Abstract
AIMS: Synthetic electrocardiograms (ECGs) for inherited cardiac diseases may overcome the issue related to data scarcity for artificial intelligence (AI)-based algorithms. This study aimed to evaluate experienced cardiologists' ability to differentiate synthetic and real Brugada ECGs. METHODS AND RESULTS: A total of 2244 ECG instances (50% synthetic generated by a generative adversarial network, 50% real Brugada patients' ECGs) were evaluated by 7 cardiologists, each with >15 years of experience. All ECGs were standard 12-lead recordings acquired with identical settings (paper speed 25 mm/s, amplitude 10 mm/mV) and randomly assigned without identifying markers. The examination was blinded and conducted in 2 rounds with at least 2 h gap between rounds to assess potential learning effects and intra-rater reliability. Each physician classified the recordings as 'real' or 'synthetic' without having any additional information. Performance metrics, including accuracy, sensitivity, specificity, and intra-rater reliability (Cohen's Kappa), were analyzed. Brugada syndrome (BrS) specialists' repeated evaluations were characterized by low accuracy (first round 40%, second round 42%), specificity (first round 22%, second round 26%) and sensitivity (first round 58%, second round 58%). Intra-rater reliability varied widely (Cohen's Kappa: -0.12 to 0.80). CONCLUSION: Synthetic Brugada ECGs cannot be adequately distinguished from real patients' ECGs by BrS specialists.