INTERPRETABLE MACHINE LEARNING UNCOVERS NOVEL PREDICTORS OF TRANSCATHETER AORTIC VALVE REPLACEMENT FUTILITY AND MID-TERM OUTCOMES

Abstract

Valvular heart disease, particularly aortic stenosis, presents significant challenges in cardiovascular care, as traditional diagnostic and prognostic methods often struggle to provide timely and accurate risk stratification. Conventional interventional risk prediction tools, which rely on a limited set of clinical features, fail to capture the patient heterogeneity that defines disease progression, leading to suboptimal clinical outcomes following intervention. This thesis demonstrates the value of machine learning as a new paradigm to address these limitations. Our core study applied a Random Forest model to a cohort of 213 aortic stenosis patients to predict 6-month transcatheter aortic valve replacement (TAVR) risk and futility, validated using repetitive 5-fold cross validation technique. Our findings show that this non-linear model significantly outperformed traditional risk scores. The findings highlight the limitations inherent in the linear modeling assumptions of conventional tools. Furthermore, by applying the SHAP interpretability method, this work uncovers novel serum predictors. These new biomarkers and their thresholds offer valuable insights into underlying pathological progressions related to patient frailty, inflammation, renal dysfunction and vascular health. In summary, this work highlights the potential of machine learning to personalize TAVR patient care. By providing an improved predictive model and revealing novel predictors linked to patient frailty, comorbidity, and vascular health, our work offers a new lens for patient assessment. This may advise clinicians to make more targeted decisions and implement personalized interventional care, ultimately reducing futility and improving patient outcomes.