Large-scale deployment of AI-enhanced ECG interpretation in telemedicine

Background Artificial intelligence (AI)–driven approaches to electrocardiogram (ECG) interpretation hold promise for enhancing diagnostic accuracy and reducing over-read times. Telemedicine has expanded rapidly in recent years and provides an ideal platform for scaling AI-based ECG tools to deliver remote ECG interpretation. However, real-world data on the performance of AI-enhanced ECG models — particularly with respect to physician overreads — remain limited. Purpose To evaluate real-world performance of an AI-based ECG interpretation system integrated within a telemedicine platform. Methods An AI-based ECG interpretation system was integrated into a telemedicine platform adopted by community pharmacies across multiple regions of Italy. Standard 12-lead ECGs were obtained onsite at each pharmacy, automatically analyzed by the AI algorithm in real time and subsequently overread by remote cardiologists. We retrospectively analyzed ECGs recorded between October 2024 and February 2025. Outcomes included the time from report opening to cardiologist overread completion and diagnostic agreement. Agreement between AI-based ECG interpretation and cardiologists was assessed using Cohen’s kappa statistic. ST-segment elevation myocardial infarction (STEMI) criteria were defined based on age- and sex-specific ST-segment elevation thresholds, in accordance with established guidelines. Results A total of 14,762 ECGs of acceptable quality were analyzed, with similar distribution across ages 18 to 44 (34.6%), 45 to 65 (32.1%), and over 65 (33.3%). The median cardiologist overread time was 18.7 seconds (IQR 9.5–34.8 seconds). ECG measurements remained unchanged in 94.7% of cases. Diagnostic agreement between the AI platform and cardiologists, as measured by Cohen’s kappa, was 0.96 for atrial fibrillation, 0.91 for atrial flutter, 0.87 for RBBB, 0.91 for LBBB, and 0.75 for higher-degree AV block. Contemporary STEMI criteria were met in 95 (0.6%) of cases, all of which were determined to be benign and not indicative of acute coronary syndrome. The AI model for STEMI (and STEMI-equivalent) detection correctly ruled out 94 of these cases (99%), generating a positive read in only 0.2% of all ECGs. Conclusions In this large-scale telemedicine implementation, an AI-based ECG interpretation system facilitated rapid physician overreads and showed high concordance with physician interpretation. These findings support the feasibility of integrating AI-driven ECG analysis into remote clinical workflows, suggesting the potential to broaden access to cardiac diagnostic services.