The GAN-augmented supervised model achieved an F1-Score of 81.4% and AUROC of 94.7% for Brugada Syndrome detection, outperforming the baseline without augmentation.
Does GAN-augmented supervised learning or self-supervised learning improve the automated detection of Brugada Syndrome from ECGs?
GAN-augmented supervised learning and self-supervised learning can improve the automated detection of rare cardiac diseases like Brugada Syndrome from ECGs by addressing data scarcity.
Absolute Event Rate: 0% vs 0%
Abstract Background Brugada Syndrome (BrS) is a rare hereditary cardiac disorder, with an estimated annual incidence of 1–5 cases per 10,000 individuals. Its diagnosis remains challenging due to subtle electrocardiographic (ECG) manifestations and its inherently low prevalence. Recent advances in deep learning (DL) offer promising opportunities for the automated detection of BrS; however, limitations related to data scarcity and class imbalance persist. Purpose This study investigates recent advancements in deep learning by evaluating two distinct frameworks for the challenging task of BrS detection. First, we examine a supervised learning approach enhanced with synthetic data augmentation using generative adversarial networks (GANs) to address data scarcity. Second, we assess a self-supervised learning (SSL) approach, where deep learning models first learn general ECG representations from large unlabeled datasets and are then fine-tuned for BrS detection. Methods We employ two datasets: (1) a Brugada dataset, with ECG recordings from 87 patients diagnosed with BrS and 207 control subjects (adults ≥ 18 years), comprising a total of 23,390 4-second ECG segments; (2) a combination of two open-source datasets (MIMIC-IV-ECG and computing in cardiology challenge 2020) reaching a total of 1,723,398 4-second ECG segments. To address class imbalance, we applied GAN-based data augmentation, training two architectures for the synthesis of Type I and Type II ECG patterns. The dataset was incrementally augmented with 25%, 50%, and 75% synthetic Brugada data. We investigated two loss functions for SSL pretraining: InfoNCE, based on contrastive leanring, and VICReg. For both supervised and SSL-based frameworks, we implemented a one-dimensional convolutional neural network architecture inspired by EfficientNet. Results The supervised EfficientNet model, trained with a 75% GAN-augmented Brugada data, achieved the best performance ( F1-Score of 81.4 ± 2.0% and an AUROC of 94.7 ± 0.8% on the test set), outperforming the baseline model without augmentation (F1-Score: 79.7 ± 1.8%, AUROC: 93.5 ± 0.7%). The best-performing SSL model was pretrained using the InfoNCE loss function, achieving an F1-Score of 78.3 ± 2.8% and an AUROC of 94.0 ± 1.7% after fine-tuning on the Brugada dataset. In comparison, VICReg-pretrained models reached an F1-Score of 75.6 ± 4.2% and an AUROC of 93.0 ± 1.9%. Conclusion Both techniques, GAN-augmented supervised learning and self-supervised learning, can improve automated Brugada Syndrome detection. GAN-based augmentation achieved the best performance, while self-supervised learning proved promising for scalable ECG representation in rare disease settings.Deep Learning architectures.Results Table
Zanchi et al. (Thu,) reported a other. The GAN-augmented supervised model achieved an F1-Score of 81.4% and AUROC of 94.7% for Brugada Syndrome detection, outperforming the baseline without augmentation.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: