Prognostic Value of Machine Learning-based Time-to-Event Analysis Using Coronary CT Angiography in Patients with Suspected Coronary Artery Disease

Purpose: To assess the long-term prognostic value of a machine learning (ML) approach in time-to-event analyses incorporating coronary CT angiography (CCTA)-derived and clinical parameters in patients with suspected coronary artery disease. Materials and Methods: The retrospective analysis included patients with suspected coronary artery disease who underwent CCTA between October 2004 and December 2017. Major adverse cardiovascular events were defined as the composite of all-cause death, myocardial infarction, unstable angina, or late revascularization (>90 days after index scan). Clinical and CCTA-derived parameters were assessed as predictors of major adverse cardiovascular events and incorporated into two models: a Cox proportional hazards model with recursive feature elimination and an ML model based on random survival forests. Both models were trained and validated by employing repeated nested cross-validation. Harrell concordance index (C-index) was used to assess the predictive power. Results: A total of 5457 patients (mean age, 61 years 11 [SD]; 3648 male patients) were evaluated. The predictive power of the ML model (C-index, 0.74; 95% CI: 0.71, 0.76) was significantly higher than the Cox model (C-index, 0.71; 95% CI: 0.68, 0.74; P = .02). The ML model also outperformed the segment stenosis score (C-index, 0.69; 95% CI: 0.66, 0.72; P < .001), which was the best performing CCTA-derived parameter, and patient age (C-index, 0.66; 95% CI: 0.63, 0.69; P < .001), the best performing clinical parameter. Conclusion: An ML model for time-to-event analysis based on random survival forests had higher performance in predicting major adverse cardiovascular events compared with established clinical or CCTA-derived metrics and a conventional Cox model.