ECG-Based Screening of Chagas Disease Using Deep Residual Networks and Feature-Based Machine Learning

Marion Taconne, stefano magni, Cristian Drudi, Sara Maria Pagotto, Valentina Corino, Pietro Cerveri, anna maria maddalena bianchi, Riccardo Barbieri, Luca Mainardi
Politecnico di Milano


Abstract

Context: Chagas disease affects millions of people in endemic regions and can lead to severe cardiac complications. Given limited access to serological testing in resource-constrained settings, scalable and accessible screening tools are urgently needed. Electrocardiograms (ECGs), which often exhibit Chagas-related abnormalities, present a promising modality for early detection. In the context of the George B. Moody PhysioNet Challenge 2025, we developed a deep learning model to identify Chagas disease from ECGs.

Methods: We designed a ResNet-based architecture to classify 12-lead ECG recordings as Chagas or non-Chagas. The preprocessing stage consisted of resampling, baseline removal, and standardization of signal length via padding or cropping. We incorporated a signal quality index based on frequency content and recording duration. During inference, low-quality signals were automatically classified as non-Chagas, reflecting diagnostic uncertainty. The processed ECGs were then fed into the proposed ResNet-based architecture (Figure 1). Training was performed using the Adam optimizer with binary cross-entropy loss.

Results: On our internal test set, the model achieved a challenge score of 0.723. In the unofficial phase of the competition, our approach yielded a score of 0.566 and placed 20th on the leaderboard under the team name Chaguys.

Conclusions: These results highlight the potential of a ResNet-based architecture, paired with careful preprocessing and signal quality checks, to detect Chagas-related ECG abnormalities. Future work will focus on refining the training dataset by discarding, reweighting, or relabeling poor-quality or non-Chagas pathological ECGs. We also plan to perform relabeling by clustering latent representations learned by the ResNet, to identify and correct potentially mislabeled data and improve overall performance.