Multimodal AI in Cardiology: A Critical Review of Diagnostic Performance and Clinical Applicability

Abstract

The article critically examines the diagnostic performance of artificial intelligence systems in the analysis of cardiovascular imaging and biomedical signals, taking into account both validation conditions and methodological limitations that hinder clinical applicability. Findings indicate that image-based studies achieve high accuracy in echocardiographic segmentation, although performance is constrained by poor image quality, small sample sizes, and limited external validation. Signal-based studies report strong performance across multiple applications, yet with variable specificity and notable declines when evaluated in external cohorts. Overall, the evidence suggests meaningful diagnostic potential, albeit conditioned by sample heterogeneity, overfitting, selection bias, reliance on incomplete metrics, and regulatory shortcomings. Such limitations underscore the need for rigorous prospective and external validation, along with continuous monitoring, prior to clinical implementation.