As part of the George B. Moody PhysioNet Challenge 2025, our team ECG_ChD_25 developed an open-source deep neural network (DNN) classifier based on 1D ConvNeXt model to detect potential Chagas cases from standard 12-lead ECGs. Training data included three labeled ECG datasets: SaMi-Trop (Chagas positive), PTB-XL (Chagas negative) and the Chagas negative subset of CODE-15% (self-reported). To address highly imbalanced training data and label noise, we combined Focal loss with a pair-wise ranking loss, applied class-weighting, and oversampled minority class to 50% per epoch. We trained the model with AdamW for 8 epochs (batch size 64, initial learning rate (LR) 2e-4, weight_decay 1e-3, dropout rate 0.3) and maintained an exponential moving average (EMA) of the weights (decay = 0.999). We used cosine annealing LR scheduler (T_max = 8 epochs, minimum LR = 5e-6) to enable more effective convergence. A 5-fold cross validation was employed, and the final model was chosen as a logit-averaged ensemble of the best-performing checkpoint from each fold. Our classifier achieved a Challenge Score (Recall@5%) of 0.372 (ranked 10th out of 65 teams) on the hidden validation set. This study demonstrates the potential of deep learning for point-of-care Chagas disease screening where serological testing capacity is limited.