Auxiliary-conditioned cross-attention with physiologically interpretable features for Chagas disease detection from 12-lead ECGs.

OBJECTIVE: Chagas disease remains a major public-health concern in endemic regions, and chronic cardiac involvement often manifests as conduction abnormalities detectable on standard 12-lead electrocardiograms (ECGs). Reliable automated screening remains challenging because of dataset heterogeneity and label uncertainty, particularly when combining strongly labeled cohorts with large weakly labeled repositories. APPROACH: We propose a hybrid architecture that integrates a 1D ResNet encoder for local ECG morphology, a bidirectional GRU for long-range temporal context, and handcrafted physiological features and demographics through an auxiliary-conditioned cross-attention module. The auxiliary vector, comprising age, sex, and QRS/conduction descriptors, is projected into a query token that selectively attends to deep sequential embeddings for feature-aware temporal aggregation. To exploit heterogeneous sources while reflecting source reliability, we further adopt a source-aware weighted binary cross-entropy objective. MAIN RESULTS: As team CAUETUMN in the PhysioNet/Computing in Cardiology Challenge 2025, the framework achieved a score of 0.347 on the organizer-held REDS-II leaderboard-validation set during the official phase and 0.218 on the final hidden test set, ranking 17th among 41 eligible teams. SIGNIFICANCE: These results suggest that conditioning detection on interpretable QRS and conduction descriptors supports a transparent and physiologically informed screening framework, while highlighting the difficulty of generalizing across heterogeneous cohorts.