Aims: Chagas disease affects millions and causes significant mortality annually, with no vaccine available; early detection is crucial. This study, developed for the CinC 2025 Challenge, aims to develop and evaluate a machine learning model for detecting Chagas disease using only a single electrocardiogram (ECG) lead. This approach is designed for potential application with low-cost, portable single-lead ECG devices to facilitate widespread screening in resource-limited settings where the disease is prevalent.
Methods: I trained a Convolutional Neural Network (CNN) using the PTB-XL and SaMi-Trop datasets with provided high-confidence Chagas labels. The training set comprised 503 remaining SaMi-Trop patients balanced by 503 patients sampled from PTB-XL for a total of 1006; 300 patients from both datasets were reserved for validation. To align with potential application in low-cost portable devices, only Lead I (extracted from standard 12-lead ECGs) was utilised. Since Chagas cardiomyopathy often involves conduction abnormalities, specificity was assessed against independent cohorts derived from PTB-XL labels: right bundle branch block (RBBB), atrioventricular block (AVB), premature ventricular contractions (PVC) and normal sinus rhythm (NSR).
Results: The model achieved promising performance on the local validation set, with an Area Under the Curve (AUC) of 0.91, a specificity of 0.93 and a sensitivity of 0.76. The overall challenge score obtained on the hidden validation set was 0.64. Specificity tests demonstrated robustness against common confounding conditions, yielding scores of 0.89 against the RBBB group, 0.88 against AVB, 0.88 against PVC and 0.94 against the NSR group.
Conclusion: This study demonstrates the feasibility of using a single-lead ECG for Chagas disease detection. This approach holds potential for improving early detection and access to care, particularly in resource-constrained areas.