Y-Net-ECG: A Multi-Lead informed and interpretable architecture for ECG segmentation across diverse rhythms

Abstract

Accurate electrocardiogram (ECG) segmentation is critical for diagnosing and monitoring cardiac conditions. However, the accuracy of ECG segmentation across different heart rhythm types remains a challenge, and its practical utility in disease diagnosis remains to be fully validated. To address these challenges, we propose Y-Net, a deep learning model designed to perform robust ECG segmentation under both single-lead and multi-lead input modes. The model incorporates a dual-branch structure and a two-stage training strategy to ensure adaptability across various clinical scenarios. We evaluated Y-Net on two 12-lead ECG segmentation datasets: LUDB, a public dataset, and RDB, a privately annotated dataset based on public data but annotated specifically by our team. Y-Net demonstrated robust performance across datasets and rhythm types, achieving F1 scores of 99.60% and 99.44% in intra-dataset evaluations, and 99.03% and 98.24% in inter-dataset tests. To improve interpretability, we introduce an intermediate feature visualization method and apply segmentation results directly to atrial fibrillation (AF) detection based on P-wave absence. This morphology-based approach achieves AUCs of 0.946, 0.971, and 0.983 on the PhysioNet2017, CPSC2018, and AFDB datasets, respectively, without the need for additional classifiers. These results highlight the effectiveness and clinical potential of Y-Net as a transparent and adaptable tool for ECG segmentation and interpretation across diverse cardiac rhythms.