Quantitative methods for stratified medicine

Stratified medicine has tremendous potential to deliver more effective therapeutic intervention to improve public health. For practical implementation, reliable prediction models and clinically meaningful categorization of some comprehensible summary measures of individual treatment effect are vital elements to aid the decision-making process and bring stratified medicine to fruitful realization. We tackle the quantitative issues involved from three fronts : 1) prediction model building and selection; 2) reproducibility assessment; and 3) stratification. First, we propose a systematic model development strategy that integrates cross-validation and predictive accuracy measures in the prediction model building and selection process. Valid inference is made possible via internal holdout sample or external data evaluation to enhance generalizability of the selected prediction model. Second, we employ parametric or semi-parametric modeling to derive individual treatment effect scoring systems. We introduce a stratification algorithm with constrained optimization by utilizing dynamic programming and supervised-learning techniques to group patients into different actionable categories. We integrate the stratification and newly proposed prediction performance metric into the model development process. The methodologies are first presented in single treatment case, and then extended to two treatment cases. Finally, adapting the concept of uplift modeling, we provide a framework to identify the subgroup(s) with the most beneficial prospect; wasteful, harmful, and futile subgroups to save resources and reduce unnecessary exposure to treatment adverse effects. The proposals are illustrated by AIDS clinical study data and cardiology studies for non-censored and censored outcomes. The contribution of this dissertation is to provide an operational framework to bridge predictive modeling and decision making for more practical applications in stratified medicine.