Artificial intelligence vs. cardiologists for PVC localization - still a long way to go
F Salvaterra, D Cazeiro, D Ferreira, M Vilela, J Cravo, J Sabido, I Araujo, J Fernandes Pedro, J Brito, J Ribeiro, A Nunes Ferreira, G Lima da Silva, N Cortez-Dias, F Pinto, J de SousaAbstract
Introduction
Artificial intelligence (AI) has shown promise in electrocardiographic (ECG) interpretation, yet its capacity to accurately localize the origin of idiopathic ventricular arrhythmias remains uncertain. Precise identification of premature ventricular complex (PVC) sites of origin is essential for ablation planning and mechanistic understanding, but the real-world performance of general-purpose AI models for this task is largely unknown.
Purpose
To assess the diagnostic accuracy of AI engines in identification of the origin of idiopathic PVCs from ECG images, and to compare their performance with that of cardiologists (both arrhythmologists and non-arrhythmologists).
Methods
We retrospectively analyzed 12-lead ECG from patients with idiopathic PVCs and structurally normal hearts who underwent invasive EP mapping and ablation. Each ECG was uploaded to ChatGPT® and Microsoft CoPilot®, and both engines were prompted to identify the most likely site of origin. An additional clarifying question was used when PVC recognition seemed uncertain.
The heart was divided into 16 anatomical segments: right ventricular (RV) outflow tract (4), aortic cusps (3), mitral papillary muscles (2), left ventricular (LV) summit (1), and remaining LV and RV wall regions (6). The “gold standard” reference for PVC origin was established by invasive EP study, defined as the site of acute successful ablation (disappearance of the targeted arrhythmia).
Results
A total of 35 patients were included (mean age 53 years, 46% male). Correct identifications occurred by both AI engines in 9 (26%) patients, exclusively within the RVOT region. Mean sensitivity, specificity, and accuracy were 26%, 95%, and 91%, respectively. For cardiologist interpretation, corresponding values were 46%, 96%, and 93%, with no statistically significant difference compared with ChatGPT® (p=0.051) or CoPilot® (p=0.07). The trend toward better human performance did not reach statistical significance, likely due to sample size limitations.
Conclusion
Both ChatGPT® and Microsoft CoPilot® showed limited sensitivity and spatial discrimination for localizing idiopathic PVC origins, performing below experienced clinicians and failing outside the RVOT. Despite high specificity, these models remain unsuitable for clinical localization tasks, underscoring the need for domain-specific ECG datasets and dedicated training. The coexistence of multiple ECG-based localization algorithms may further hinder accuracy, as general-purpose AI cannot yet identify or apply the appropriate diagnostic criteria.