The 2025 George B. Moody PhysioNet Challenge focuses on detecting Chagas disease from standard 12-lead electrocardiogram (ECG) recordings, where the main difficulties include heterogeneous data sources, weak versus strong labeling, and serious class imbalance. We propose a deep learning framework based on a Res2Net–Transformer backbone that integrates curriculum learning and hard sample mining to improve both convergence and sensitivity to rare cases. The architecture consists of a main classification head for Chagas prediction, auxiliary branches for demographic and physiological variables (age, sex, heart rate, QRS duration) trained with proxy constants acting as weak regularizers, and an adaptive threshold head that refines the decision boundary. The composite loss function was designed to guide learning under imbalance: focal loss emphasizes difficult positive cases by down-weighting easy negatives, ranking loss improves the ordering of predictions to enhance sensitivity in the top-risk subset, auxiliary losses act as regularizers to stabilize shared representations, and threshold regularization constrains the adaptive head toward a clinically meaningful prior. Training follows a curriculum that progresses from easy to hard samples, while hard sample mining dynamically emphasizes high-loss cases to strengthen discriminability under borderline conditions. Finally, we achieved a Challenge score of 0.299 (ranked 31th out of 66 teams) on the hidden validation set.