Fine-tuning a Pretrained ECG Foundation Model for Chagas Disease Detection

Yongchao Long1, Jinshuai Gu2, Mingke Yan3, Deyun Zhang4, Shijia Geng4, Jun Li5, qinghao zhao6, Yuxi Zhou7, Shenda Hong8
1Department of Computer Science, Tianjin University of Technology, 2National Institute of Health Data Science, Peking University, Beijing, China Department of Computer Science, Tianjin University of Technology, Tianjin, China., 3+86 13537539857, 4Heartvoice Medical Technology, 5Jilin University, 6Department of Cardiology, Peking University People's Hospital, 7Peking University, 8Georgia Institute of Technology

Abstract

Although large-scale acquisition of electrocardiograms (ECGs) with common disease labels is feasible, the scarcity of strongly labeled data for uncommon conditions presents significant challenges for AI diagnostic systems. As part of the George B. Moody PhysioNet Challenge 2025, we propose a series of effective deep-learning techniques for transferring ECG knowledge to detect Chagas disease. First, we trained exclusively on limited strongly labeled data. Subsequently, we extracted high-level ECG features from ECGFounder, a fully frozen foundational model pre-trained on the 10-million-scale Harvard-Emory ECG database. Finally, we concatenated these ECG embeddings with demographic features and implemented classification through a lightweight machine-learning model. Our team, ChagasExplorers, received a Challenge score of 0.988 (Ranked No. 1 out of 71 teams) on the hidden validation set