Content area
Abstract
Checkpoint immunotherapy drugs, either individually or in combination with other drugs, have become standard of care for many cancers. The long-term survival impacts of these drugs on a subset of treated population and their unique survival dynamics have challenged traditional statistical methods to model long-term survival impacts and estimate individualized treatment rules. In this dissertation, I proposed novel machine learning techniques that can be used to tackle some of these challenges.
This research addresses the following three aims using patient-level data from a checkpoint immunotherapy clinical trial for advanced melanoma: (1) Develop a novel individual-level survival extrapolation method for right-censored observations, and compare the predictive accuracy of the proposed method with population-level standard parametric models. (2) Compare the accuracy of survival extrapolation models that directly model heterogeneity of treatment response to the accuracy of the proposed survival extrapolation model from Aim 1 that incorporates cure fraction models at the individual level. (3) Estimate individualized treatment rules (ITRs) and calculate survival and cost impacts associated with implementing them in the trial cohort compared to the survival and cost impacts associated with universal use of the trial-recommended treatment.
The Aim1 paper provides a tutorial that introduces the kernel-weighted survival forest (KWSF) model, a novel survival extrapolation method that uses patient-level characteristics to estimate individualized survival function. The findings showed that compared to standard parametric models, KWSF more accurately predicted survival beyond the available trial follow up. The results of the Aim2 paper showed that compared to models that use standard parametric extrapolation, cure fraction models and KWSF with cure fraction extrapolation function were more accurate in predicting survival in the immunotherapy arm of the trial. The KWSF model with a cure fraction survival extrapolation function demonstrated comparable accuracy with cure fraction models, while uniquely allowing for estimating individual-level survival functions. The findings of Aim3 paper showed that compared to allocating treatment based on the average treatment effect from a clinical trial, treatment allocation based on the estimated ITRs resulted in higher survival gains and lower direct treatment costs, which is likely to persist even when considering the cost of implementing individualized treatment.