Diagnostic accuracy of artificial intelligence (AI)-based models for acute myocardial infarction: a systematic review and meta-analysis.
TL;DR
AI-based diagnostic algorithms for AMI diagnosis were associated with particularly efficient performance and has the potential to enhance existing tools by identifying subtle ECG features and integrating multiple data streams.
Abstract
Aims
Despite continuous improvements in diagnostic tools, the diagnosis of acute myocardial infarction (AMI) is still based on electrocardiogram (ECG) and cardiac biomarkers supported by other non-invasive imaging techniques. However, ECG has limited sensitivity, while high-sensitivity troponin assays may show reduced specificity, particularly in older or multimorbid patients. Widely adopted clinical pathways (e.g. ESC 0/1 h algorithm) have reported high diagnostic discrimination in external cohorts (with AUC ranging from 0.93 to 0.96) which provides clinical context for interpreting our findings. These diagnostic tests have limitations that can be addressed through automated computer-based diagnostic systems using Artificial Intelligence (AI). Several studies in the literature have presented machine learning (ML) or deep learning (DL) models to quickly and accurately assess the diagnosis of AMI to optimize treatment.
Methods
AND
Results
The aim of our analysis was to evaluate the performance of AI-based models for the diagnosis of AMI using data from all validation and test cohorts available in literature. Sensitivity and specificity for all MIs, and for STEMI and NSTEMI separately, were analysed for different types of AI approach (non-DL or DL) and for different data sources (ECG or troponin-based). Hierarchical models were used to estimate the summary receiver operating characteristic curve, sensitivity and specificity.A total of 24 studies evaluating 27 AI-based diagnostic algorithms were included encompassing 166.052 patients, of which 18.850 (11.3%) had a final diagnosis of AMI. Patients were predominantly male (61.8%) with a median age of 60.7 (IQR 59.0-64.1) years. Discrimination of AI-based algorithms for the diagnosis of AMI was high with an AUC of 0.97 (sensitivity 0.93, specificity 0.92). This result remained consistent when considering the ability to diagnose STEMI (AUC 0.96, sensitivity 0.91, specificity 0.94) or NSTEMI (AUC 0.97, sensitivity 0.95, specificity 0.90) separately. No significant difference was found regarding type of AI. ECG and troponin-based models were both extremely accurate, with the first ones being less sensitive and the latter less specific.
Conclusions
AI-based diagnostic algorithms for AMI diagnosis were associated with particularly efficient performance. These results were consistent for both STEMI and NSTEMI considered separately. Rather than replacing conventional diagnostic pathways, AI has the potential to enhance existing tools by identifying subtle ECG features and integrating multiple data streams. While direct head-to-head comparisons with traditional statistical models are currently limited, future prospective studies evaluating AI and conventional approaches within the same clinical populations are warranted.AI did not lose accuracy depending on the data source or the type of AI applied.