Large-scale ECG preliminary screening can efficiently identify high-risk individuals for targeted confirmatory testing, combating the widespread under-diagnosis of Chagas disease due to limited serological test coverage. It is this potential that provides the fundamental motivation for developing automated ECG screening. We developed a computational approach to detect Chagas disease from electrocardiograms (ECGs). Our team, \textit{CinCo Amigos}, developed a two-stage domain-adversarial training process to address the key issues of significant label noise, extreme class imbalance, and substantial domain shift.
Our two-stage framework first pre-trains a custom convolutional neural network on a large, noisy dataset. We integrate Early Learning Regularization (ELR) to mitigate label errors and a Domain-Adversarial Neural Network (DANN) to encourage domain-invariant features. To handle class imbalance, we employ LMFLoss, a composite objective combining Focal Loss and Label-Distribution-Aware Margin (LDAM) Loss. In the second stage, the model is fine-tuned on high-quality datasets using feature distillation to retain generalisable features.
Our model achieved a Challenge score of 0.338 on the validation set. Our official Challenge score was -- (ranked -- out of -- teams) on the hidden test set. 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.