Chagas disease is a leading cause of non-ischemic cardiomyopathy, and its long asymptomatic phase makes timely case detection challenging. We develop an end-to-end deep learning model for automated screening from 12-lead ECGs and compare two inputs: ECG-only and a dual-domain input that combines ECG with per-lead power spectral density (PSD) to capture complementary frequency information. We also propose a lightweight multi-scale CNN-Transformer model that couples local morphology with longer-range temporal patterns. To overcome the severe class imbalance in Chagas datasets, we apply age- and sex-aware negative sampling that is repeated during training. On a balanced validation set, the dual-domain model outperformed the ECG-only baseline across AUROC, AUPRC, F1, and the Challenge score. On a held-out test set with the original prevalence, it achieved AUROC 0.92, AUPRC 0.77, F1 0.68, and a Challenge score of 0.23. The official phase test reported a score of 0.317. These results indicate that PSD adds learnable features and improves sensitivity over ECG alone. Furthermore, targeted negative sampling enhances model generalization. The compact CNN-Transformer achieves competitive performance with lightweight computation, supporting practical large-scale screening.