AI-Assisted ECG Interpretation Systems

Abstract

The 12-lead electrocardiogram remains the most widely performed cardiac diagnostic test, with over 300 million recorded annually worldwide, yet interpretation accuracy varies dramatically depending on who is reading it. Cardiologists achieve 75-85% diagnostic accuracy on complex tracings, emergency physicians 55-70%, and general practitioners 35-55% -- a variability that translates directly into missed diagnoses, delayed treatment, and avoidable patient harm. AI-assisted ECG interpretation promises to narrow this gap by providing consistent, real-time algorithmic analysis that augments human expertise. We evaluated five AI-ECG system architectures -- rule-based expert systems, traditional machine learning classifiers, convolutional neural networks, recurrent neural networks, and transformer-based models -- across four clinical interpretation tasks: rhythm classification (12 rhythms), acute MI detection, left ventricular dysfunction screening, and multi-label 24-condition diagnosis. Our evaluation used 186,422 ECGs from three sources: the PTB-XL dataset (21,837 ECGs), an institutional archive from three European hospitals (Madrid, Paris, and Rome; 142,485 ECGs), and a prospective validation set of 22,100 ECGs collected between January and June 2021. We developed an ECG-AI Performance Index (EAPI) integrating diagnostic accuracy, clinical concordance with cardiologist consensus, false-critical rate, interpretability, and processing speed. The transformer-based model achieved the highest EAPI (0.932), with 93.2% accuracy on rhythm classification and AUROC 0.96 for acute MI detection. The CNN architecture (EAPI 0.908) offered the best accuracy-to-speed trade-off for real-time bedside deployment.