Reliable Pseudo-Labeling for semi-supervised Chagas Diagnosis from Noisy-Labeled ECG

Weifeng Liu, Qiu Miaohan, Li Jing, Li Yi, Han Yaling
Department of Cardiology, General Hospital of Northern Theater Command


Abstract

Background: Chagas disease remains a significant public health concern in America. Although serological testing is the gold standard for diagnosis, its high cost and labor-intensive nature hinder large-scale screening efforts. Electrocardiography (ECG) presents a promising, low-cost, non-invasive, and widely accessible alternative. However, the disease's low prevalence (<5%) results in severe class imbalance, complicating the development of accurate ECG-based diagnostic models. To address this challenge, we propose a three-stage semi-supervised learning framework for the George B. Moody PhysioNet Challenge 2025. Methods: In Stage 1, a Multi-Scale Deep Neural Network (MSDNN) is trained using five-fold cross-validation on two high-quality labeled datasets: Sami-Trop and PTB-XL. Ensemble predictions are employed to assign confidence scores to noisy-labeled samples from the Code-15% dataset. In Stage 2, the top 50% of Code-15% samples with the highest confidence scores are selected as pseudo-labeled data and incorporated into the training set, while the remaining samples are set aside for validation. In Stage 3, the final model is trained on the combined dataset (Sami-Trop, PTB-XL, and Code-15%) and evaluated on the official online server. The MSDNN architecture consists of eight multi-scale, multi-channel convolutional blocks integrated with batch normalization, dropout, and squeeze-and-excitation (SE) modules to effectively extract discriminative ECG features. Result: Our team AI-IN-Cardio's approach achieved a local validation Challenge Score of 0.612 and an official online score of 0.527, demonstrating the framework's potential in Chagas disease screening. Conclusion: Our semi-supervised learning framework offers a promising solution for leveraging noisy and imbalanced ECG data to enhance Chagas disease diagnosis. Despite the challenges posed by class imbalance and limited labeled data, the proposed method demonstrates a strong ability to improve diagnostic performance.