Conference Program

A Case Study with RCT of Varenicline Using Two Machine Learning Approaches (310244)

Wave-Ananda Baskerville, University of California, Los Angeles 
*Alondra Cruz, University of California, Los Angeles 
Suzanna Donato, University of California, Los Angeles 
Erica N Grodin, University of California, Los Angeles 
Amanda Kay Montoya, University of California, Los Angeles 
Lara A Ray, University of California, Los Angeles 

Keywords: Machine Learning, Treatment Responders, Qualitative Interactions, Lasso

Clinical researchers conducting randomized clinical trials (RCTs) are often interested in identifying who responds to treatments. Recently, machine learning (ML) has become a popular method for identifying treatment responders. However, the application of ML methods has been limited within alcohol use disorders (AUD) treatment. QUalitative INteraction Trees (QUINT; Dusseldorp & Van Mechelen, 2014) and group-lasso interaction net (glinternet; Lim & Hastie, 2015) seem especially promising in identifying treatment responders. QUINT is a tree-based method that is designed to create subgroups that are involved in qualitative interactions by maximizing prediction accuracy of the treatment effect within a group instead of the outcome. Glinternet is a lasso-based regularization method that is designed to capture interactions that have a strong hierarchy by placing constraints on the main effects and interactions. The two methods were compared using data from the NCIG RCT of Varenicline to treat AUD. We examined the primary drinking outcome percent heavy drinking days (PHDD). The QUINT model resulted in a pruned tree with 10 leaves. Each leaf represented a subgroup of patients and results indicated that in 2 of the leaves, Varenicline was more effective at reducing PHDD compared to placebo. Some of the variables in the model included participant age and sitting heart rate at week 0. The glinternet model selected features that interacted to represent treatment responders, like participant age and drinking goal after treatment. Participant age was a common variable selected in both models. Both methods appear to be promising at identifying treatment responders in AUD treatment. Researchers interested in using these methods to identify treatment responders should consider the strengths and limitations of each method. When prioritizing qualitative interactions use QUINT and when prioritizing strong prediction accuracy use glinternet.