As part of the George B. Moody PhysioNet Challenge 2025, team MIWEAR developed an approach for detecting Chagas disease from 12-lead electrocardiograms (ECGs). Using ResNet-18 with five-fold cross-validation, we estimated sample confidence and curated a subset by retaining only high-confidence positives and negatives. A large-scale ECG foundation model pretrained on over ten million recordings was then fine-tuned, alongside EfficientNet-B0 and ResNet-18 trained on the curated data. Model predictions were fused by averaging. Cross-validation confirmed that confidence-based sampling improved the performance. The standalone ECG foundation model achieved a Challenge score of 0.379 on the hidden validation set, ranking 7th, underscoring strong transferability under distribution shifts. A fusion model guided by the foundation model reached the highest score of 0.400 on the training set, demonstrating the value of integrating complementary architectures to boost accuracy and reduce variance. These findings show that foundation models provide a reliable backbone, while fusion enhances stability, offering a competitive strategy for ECG-based Chagas disease detection.