AI models show high accuracy for myocardial infarction diagnosis in meta-analysis

A systematic review and meta-analysis assessed validation and test cohorts from studies of artificial intelligence (AI)-based algorithms for diagnosing acute myocardial infarction (AMI). The 24 included studies evaluated 27 algorithms in 166,052 patients; 18,850 (11.3%) had a final AMI diagnosis. The analysis examined sensitivity and specificity overall and separately for STEMI and NSTEMI, as well as by AI approach and data source.

Pooled performance for AMI was an area under the curve (AUC) of 0.97, with sensitivity of 0.93 and specificity of 0.92. For STEMI, the AUC was 0.96 (sensitivity 0.91, specificity 0.94); for NSTEMI, it was 0.97 (sensitivity 0.95, specificity 0.90). Performance did not differ significantly between deep-learning and other AI approaches. ECG-based models were less sensitive, while troponin-based models were less specific.

The findings indicate potential for AI to complement existing diagnostic pathways, rather than replace them. Direct head-to-head comparisons with conventional statistical models were limited, and the authors call for prospective studies comparing AI and conventional approaches in the same populations. The reported accuracy alone does not establish improved clinical outcomes or support changing current diagnostic practice.