Precision Medicine aims to deliver patient-centered care beyond one-size-fits-all treatment rules, by accounting for heterogeneity in treatment effect to determine patient subgroups. In- dividualization promises two benefits: (i) to extend the population mean survival (ii) to spare patients from aggressive treatment who are unlikely to benefit. The "Direct Value Search" approach frames the estimation of Individualized Treatment Rules as an Empirical Risk Mini- mization problem, allowing for the use of classification techniques which avoid strict assump- tions of parametric regression models, and therefore minimize the possibility of misspecifica- tion.
This thesis reviews a series of developments in Outcome Weighted Learning (OWL) meth- ods that handle right-censored survival outcomes. We evaluate the proposed efficiency-gain and distributional robustness of the Multistate Outcome Weighted Learning (MSOWL) method. MSOWL integrates Inverse-Probability-Censoring-Weighted individual stochastic benefit pro- cesses, including right-censored cases, at the cost of cubic computational complexity, which can be ameliorated via divide and conquer. We analyze a novel data set from an oncologi- cal clinical trial, which assesses the inclusion of immunotherapy to standard chemotherapy for Pancreatic Cancer.
Precision Medicine, Survival Analysis, Causal Inference, Clinical Trials, Missing Data, Statistical Learning, Support Vector Machine
The clinical trial data are available via projectdatasphere.org.
This thesis investigates the method of Bakoyannis Multistate Outcome Weighted Learning, Biostatistics 2023, code available on Github