A Cascading Multi-Stage Deep Learning Approach for Detecting Chagas Disease from Electrocardiograms

Jonas Julius Sandelin1, zoher orabe2, Ismail M. Elnaggar1, Yangyang ZHAO1, Katri Karhinoja1, Chito PatiƱo1, Matti Kaisti1, Antti Airola3
1University of Turku, 2PhD, 3


Abstract

Aims: For the PhysioNet Challenge 2025, our team "bug busters" developed an approach to detect Chagas disease from electrocardiograms. This parasitic infection can be life-threatening when untreated, and ECG-based screening could direct limited resources more efficiently. Methods: We implemented a novel multi-stage cas- cading approach using five deep learning models: two ResNet18 variants with attention mechanisms (SE and CBAM), two SimpleCNN models, and an AttentionCNN. Our key innovation is a progressive filtering pipeline that ranks healthy samples by their prediction scores and re- moves those most confidently classified as healthy, creating increasingly focused training sets. Results: Our approach scored 0.369 in the official stage. Our internal testing yielded improved performance and the best model reached an average Challenge score of 0.495 with 5-fold cross-validation. Conclusion: The cascading multi-stage methodology shows promise for Chagas disease detection, overcoming the limitations of single-model approaches. Future work should investigate performance across diverse patient pop- ulations and explore interpretability of model decisions.