than task-specific patterns. • Architectural inductive bias via lead grouping (LGTemporalNet) consistently enhances performance, especially for the KCL task, indicating that explicitly modeling inter-lead structure improves representation quality beyond standard temporal convolutional models. • Model adaptation to the target task remains important for clinical deployment. While pretrained models provided strong initial representations, fine-tuning on task-specific data yielded the highest performance, emphasizing the need for downstream calibration before clinical use.
Keywords:
Machine learning
Electrocardiogram
Self-Supervised Learning
Contrastive Learning
Representational Learning
Pre-Training
Cardiac Classification
Foundational Models
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.