Using Statistical Learning to Develop a More Sensitive Outcome for Progressive Multiple Sclerosis

• Activity Number: 175 - Contributed Poster Presentations: Section on Statistical Learning and Data Science
• Type: Contributed
• Date/Time: Monday, July 31, 2017 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Christopher Barbour* and Mark Greenwood and Peter Kosa and Danish Ghazali and Makoto Tanigawa and Blake Snyder and Bibiana Bielekova
• Companies: Montana State University and Montana State University and National Institute of Neurological Disorders and Stroke, National Institutes of Health and National Institute of Neurological Disorders and Stroke, National Institutes of Health and National Institute of Neurological Disorders and Stroke, National Institutes of Health and National Institute of Neurological Disorders and Stroke, National Institutes of Health and National Institute of Neurological Disorders and Stroke, National Institutes of Health
• Keywords: multiple sclerosis ; scale construction ; evolutionary algorithm ; statistical learning ; clinical endpoints ; longitudinal scale development

Abstract

Scales are often constructed from multiple outcome measures to create a combined metric that is a better measure of the true trait of interest than the any of the original components. These methods typically focus on explaining cross-sectional variation in the responses to define the combinations of variables. When the interest is in creating a scale that is sensitive to changes over time, developing it using cross-sectional data may not tune the projection to detect changes over time optimally. This research focuses on the creation of a new scale that, instead of explaining cross-sectional variation, is optimized to detect variation over time in longitudinal data. This method is performed on a motivating dataset of multiple sclerosis (MS) patients, where interest lies in constructing a better clinical endpoint to detect small changes in disease progression. An evolutionary algorithm, as implemented in the R package GA, is used to generate the optimal coefficient weights. The assumptions, computational challenges in developing this scale, preliminary results from a simulation study, and other potential applications are also discussed.