The electrocardiogram (ECG) has emerged as a vital tool for diagnosing cardiac abnormalities associated with Chagas disease, enabling timely interventions. Despite advancements in data-driven methods, the generalization ability remains a critical challenge. Performance often degrades when models encounter unseen target datasets, primarily due to the dynamic and non-stationary nature of ECG signals. ECG signals exhibit substantial spatial variability across individuals and significant temporal variability within the same individual due to physiological state changes. The failure to adapt to these variations limits the robustness of conventional models.
To address these challenges, this study proposes a latent domain-invariant representation framework for automated detection of Chagas disease from 12-lead ECG signals. The ECG records are standardized by random cropping or zero-padding to a consistent length of 4096 samples. This framework directly utilizes raw ECG signals and integrates ResNet with a Squeeze-and-Excitation (SE) module to enhance the model's representative learning capacity. During feature learning, adversarial clustering enables self-supervised discovery of latent domain structures, allowing the model to capture complex intra-patient dynamics. Subsequently, adversarial learning is applied to the identified latent domains, adaptively addressing variations in spatial and morphological patterns between patient groups. This design effectively handles both inter-patient spatial variability and intra-patient temporal variability induced by physiological or pathological state changes. Initial results from the official testing for the PhysioNet Challenge indicate a challenge score of 0.181. This method demonstrates potential in improving the generalization and reliability of automated ECG-based screening for Chagas disease across diverse patient populations.
The performance remains suboptimal due to the extreme imbalance in sample distribution. As a further improvement, the model will be fine-tuned in the future using the Outlier Exposure methodology, which will improve the model's performance in detecting rare abnormalities.