Brugada syndrome: a multicenter study on ECG image-based self-supervised learning for non-invasive prediction of sodium channel blocker response

Introduction Brugada syndrome (BrS) is an inherited channelopathy predisposing to ventricular fibrillation and sudden cardiac death. Definitive diagnosis often requires provocation with sodium channel blockers, yet this procedure carries inherent risks. Predicting challenge test outcomes from baseline electrocardiograms (ECGs) could facilitate patient triage; however, most deep learning approaches require digitized waveform signals, which are unavailable as ECGs are in form of printed or scanned images archived. We aimed to develop a deep learning model that predicts challenge test outcomes directly from baseline ECG images for non-invasive pre-screening of patients with suspected BrS. Methods We compared two pretraining strategies for a ResNet18 classifier: (1) domain-specific pretraining using SimCLR self-supervised learning on approximately 187,000 unlabeled ECG images from a health check-up institution (Japan), and (2) general-purpose pretraining on ImageNet. Both models were fine-tuned on 293 labeled ECG images from 290 patients referred for ajmaline challenge testing at a university hospital (UKM, Germany; 87 patients [30.0%] test-positive). Patient-aware cross-validation (5-fold x 3-repeat) was performed with thresholds optimized by Youden's J statistic. External validation used 111 ECG images from 46 patients undergoing pilsicainide challenge testing at a medical university (JMU, Japan; 32 patients [69.6%] positive), independent of the pretraining dataset. Internal AUC comparison used the corrected resampled t-test; external validation employed the DeLong test, McNemar's test, and stratified bootstrap (10,000 replicates). Results On internal cross-validation, the domain-pretrained model achieved a mean AUC of 0.779 ± 0.040 versus 0.710 ± 0.053 for ImageNet (p = 0.031). On external validation, the domain-pretrained model achieved image-level and patient-level AUCs of 0.728 and 0.770, versus 0.427 and 0.462 for ImageNet. Stratified bootstrap confirmed a significant AUC difference at patient level (ΔAUC = +0.308; 95% CI, +0.089 to +0.516; p = 0.006). The domain-pretrained model achieved sensitivity of 0.375, specificity of 1.000, PPV of 1.000, and NPV of 0.412, whereas the ImageNet model yielded zero sensitivity (McNemar p < 0.001) with specificity of 1.000, indicating that it classified all cases as negative. Conclusion Domain-specific self-supervised pretraining on ECG images significantly outperformed ImageNet pretraining, particularly in cross-institutional generalization. The ImageNet model failed in external validation, likely due to overfitting to institution-specific visual artefacts. Operating directly on ECG images rather than digitized waveform signals, as an approach applicable to the ECG format commonly encountered in clinical practice and retrospective archives, may enable non-invasive pre-screening for Brugada syndrome by predicting pharmacological provocation test.