Machine learning models for predicting coronary in-stent restenosis after percutaneous coronary intervention: A systematic review and meta-analysis

Abstract

Background: Machine-learning (ML) models are increasingly used to predict coronary in-stent restenosis (ISR), but evidence combines prognostic, diagnostic-radiomics, and imagereconstruction tasks and often lacks uncertainty around area-under-the-curve (AUC) estimates. We conducted a task-aware systematic review and meta-analysis with study-level auditing of AUC provenance. Methods: PubMed, Embase, and Scopus were searched through 3 November 2024. Tabular models predicting subsequent coronary ISR after PCI were pooled within model families using logit-AUC random-effects REML, with one estimate per cohort and family. Missing 95 confidence intervals (CIs) were reconstructed using Hanley-McNeil variance and the logit-delta method when validation/test case and non-case counts were available or transparently approximated. Modified Hartung-Knapp and sensitivity analyses assessed robustness; diagnostic radiomics was analyzed separately. Results: Twelve studies represented 16,964 nominal participants/observations. Pooled AUCs for subsequent ISR prediction were 0.84 (95 CI 0.67-0.94; I2 = 98.4) for random forest (RF), 0.74 (0.71-0.77; I2 = 0) for logistic regression (LR), 0.59 (0.46-0.71; I2 = 23.1) for support-vector machines, and 0.73 (0.70-0.76; I2 = 0) for DNN/MLP. The modified Hartung-Knapp RF interval widened to 0.52-0.96. Restricting analysis to independent patient-level cohorts reportingconventional published 95 CIs yielded AUCs of 0.89 (0.54-0.98) for RF and 0.74 (0.70-0.77) for LR. RF subgroup AUCs were 0.91 (0.70-0.98) for DES-only and 0.72 (0.71-0.73) for mixed BMS/ DES cohorts. Age and lipid-related variables appeared in 50 of studies. Conclusions: RF showed the highest discrimination but substantial instability, whereas LR was more consistent. Prospective external validation is required before algorithm ranking or clinical deployment.