The George B. Moody PhysioNet Challenge 2025 focused on detecting potential cases of Chagas disease from standard 12-lead electrocardiograms (ECGs). Our team, Kust_MeAI, proposed an approach to predict Chagas disease by combing an 18-layer residual neural network and a joint loss consisting of focal loss and precision loss. Firstly, we retain all samples from the SaMi-Trop and PTB-XL datasets and select a portion of samples from the CODE-15% dataset as the training set. All selected recordings then underwent the same series of preprocessing steps including denoising, cropping and normalization. Secondly, we built an 18-layer residual network with multi-head self-attention to extract the Chagas disease's complex pathological patterns from ECG records. Thirdly, to alleviate the severe class imbalance problem and achieve accurate detection of Chagas disease, we perform mixup data augmentation operation on all samples and then jointly optimize the proposed model using focal loss and a custom precision loss. Preprocessed 12-lead ECG segments were used as model inputs for end-to-end training, and the prediction probabilities of positive and negative categories were produced as model outputs. We evaluated the proposed approach on the public training datasets and achieved an average 5-fold cross-validation Challenge score of 0.216.And we achieved a challenge score of 0.201 on the hidden validation dataset (ranked 49th out of more than 66 teams). Finally, our proposed approach received the Challenge score of X.XX (ranked XXth out of XX teams) on the hidden test set.