Measuring Individual Treatment Benefits Using Longitudinal Outcomes from Clinical Trials or Hospital Data

Abstract

It is increasingly recognized that a patient’s response to a medical treatment is a statistically heterogeneous phenomenon. The average treatment effects may not represent a heterogeneous population of patients. The benefits each patient receive from the treatment could differ, requiring measurement of treatment benefits at the patient level. Despite of the development of methods in this field, new methods are needed for predicting individual treatment benefits using longitudinal binary outcomes or hospital data with nonignorable missingness. This dissertation has three main chapters. Chapter 1 introduces a method for predicting individual treatment benefits based on a personalized medicine model that implements random effects logistic regression of binary outcomes that may change over time. The method uses empirical Bayes (EB) estimators based on patients’ characteristics and responses to treatment. The prediction performance is evaluated in simulated new patients using correlations between the predicted and the true benefits as well as relative biases of the predicted benefits versus the true benefits. As an application, the method is used to examine changes in the disorganized dimension of antipsychotic-naïve patients from an antipsychotic randomized clinical trial. Chapter two of the dissertation presents a method for predicting individual treatment benefits with a novel 2-dimensional personalized medicine model that handles non-ignorable missingness due to hospital discharge and evaluate its reliability and accuracy by simulations. The longitudinal outcome of interest is modeled simultaneously with the hospital length of stay through a joint mixed model. The method is illustrated with an application assessing individual pain management benefits post spine fusion surgery. EB-Predicted individual benefits are compared with Monte-Carlo computed benefits. Pearson’s correlations and relative biases are used to assess the prediction accuracy. Finally, Chapter three of the dissertation applies the methodology developed in Chapter two to analyze with more clinical detail the impact of depression and age on individual benefits of postoperative pain management in lumbar spinal fusion patients using Cerner HealthFacts® electronic health records. The developed joint multivariate mixed model of pain scores and length of hospital stay is used to analyze individual benefits. The effects of depression and age on the amount and rate of change of the pain management benefits are evaluated, as well as the association between individual benefits and post-surgical hospital length of stay. We conclude that the utilization of the EB prediction of individual treatment benefits is useful in the analyses of treatment effects using not only clinical trial data but also electronic health records. Predicted individual treatment benefits are accurate when model parameters are reliably estimated.