Machine Learning Techniques for Constructing Tailoring Variables for Sequential Decisionmaking

• Activity Number: 155
• Type: Topic Contributed
• Date/Time: Monday, July 30, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Daniel Almirall*+
• Companies: University of Michigan
• Keywords: adaptive intervention ; generalized boosted models ; unsupervised learning ; supervised learning

Abstract

Chronic conditions often require that clinicians make sequential, individualized treatment decisions. A dynamic treatment regime (DTR) is a sequence of individually-tailored decision rules that use demographic information, severity and illness history, co-morbidities, and response to prior treatment as inputs and then recommend the type, modality, intensity or delivery of subsequent treatment(s) as outputs. A key step in developing effective DTRs is identifying tailoring variables to use as inputs in the decision rules. Tailoring variables are special types of pre-treatment moderators that pinpoint which treatment is best for whom. Existing methods do not explicitly (a) identify which measures (variable selection), or (b) combine them (feature construction) for use as tailoring variables. In this talk, we propose a framework for adapting and extending machine learning techniques (originally-developed for prediction) for use in tailoring variable selection and feature construction. Using an observational, longitudinal data set, we illustrate the approach by developing a DTR for determining the optimal duration/discontinuation of adolescent substance use treatment.