ResNet-BiGRU with Conditioned Query-Based Cross-Attention and Weighted Loss for Automated Chagas Disease Detection from 12-Lead ECG

Hyuno Im1, Nahyun Lee2, TaeYoung Kang1, Kim TaeHwan1, DONGGUN KIM3, DONGGYU LEE3, Seungsang Oh3, Wuming Gong4, Il-Youp Kwak1
1Chung-Ang University, 2Chung Ang University, 3korea university, 4University of Minnesota


Abstract

Our team CAUETUMN, a participating team in the 2025 PhysioNet/CinC Challenge, investigates whether integrating physiologically interpretable features with deep sequence representations, enhanced by conditioned Query-based cross-attention, can improve Chagas disease detection from standard 12-lead ECGs. We combine three information streams: (i) a 4-layer 1D ResNet backbone for local morphology extraction; (ii) a bidirectional GRU with gated attention for long-range temporal context; and (iii) handcrafted R-peak morphology feature and demographic features (age, sex). The auxiliary features are projected into a high-dimensional query space to conditionally attend over sequence embeddings, enabling selective integration of relevant temporal patterns. Raw ECGs undergo baseline-wander removal with an OC/CO morphological filter. The model is trained on a combination of SaMi-Trop, PTB-XL, and CODE-15\% dataset using a loss function that incorporates both class-specific weights to address label imbalance and group-specific weights to account for dataset-level distribution differences. Our official challenge score on the leaderboard is 0.347.