Chagas disease is a parasitic infection that affects approximately 6.5 million individuals, primarily in Latin America. It often produces distinctive patterns in ECG signals, making ECG a valuable tool for its detection. Conventional deep learning models pre-trained on large ECG datasets have shown promise in identifying various ECG abnormalities. However, because Chagas disease predominantly occurs in a specific region, models trained solely on broadly sourced ECG data may face distributional mismatches that limit the effectiveness of their learned representations. In contrast, time-series foundation models, which are trained on diverse time-series data, can extract robust, domain-independent features that help overcome these challenges.
In this work, a hybrid approach is proposed to leverage both strategies. For ECG-specific analysis, a ConvNeXt V2 architecture is adapted by replacing 2D convolutions with 1D convolutions and reducing the kernel size to better capture subtle variations in ECG signals. Simultaneously, a transformer-based time-series foundation model extracts lead-specific representations, and an attention module dynamically weighs these features. The weighted time-series features are concatenated with the ECG-specific features, and the resulting combined vector is used to train a shallow neural network for final classification.
The code was submitted under the name SharifAIPlatform and achieved a PhysioNet Challenge score of 0.127. Additionally, local cross-validation yielded an AUROC of 0.765, an AUPRC of 0.152, an accuracy of 0.971, and an F-measure of 0.065.