Prognostic modeling from medical data holds the promise of informing personalized care and improving clinical decision-making. This thesis explores two applications of survival analysis and deep learning to estimate long-term patient risk and treatment benefit in high-impact cardiopulmonary settings. In the first study, we develop a deep multimodal time-to-event prediction framework that estimates patient-specific mortality risk using chest radiographs and demographic features. Unlike traditional binary classifiers, our approach, leveraging models such as Cox proportional hazards and deep survival machines, accounts for right-censoring and allows for risk estimation at arbitrary time horizons, offering greater flexibility and clinical utility. In the second study, we apply individualized treatment effect estimation to determine which patients with stable ischemic heart disease are most likely to benefit from coronary artery bypass grafting (CABG). Using a recently proposed machine learning algorithm, Cox Mixtures with Heterogeneous Effects (CMHE), we stratify patients based on their predicted survival gain from CABG versus optimal medical therapy alone and validate these predictions on an external surgical cohort. Together, these works demonstrate the potential of survival-based machine learning models to enhance personalized risk prediction and treatment optimization in real-world clinical scenarios.

Prof. Artur Dubrawski (advisor)

Prof. George H. Chen

Angela Chen

Meeting ID: 913 6827 8661
Passcode: 597766