Towards Precision Oncology: A Predictive and Causal Lens

Abstract

Precision oncology promises personalized care for each patient based on a holistic view of their data. However, several methodological and translational advances are required for successful implementation of this vision in the clinic. These include building temporal models to predict a patient’s survival outcomes in response to therapy, validating these methods with experimental data from Randomized Controlled Trials (RCTs), quantifying the uncertainty in the predictions, and finally, exploring how these elements can be woven together into a clinical decision support tool. In this thesis, I explore each of these aspects in turn: i) first, I build different models of clinical time-series data, with a focus on prediction of survival outcomes and forecasting of core biomarkers, ii) next, I design methods to give additional “context” for these models, including uncertainty quantification of causal estimates and validation of these estimates using RCT data, and iii) finally, I study how these elements affect treatment decision-making via a controlled user study of a decision support tool prototype.