Artificial Intelligence in Cardiology: A Critical Narrative Review of Current Applications, Clinical Evidence and Translational Priorities

Abstract

Background and Significance: Artificial intelligence (AI) has become one of the most heavily investigated technologies in cardiovascular medicine, with algorithms now reported for electrocardiographic interpretation, echocardiographic quantification, coronary plaque characterisation, arrhythmia detection from consumer devices, and risk prediction from routinely collected clinical data. The volume of publication has grown far more rapidly than the volume of evidence demonstrating patient benefit. Purpose and Scope: This critical narrative review evaluates the strength, consistency and translational maturity of the evidence supporting AI in adult and paediatric cardiology, with particular attention to the gap between reported discriminative accuracy and demonstrated clinical utility. The review covers electrocardiographic AI, cardiovascular imaging, consumer wearables, risk prediction and phenotyping, and generative language models. Literature Selection: Peer-reviewed publications were identified through structured searching of major open scholarly databases and indexes, supplemented by backward and forward citation searching and by examination of authoritative institutional documents. Sources were selected for methodological contribution, evidential weight and relevance to the appraisal questions rather than for citation count alone. Principal Findings: The evidence base is markedly uneven. Electrocardiographic AI for left ventricular systolic dysfunction and for atrial fibrillation risk stratification, and echocardiographic AI for ejection fraction estimation, are supported by prospective or randomised evaluation. Most other applications rest on retrospective, single-centre, internally validated development studies. Where randomised evidence exists, absolute effect sizes are modest and are mediated by clinician behaviour rather than by algorithmic accuracy alone. Reported performance frequently degrades under distributional shift, and external validation in demographically and geographically distinct populations remains sparse. Reporting quality, although improving under dedicated guidelines, remains inconsistent, and regulatory clearance is often granted on retrospective evidence. Unresolved Questions and Implications: It remains unclear whether AI-driven earlier detection of asymptomatic cardiac disease improves long-term outcomes, whether performance is preserved in health systems unlike those in which models were derived, and how algorithmic outputs should be integrated into guideline-directed care pathways. Priorities are pragmatic randomised trials with patient-centred endpoints, prospective multinational external validation, prespecified subgroup performance analysis, and post-deployment surveillance for performance drift.