Percutaneous coronary intervention (PCI) is a cornerstone of coronary artery disease management, yet predicting post-procedural outcomes remains challenging. Machine-learning (ML) models applied to cardiac imaging have emerged as tools to improve risk stratification and prognostic assessment. To systematically evaluate the performance of ML models derived from cardiac imaging, primarily coronary CT angiography (CCTA) and single-photon emission computed tomography (SPECT), for predicting PCI-related clinical outcomes, including major adverse cardiovascular events (MACE), mortality, repeat revascularization, procedural success, and functional recovery. PubMed, Scopus, and Web of Science were searched for studies developing or validating ML models using CCTA and/or SPECT to predict PCI-related outcomes. Extracted performance metrics included area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. Risk of bias was assessed using PROBAST. Where methodologically appropriate, random-effects meta-analyses were performed within modality-specific and endpoint-specific strata. Ten studies met inclusion criteria, nine of which contributed quantitative data. Most models were CCTA-based, with fewer SPECT-only or combined-modality approaches. Across pooled analyses, ML models demonstrated strong discriminative performance for PCI-related outcomes, with pooled estimates of accuracy 0.89, sensitivity 0.78, specificity 0.86, and AUC 0.87. CCTA-based models showed more consistent performance across outcomes, while SPECT-based models demonstrated promising but more heterogeneous results. ML models applied to cardiac imaging show substantial potential for predicting PCI-related outcomes, particularly for CCTA-based approaches. Evidence for SPECT-based models remains emerging and heterogeneous. Standardized outcome definitions, modality-specific reporting, and prospective validation are essential before routine clinical implementation.
Erabi et al. (Sun,) studied this question.