Development of a long-term survival prediction model for patients undergoing invasive coronary angiography using ensemble-based machine learning in time-to-event analysis.

BACKGROUND: Prognosis prediction for high-risk patients undergoing invasive coronary angiography (ICA) is crucial for clinical decision-making. Despite machine learning (ML) advancements, time-to-event survival prediction remains limited. OBJECTIVES: This study developed an ensemble ML model based on survival analysis to predict long-term outcomes in ICA patients. METHODS: A total of 9517 ICA patients (2008-2020) were retrospectively analyzed. The primary outcome was all-cause mortality, with follow-up until December 31, 2021. Using 8 ML algorithms, we developed a model comprising 80 variables. Model performance was assessed using time-dependent C-index and Brier score, with variable importance analyzed using permutation-based and partial dependent plots. RESULTS: Survival Quilts model achieved the highest time-dependent C-index (0.920 at 30 days, 0.897 at 365 days), outperforming other ML algorithms. Time-dependent Brier scores generally increased, which remained stable. ICA-related characteristics had the greatest impact on mortality, while laboratory results, comorbidities, and patient characteristics gained influence over time. By day 365, patient characteristics and laboratory results became more prominent predictors. Among the domains, key variables included catheterization status, C-reactive protein, smoking, and chronic kidney disease. CONCLUSION: Survival analysis-based ensemble ML models, such as Survival Quilts, improve survival prediction by capturing time-varying influences of key predictors, offering a foundation for more precise cardiovascular care.