Strong for Atrial Fibrillation, Weak for Other SVTs: ECG Interpretation Using ChatGPT

Abstract 2311410, presented at Western Atrial Fibrillation Symposium 2026

OpenAI's ChatGPT is a widely accessible artificial intelligence platform increasingly used for ECG interpretation by clinicians, healthcare staff, and patients. Despite this growing real-world use, its diagnostic accuracy across supraventricular tachycardias (SVTs) and implications for clinical management remain unclear. We evaluated the diagnostic performance of ChatGPT across common atrial arrhythmias to define its strengths, limitations, and relevance to clinical settings.We retrospectively analyzed 500 ECGs from a single center (100 each: atrial fibrillation [AFib], atrial flutter, sinus tachycardia, atrial tachycardia, and AV nodal reentrant tachycardia [AVNRT]). All rhythms were adjudicated by cardiologists, with electrophysiology study confirmation when available. De-identified raw ECG tracings were uploaded to GPT-5 (OpenAI), which classified each ECG into one of five rhythm categories without class distribution input. Diagnostic performance was assessed using sensitivity, specificity, predictive values, likelihood ratios, Matthews Correlation Coefficient (MCC), Youden index, overall accuracy, and area under the ROC curve, all reported with 95% confidence intervals. Statistical analyses performed using SPSS v30.ChatGPT demonstrated strong performance for AFib, with sensitivity 0.82, specificity 0.95, LR+ 15.2, MCC 0.76, overall accuracy 0.92, and AUC 0.89. Atrial flutter showed moderate performance (specificity 0.93, sensitivity 0.51; MCC 0.46; AUC 0.72). In contrast, sinus tachycardia, atrial tachycardia, and AVNRT demonstrated poor performance, with low positive predictive values (0.14–0.26), low MCC (−0.15 to 0.15), AUCs ≤0.60, and accuracy 45–54%. Misclassifications were systematic, most commonly labeling atrial tachycardia or AVNRT as AFib or atrial flutter. Global diagnostic accuracy across all rhythms was 67%.Open AI's, easily accessible, ChatGPT demonstrates strong capability for AFib detection but performs poorly for more nuanced SVTs. Misclassification of rhythms such as sinus tachycardia, atrial tachycardia, or AVNRT as AFib or atrial flutter may result in inappropriate downstream management, including unnecessary anticoagulation or referral for ablation procedures. While ChatGPT shows promise for AFib identification, its limitations in SVT discrimination underscore the need for clinician oversight and targeted refinement before clinical integration.