As part of the classification of Chagas disease for the George B. Moody PhysioNet Challenge 2025, we have developed a deep learning approach for classifying Chagas disease based on electrocardiograms (ECG). We achieved a challenge score of 0.690 (Submission ID 1235) using the Short-Time Fourier Transform (STFT) in conjunction with a 2D Convolutional Neural Network (CNN), highlighting the potential of our methodologies.
We are excited to explore a hybrid approach that CNNs and transformers, leveraging the strengths of both architectures. This framework will incorporate the Stockwell transform for feature extraction, which provides a time-frequency representation of the ECG signals. This transformation is particularly valuable for capturing the dynamic frequency content of the signals, crucial for identifying specific patterns associated with Chagas disease.
Additionally, we plan to utilize pre-trained models, allowing us to leverage existing knowledge from similar tasks. By fine-tuning these models on our dataset , we aim to enhance their performance and reduce training time. We will systematically evaluate various pre-trained models to identify those that yield the best results.
We also intend to implement unsupervised learning with autoencoder and clustering algorithms, to create meaningful embeddings. Analyzing this data will help uncover specific features indicative of Chagas disease, enhancing the following models ability to differentiate between normal and abnormal signals.
Our training strategy will minimize cross-entropy loss, adjust class weights, and employ the AdamW optimizer over 30 epochs, with early stopping to prevent overfitting. A 5-fold cross-validation will be performed, and the final model will be an ensemble of the best-performing models from each fold, aiming to improve the diagnostic accuracy of Chagas disease significantly.