Effectiveness of AI Models for Cardiovascular Disease Diagnosis: A Systematic Review and Meta-analysis

Abstract

Introduction. Cardiovascular disease represents a substantial global health burden, and delays in diagnosis can further compromise clinical outcomes. Objective. This review aimed to evaluate the diagnostic effectiveness of artificial intelligence models across major cardiovascular conditions, with a specific focus on coronary artery disease, acute coronary syndromes and myocardial infarction, and heart failure. Method. Following PRISMA guidelines, we conducted a systematic review and applied meta-analytical techniques to pool measures of diagnostic performance. Random effects models with restricted maximum likelihood and Hartung–Knapp adjustments were used to calculate summary estimates, and analyses were stratified by cardiovascular condition. Only effects with variance (reported 95% CI or derivable denominators) were pooled. Areas under the receiver operating characteristic curve without variance were synthesized narratively. Results. Out of 39,383 records, thirty-five studies were eligible. The pooled area under the receiver operating characteristic curve was 0.823 (95% confidence interval 0.754–0.892). Heterogeneity was considerable (I² = 98.4%), and the 95% prediction interval was 0.59 to 1. However, 9 of 35 included studies reported external/independent validation. The pooled estimate therefore represents an overall descriptive average, reflecting the expected range of diagnostic performance rather than a single universal value. Stratified results showed pooled values of 0.88 for acute coronary syndromes and myocardial infarction, 0.87 for heart failure, and 0.86 for coronary artery disease. Sensitivity often reached 0.90 or higher in electrocardiography deep learning models, while specificity remained variable. Conclusion. Artificial intelligence models demonstrated promising discrimination across cardiovascular conditions, with variable specificity and predictive value, indicating that clinical utility requires external validation and local calibration.

Ähnliche Arbeiten