300 Section - Multi-scale Self-supervised Representation Fusion with Clinical Features for Enhanced ECG Classification in Cardiac Rehabilitation

Wednesday, May 20, 2026

6:00 PM - 7:30 PM PT

Presenting Author/Autor expositor(s)

FF

Fan Fu, PhD

Post-doctoral
Nantong First People's Hospital affiliated to Southeast University
Nantong, Jiangsu, China (People's Republic)

Co-Author/Coautor(s)

JL

Jiamin Liu, MS

Assistant Researcher
Nantong First People's Hospital affiliated to Southeast University
Nantong, Jiangsu, China (People's Republic)
WC

Weiguan Chen, MD

Director of Department of Rehabilitation Medicine
Nantong First People's Hospital affiliated to Southeast University
Nantong, Jiangsu, China (People's Republic)

Corresponding Author/Autor a quien dirigir la corr(s)

HL

Hongjian Lu, PhD

Professor
Nantong First People's Hospital affiliated to Southeast University
Nantong, Jiangsu, China (People's Republic)

Objectives : Electrocardiogram (ECG) analysis plays a pivotal role in cardiovascular disease diagnosis, risk stratification, and continuous monitoring during cardiac rehabilitation. However, existing self-supervised learning (SSL) approaches for ECG classification often lack explicit evaluation of the pretraining process and are restricted to a single representation scale, limiting their effectiveness in real-world rehabilitation scenarios. This study aims to develop a multi-scale and clinically informed framework that enhances ECG representation quality, supports individualized exercise prescription, and improves early detection of rehabilitation-related cardiac risks.

Design: We propose the Multi-scale Self-supervised Representation Fusion Network (MSRFN), which integrates multi-scale SSL representations with clinically guided features. The model employs multi-scale temporal masking and reconstruction-based optimization to enhance pretraining effectiveness and broaden representational diversity. A fuzzy attention–based fusion (FAF) module is further designed to adaptively integrate multi-scale physiological representations with clinical indicators, enabling robust, interpretable decision-making for long-term ECG monitoring and cardiac rehabilitation applications.

Results: MSRFN was evaluated on three benchmark ECG datasets—CODE-15, PTB-XL, and CPSC2018—and consistently outperformed state-of-the-art models including STMEM, TCN, and Medformer. Improvements were observed across AUROC, accuracy, precision, recall, and F1 score. Ablation analyses confirmed the necessity of both multi-scale SSL representations (MSR) and the FAF module. The model demonstrated enhanced sensitivity to subtle arrhythmic and ischemic features, supporting safer and more personalized rehabilitation management, including early warning of adverse events and continuous home-based monitoring.

Conclusion: MSRFN provides a robust, interpretable, and clinically relevant framework for multi-scale ECG representation learning, significantly improving downstream classification performance. Its capacity to enhance cardiac risk identification and support continuous physiological monitoring highlights its potential as an intelligent decision-support tool for personalized cardiac rehabilitation. The proposed model advances the integration of AI-driven ECG analytics into rehabilitation workflows, contributing to safer, adaptive, and patient-centered recovery pathways.