Predicting long-term major adverse cardiac events after percutaneous coronary intervention in individuals with acute coronary syndrome and nonvalvular atrial fibrillation: a machine learning study.

backgroundThis study was designed to identify risk factors for long-term major adverse cardiac events (MACE) in individuals with high-risk acute coronary syndrome (ACS) and concomitant nonvalvular atrial fibrillation (NVAF) after percutaneous coronary intervention (PCI), and to present these findings in a visual predictive format.

methodsThe least absolute shrinkage and selection operator (LASSO) regression was applied to identify predictors of long-term MACE. Using these predictors, three models were developed: a Cox proportional hazards regression model, a random survival forest model, and a CoxBoost model. Model performance was assessed using receiver operating characteristic curves, Kaplan-Meier survival analysis, and the net reclassification improvement (NRI). The optimal model was subsequently visualized in a nomogram.

resultsOver a median follow-up period of 36 months, 159 out of 754 participants experienced MACE, including 58 deaths. All variables met the proportional hazards assumption, with no evidence of multicollinearity. In the training set, the C-index values for the LASSO-Cox, LASSO-random survival forest, and CoxBoost models were 0.791, 0.784, and 0.815 respectively; corresponding values in the validation set were 0.736, 0.714, and 0.725. The CoxBoost model achieved the highest C-index; however, given its inclusion of 18 variables, the more parsimonious LASSO-Cox model with five variables—diabetes, number of vessel lesions, urea, Gensini score, and GRACE score—was selected. Compared with the GRACE score alone, the LASSO-Cox model demonstrated superior predictive performance, with both NRI values below 0.

conclusionA nomogram derived from a LASSO-Cox regression model incorporating five key predictors provides a potential tool for estimating long-term MACE risk in patients with high-risk ACS and NVAF after PCI. In our analysis, the model demonstrated improved predictive performance compared with the GRACE score. However, its clinical utility for guiding management decisions requires further validation in prospective investigation.