Deep Learning Architectures for Brugada Syndrome Detection: a Comparative Analysis with GAN-Generated ECG Data
This HF repository contains the pretrained models developed in our paper: "Deep Learning Architectures for Brugada Syndrome Detection: a Comparative Analysis with GAN-Generated ECG Data".
Authors: Beatrice Zanchi, Giuliana Monachino, Giulio Conte, Francesca D. Faraci
Abstract
This study investigates the potential of deep learning models for the automated detection of Brugada syndrome, a rare and life-threatening cardiac condition. Three neural network architectures—Conformer, EfficientNet, and ResNet18— are exploited using data from 87 Brugada syndrome patients and 207 controls. Given the limited availability of real-world electrocardiogram data for Brugada syndrome, we investigate the benefits of synthetic GAN-generated data to address data scarcity. Without data augmentation, EfficientNet emerged as the most balanced performer, (F1-score = 79. 7%, AUROC = 93.5%), while ResNet18 and Conformer exhibited trade-offs between sensitivity and precision. The enhancement of training data with synthetic Brugada samples improved performance in both Conformer (F1 = 79.4%, AUROC = 92.3%) and EfficientNet (F1 = 81.4%, AUROC = 94.7%), although ResNet18 showed a decrease in precision with an increased level of synthetic data. These results emphasize the importance of selecting an appropriate model architecture and augmentation strategy to optimize performance in automated medical diagnosis. Our findings highlight EfficientNet as a robust and effective model for detecting Brugada syndrome.
Github repository
Official code implementation can be found at:
https://github.com/MeDiTech-SUPSI/brugada_detection