UCLA

This dissertation evaluates when and how machine learning (ML) adds value beyond traditional statistical methods for cardiovascular risk prediction across three electronic health record (EHR)-derived cohorts. It consists of three projects. Project 1 examined major adverse cardiovascular events (MACE) in 5,766 patients treated with immune checkpoint inhibitors (ICIs). MACE occurred in 19.6% of patients and was associated with traditional risk factors, cancer exposures, and treatment intensity. ML (XGBoost, AUROC 0.71) identified non-cardiovascular immune-related adverse events requiring prednisone as a novel, high-impact predictor. Project 2 investigated new-onset left ventricular systolic dysfunction (LVSD) after orthotopic liver transplantation in 956 patients. LVSD developed in approximately 7% of patients, typically early in the post-transplant course, and was often reversible. Risk markers included mineralocorticoid receptor antagonist use, lower pre-operative hemoglobin, intraoperative dialysis, and diastolic/pulmonary hemodynamic features; ML (XGBoost, AUROC 0.73) revealed important non-linear interactions among echocardiographic variables. Project 3 analyzed predictors of successful decannulation in veno-arterial extracorporeal membrane oxygenation (VA-ECMO). Among 199 patients, successful decannulation occurred in 48.5%. ML (random forest, AUROC 0.94) and regression converged on higher mixed venous oxygen saturation as the dominant predictor. Across these three projects, ML complemented logistic regression by uncovering novel predictors in Project 1, capturing complex interactions in Project 2, and isolating a single dominant predictor in Project 3. Together, the combined use of traditional statistical methods and ML enhanced risk stratification across diverse cohorts and provides a framework for future applications in cardiovascular prediction.