Can ChatGPT accurately detect atrial fibrillation using smartwatch ECG?

Background: Current guidelines require physician confirmation for smartwatch-diagnosed atrial fibrillation (AF), increasing telemedicine workloads. The newest ChatGPT-4o (GPT-4o) incorporates advanced image input capabilities. Objective: To assess GPT-4o's performance in identifying AF from smartwatch recordings. Methods: Consecutive 120 patients with AF and 60 controls with sinus rhythm (SR), confirmed by conventional 12-lead ECG, recorded single-lead ECGs using an Apple Watch (AW) Series 6 (R). Two blinded cardiologists independently classified the smartwatch recordings as AF, SR, or undetermined. GPT-4o was subsequently prompted to analyze all smartwatch ECGs. Results: Six AF cases were excluded due to undetermined AW-ECG recordings, leaving 114 AF patients (mean age: 73.4 +/- 10.4 years) and 60 controls. The AW algorithm achieved 97.3 % and 100 % accuracy for AF and SR, respectively, while GPT-4o correctly analyzed 47.3 % of AF and 71.6 % of SR tracings. None of the AF characteristics-chronicity, heart rate, QRS width, fibrillatory wave amplitude, or R-wave amplitude and polarity-were predictive of GPT-4o's diagnostic accuracy. Conclusion: The current capabilities of GPT-4o are insufficient to make a reliable diagnosis of AF from smart-watch ECGs. Despite the theoretical appeal of leveraging this innovative technology for such purpose, the findings highlight that human expertise remains indispensable. Consumers must remain aware of the current limitations of this technology.