Machine Learning to Predict Long-Term Cardiac-Relative Prognosis in Patients With Extra-Cardiac Vascular Disease

Abstract

Aim: Patients with ischemic stroke (IS), transient ischemic attack (TIA), and/or peripheral artery disease (PAD) represent a population with an increased risk of coronary artery disease. Prognostic risk assessment to identify those with the highest risk that may benefit from more intensified treatment remains challenging. To explore the feasibility and capability of machine learning (ML) to predict long-term adverse cardiac-related prognosis in patients with IS, TIA, and/or PAD. Methods: We analyzed 636 consecutive patients with a history of IS, TIA, and/or PAD. All patients underwent a coronary CT angiography (CCTA) scan. Thirty-five clinical data and 34 CCTA metrics underwent automated feature selection for ML model boosting. The clinical outcome included all-cause mortality (ACM) and major adverse cardiac events (MACE) (ACM, unstable angina requiring hospitalization, non-fatal myocardial infarction (MI), and revascularization 90 days after the index CCTA). Results: During the follow-up of 3.9 ± 1.6 years, 21 patients had unstable angina requiring hospitalization, eight had a MI, 23 had revascularization and 13 deaths. ML demonstrated a significant higher area-under-curve compared with the modified Duke index (MDI), segment stenosis score (SSS), segment involvement score (SIS), and Framingham risk score (FRS) for the prediction of ACM (ML:0.92 vs. MDI:0.66, SSS:0.68, SIS:0.67, FRS:0.51, all P < 0.001) and MACE (ML:0.84 vs. MDI:0.82, SSS:0.76, SIS:0.73, FRS:0.53, all P < 0.05). Conclusion: Among the patients with IS, TIA, and/or PAD, ML demonstrated a better capability of predicting ACM and MCAE than clinical scores and CCTA metrics.