Methods for Improving Inference in Clinical Outcomes

Abstract

Advances in technology have allowed for the collection of diverse data types along with evolution in computer algorithms. This dissertation focuses on the development and application of novel methodologies to model and improve inference on clinical outcomes. First, a new prognostic approach of modeling time-to-event data using Bayesian Networks (BNs) is developed and illustrated using publicly available cancer data. This approach allows for flexible modeling of different structural relationships that might exist between variables at different periods, hence, improving our understanding of critical prognostic factors that can inform patient care and development of targeted interventions. As a prognostic model, BNs demonstrated better or comparable performance as compared to other equivalent models for bladder and lung cancer data. In this dissertation, we also reviewed application of predictive modeling algorithms in randomized clinical trials (RCTs). RCTs are costly and time-consuming. Predictive modeling has the potential to mitigate challenges associated with clinical trial failures and facilitate efficient clinical trial conduct in areas such as patient recruitment, trial optimization, and safety & efficacy evaluations. Finally, we present a new approach for estimating causal treatment effect in RCTs that are prone to post-randomization intercurrent events (ICE). Examples of ICE include treatment switch, treatment discontinuation, or adverse events. Here, we adopt the principal stratification framework where we first predict the latent strata membership using baseline covariates and then estimated causal treatment effects using appropriate stratum having a homogeneous group of subjects. Using simulations, our approach demonstrated a better performance in estimating treatment effects as compared to the standard intent-to-treat (ITT) strategy.