Robust detection of Chagas disease via spatio-temporal masked autoencoder

OBJECTIVE: Chagas disease affects an estimated 6.5 million people in endemic countries, and causes nearly 10,000 deaths annually. Serological testing is effective for diagnosing individual patients, but its limited capacity makes large-scale screening impractical. The electrocardiogram (ECG) provides a low-cost, noninvasive alternative. Our objective is to develop an ECG-based algorithm for the early detection of Chagas disease. APPROACH: Firstly, we proposed a novel framework based on the Spatio-Temporal Masked Autoencoder (ST-MAE) to detect Chagas disease from standard 12-lead ECG signals. Secondly, we introduced a two-stage curriculum masking strategy. Thirdly, we introduced a soft label strategy to alleviate the issue of overfitting and handle potential label ambiguity. Fourthly, we incorporated a Gated Attention mechanism into the encoder architecture. MAIN RESULTS: As a follow-up entry after PhysioNet/Computing in Cardiology Challenge 2025, our proposed approach achieved a Challenge score of 0.255 on the official hidden test set. SIGNIFICANCE: We present a novel framework for the early detection of Chagas disease. We preliminarily verified that our two-stage curriculum masking strategy may enhance the model's ability to capture robust features. The excellent performance suggests that our framework has great potential as an assistive tool for early screening of Chagas disease in high-risk regions.