Predicting repeat percutaneous coronary intervention more than 30 days after index PCI: exploratory internal validation of clinical and machine-learning models

Abstract Background Repeat percutaneous coronary intervention (PCI) may reflect staged treatment, restenosis, progression in another vessel, or a new coronary event. We evaluated whether information available during the index hospitalization could identify patients with a repeat PCI record more than 30 days after index PCI. Methods This single-center retrospective study retained the original PCI inclusion algorithm and required at least 30 days of potential administrative follow-up. A repeat PCI in a distinct encounter more than 30 days after index PCI was the outcome; procedures within 30 days were treated as potentially staged and were not counted as events unless a later repeat PCI occurred. Clinical risk factors, comorbidities, index-PCI descriptors, medications, and index-encounter laboratory means were considered. Patients were assigned by a fixed stratified 70/30 split. Predictor screening, block selection, model tuning, score-size selection, and probability calibration were confined to the training data. A chronological 70/30 split was performed as a temporal sensitivity analysis. Results Among 1,079 patients in the original PCI cohort, 751 had at least 30 days of potential follow-up and 50 (6.7%) had repeat PCI after 30 days. The random split included 526 training patients (35 events) and 225 validation patients (15 events). A training-selected, equal-weight 12-variable score had a validation area under the receiver operating characteristic curve (AUC) of 0.711 (95% confidence interval [CI], 0.574–0.849), area under the precision–recall curve (AUPRC) of 0.131 (95% CI, 0.085–0.301), Brier score of 0.0608, and calibration slope of 0.906. LASSO logistic regression had a lower AUC (0.685; 95% CI, 0.538–0.832) but a higher AUPRC (0.193; 95% CI, 0.080–0.393). In the temporal analysis, the pipeline independently selected a two-variable score ( $$K=2$$ K = 2 ), which had an AUC of 0.517; the best model, LightGBM, had an AUC of 0.544 (95% CI, 0.365–0.723). Median potential follow-up was 89.2 days in the earlier training cohort and 44.1 days in the later validation cohort, so this comparison was not a fixed-horizon validation. Conclusions The 12-variable score showed moderate discrimination in an exploratory random validation set, but the confidence interval was wide and performance did not persist in temporal validation. These models are not ready to guide treatment. External validation with adjudicated lesion-level outcomes, longer follow-up, and complete anatomical data is required.