Introduction: This study aims to develop a robust algorithm for the automated detection of Chagas disease using 12-lead ECG signals as part of the George B. Moody PhysioNet Challenge 2025. The primary goal of this work is to improve early detection and screening by utilizing ECG data as a quick and affordable alternative to serological testing, which would enable timely diagnosis and treatment of the disease.
Methods: Our team, BuckeyeAI, has developed a deep learning based pipeline for the automated detection of Chagas disease from 12-lead ECG recordings. We first implemented a custom data processing method to efficiently load and preprocess ECG signals. This step includes missing data imputation, lead reordering, resampling to a designated sampling rate, and standardizing signal length. For classification, we designed a 1-D Convolutional Neural Network (CNN) architecture. The model consists of two convolutional layers, each followed by batch normalization, ReLU activation, and max-pooling. The first convolutional layer transforms the 12 input channels into 16 feature maps, while the second increases this to 32. The resulting feature maps are flattened and passed through a fully connected layer with a log-softmax activation to produce the binary classification output.
Results: In the current challenge stage, our classification model achieved an AUC of 0.73 and a True Positive Rate (TPR) of 0.195 during internal cross-validation, along with a TPR of 0.153 on the hidden challenge test sets. We are actively exploring strategies to enhance this performance and refine our approach.
Conclusion: Our initial results show the potential of deep learning methods for detecting Chagas in ECG data. To improve performance, we plan to address class imbalance (<5% positive cases) using data resampling and weighted loss functions. Additionally, we plan to explore attention-based models to better capture ECG features specific to Chagas disease.