Two-Stage Domain Adversarial Learning to Identify Chagas Disease from ECG and Patient Demographic Data

Large-scale automated ECG screening can combat the widespread underdiagnosis of Chagas disease due to limited serological test coverage. To this end, our team, CinCo Amigos, developed a computational approach to detect Chagas disease from electrocardiograms (ECGs) - a two-stage domain-adversarial training process to address key issues of significant label noise, extreme class imbalance, and substantial domain shift.

Our framework first pre-trains a custom neural network on a large, noisy dataset. Early Learning Regularization (ELR) and Domain-Adversarial Neural Network (DANN) were integrated to mitigate label errors and encourage domain-invariant features respectively. To handle class imbalance, we employed a objective combining Focal Loss (LMFLoss) and Label-Distribution-Aware Margin (LDAM) Loss. In the second stage, the model was finetuned on high-quality datasets using feature distillation.

Our model achieved an official Challenge score of 0.250 (ranked 7 of 40 teams), and was the best performing on one of the three test sets. This work suggests that our integrated approach provides a robust framework for automated ECG-based diagnosis and can improve generalisation in challenging real-world scenarios.