Detection of Chagas Disease Using Digital Electrocardiogram by Deep Transfer Learning of the InceptionTime Model

Arnaud Champetier
Cardiovascular Research Institute Basel (CRIB)


Abstract

This paper introduces the solution of team cardio_Basel25 to the PhysioNet Challenge 2025. Chagas disease, caused by Trypanosoma cruzi, remains underdiagnosed in low-resource settings where serological testing is limited. Given the widespread availability of electrocardiography (ECG) and the conduction abnormalities characteristic of Chagas cardiomyopathy and ECG-based artificial intelligence offers a scalable alternative for early detection. A deep transfer learning model was developed based on a pretrained InceptionTime architecture, and fine-tuned on Brazilian (CODE-15%, Sami-Trop) and European (PTB-XL) ECG datasets. Recordings were preprocessed with filtering, downsampling, normalization, and 5-second segment extraction, with data augmentation applied during training. The training loss consisted of binary cross-entropy with a penalization term to emphasize the challenge metric. Inference combined predictions across multiple ECG segments and models. The approach achieved a cross-validation score of 0.42 and a score of 0.382 on the validation set, ranking 27th on the official leaderboard. These findings demonstrate the feasibility of deep transfer learning for ECG-based Chagas screening and its potential to expand diagnostic access in underserved regions.