Predicting differential treatment outcomes in randomized clinical trials: A comparison of model-based and machine learning approaches

Abstract

Precision medicine refers to the practice of improving health outcomes by subgrouping patients and tailoring care to the individual rather than treating patients as a homogenous group (National Research Council, 2011). This initiative is particularly relevant to clinical psychology and psychiatry, where theories of psychopathology commonly posit complex relations among biological, psychological, and social variables, and effects of evidence-based interventions are marked with significant heterogeneity in individual treatment response. To identify patient characteristics and/or patient subgroups predicting differential treatment outcomes, researchers have traditionally examined potential moderators of treatment outcome in model-based regression frameworks. Model-based approaches require a priori specification of hypothesized predictors and moderators, often leading to overly simplistic regression models that omit relevant predictors and insufficiently account for the underlying complexity of psychopathology. Thus, flexible, data-driven machine learning approaches may greatly enhance progress in precision psychiatry initiatives. The current study compared model-based and machine learning approaches for predicting differential treatment outcomes in a randomized clinical trial (RCT) assessing the effectiveness of cognitive behavior therapy (CBT) for youth with chronic abdominal pain. Modeling approaches were selected to handle two analytic tasks relevant to precision medicine: 1) Identifying patient predictors relevant to treatment outcome, and 2) Modeling complexity and underlying interaction effects among relevant variables. Results indicated that the machine learning algorithm (i.e., conditional random forest) significantly outperformed the model-based approaches, due in part to the presence of underlying interaction effects that were unaccounted for by model-based approaches. Higher-order interactions were subsequently probed by fitting individual conditional inference trees, further characterizing patient subgroups associated with differential treatment outcomes. Findings support the incremental utility of machine learning algorithms for improving prediction of treatment outcomes and tailoring interventions to the individual.