Machine Learning-Based Prediction of Long-Term Outcomes in Patients With Chronic Total Occlusion of the Coronary Artery.

Abstract

Background and objectives: Precise prediction of long-term outcomes in patients with chronic total occlusion (CTO) of the coronary artery is crucial for cardiovascular care. The recent development of advanced machine learning (ML) models has opened up new possibilities in medical prognostics. This study aimed to develop ML models and validate their performance in predicting long-term clinical outcomes in patients with CTO.

Methods: This study retrospectively analyzed 3,248 patients listed in the Asan Medical Center CTO Registry (2003-2018). Patients underwent coronary artery bypass grafting, percutaneous coronary intervention, or optimal medical therapy and were followed up for a median period of 5.3 years. The study population was randomly split into training (n=2,598) and test (n=650) sets. Three ML algorithms-namely, L2-penalized logistic regression, artificial neural networks, and CatBoost-were employed to develop a prognostic model for 5-year cardiac death (primary endpoint) as well as 5-year all-cause mortality and target vessel revascularization (TVR) (secondary endpoints). Model performance was assessed using area under the receiver operating characteristic curves (AUCs), and feature importance was evaluated using SHapley Additive exPlanations values.

Results: The three ML algorithms exhibited comparable performance in predicting 5-year cardiac death (AUC: 0.80). Additionally, these three ML algorithms successfully predicted 5-year all-cause mortality (AUC: 0.83-0.84) and TVR (AUC: 0.65-0.74), showing good predictive performance. Patient demographics and comorbidities, rather than treatment modality, were the leading predictors of outcomes.

Conclusions: The ML models are reliable in predicting 5-year clinical outcomes in patients with CTO, demonstrating their potential for clinical application.