Document resource
Introduction The electrocardiogram (ECG) remains one of the most widely used investigations in cardiology, forming the basis of diagnostic and management decisions across multiple cardiac conditions. Accurate ECG interpretation requires clinical experience and remains subject to inter-observer variability. More recently, artificial intelligence (AI) based ECG interpretation tools have been developed and increasingly evaluated as potential aids to clinical diagnosis. This review synthesises recent evidence on the diagnostic performance of AI-based ECG interpretation relative to clinician assessment across a range of cardiac and ECG-detectable conditions.Methods A structured literature review was conducted using MEDLINE and Web of Science to identify studies published from 2021 onwards assessing AI-based ECG interpretation against clinician performance and/or expert-validated reference standards. A PICO framework was used to define eligibility criteria. Studies involving adult populations that reported diagnostic performance outcomes, including sensitivity, specificity, accuracy or area under the receiver operating characteristic curve (AUROC), agreement with reference standards, or measures of clinical efficiency and decision support, were included. Records were screened, duplicates removed, and eligible studies included for qualitative synthesis.Results Across the included studies, AI-based ECG interpretation demonstrated diagnostic performance comparable to or exceeding clinician interpretation across multiple ECG-detectable conditions. In rhythm classification tasks, deep-learning models achieved high diagnostic accuracy, with reported AUROC values up to 0.987 and F1 scores up to 0.97, outperforming cardiologists, emergency physicians, and internists in several rhythm categories.For acute coronary syndromes, AI models demonstrated strong performance for STEMI and occlusive myocardial infarction, often exceeding clinician diagnostic accuracy. Domain-specific ECG AI models demonstrated more balanced performance than general purpose large language model-based tools.AI-based ECG interpretation also showed utility in detecting cardiac abnormalities and in supporting clinical workflows, demonstrating near-perfect agreement with physician-derived atrial fibrillation burden measurements, reduced interpretation time, and improved guideline-consistent anticoagulation recommendations. Most studies were retrospective and single-centre, with limited prospective validation.Conclusion Current evidence suggests that AI-based ECG interpretation can match or outperform clinician interpretation across multiple diagnoses and improve clinical efficiency. However, important limitations remain, highlighting the need for prospective, multi-centre evaluation before widespread clinical implementation. Collaboration between AI-based ECG systems and clinician expertise offers a promising approach to improving diagnostic accuracy and clinical efficiency at scale.Abstarct9 Table 1Characteristics of included cohort studies comparing pregabalin and gabapentinStudyStudy populationBaseline HF statusHF outcome assessedEffect estimatesHo et al., 2017Seizure disorder patients initiating gabapentinoids(total n = 246,237)Mixed (stratified by HF history)HF hospitalisation or ED visit within 90 daysUnadjusted HRsLund et al., 2020Established HF patients initiating gabapentinoids(total n = 2,790)Established HFWorsening HF within 90 days (HF hospitalisation or death)Adjusted HRsPark et al., 2020Chronic non-cancer pain patients initiating gabapentinoids(total n = 19,710)No prior HFHospital admission or ED visit with HF (de novo HF)Adjusted HRsAbbreviations: HF, heart failure; HR, hazard ratio; ED, emergency department.