Introduction: We present ElectroDoChagas's team contribution to the CinC-PhysioNet 2025 challenge to detect Chagas disease on 12-lead electrocardiogram (ECG) signals.
Materials & Methods: Three datasets were provided. To account for class imbalance, dominant class was divided into three subsets and the recessive class was reproduced across all of them. Each subset was used to train a model integrated into a major voting ensemble. We focused on mitigating database, age, and sex-dependence biases, by stratifying them during subset, validation and holdout-test splitting. PTB-XL database was excluded during training.
Signals were resampled to 400 Hz, standardized and zero-padded or crippled to 3000 samples. To mitigate overfitting lead inversion, interchange, and Gaussian noise was applied with probabilities from 0.02 to 0.06. A validation and holdout test of 25% and 20% of train and total data was created, respectively.
The model was a deep Residual Network (ResNet) with 8 residual blocks enhanced with Convolutional Block Attention Modules (CBAM). The 1D convolutional layers had a kernel size of 15 at outer and 7 at deeper layers with L2 regularization and 64 filters that doubled every 2 blocks. A dropout layer of 0.3 was added at each residual block. Output was provided by a dense layer with sigmoid activation. While CBAM applied lead and time attention mechanisms, residual connections allowed a deeper model and more complex feature extraction.
Adam optimizer, weighted focal loss, learning rate of 0.001 with a 0.1 factor decrease after 5-epoch F1-score validation plateaus and early stopping with patience 10 was used for training for 100 epochs, batch size of 64.
Results: Last model version obtained a local validation challenge score (CS) of 0.664. A preliminary submitted version obtained a local validation CS of 0.401 and of 0.442 on the hidden test set (provided by the Challenge submission system).