Abstract
Elastic net regularization is a popular statistical tool for variable selection that combines lasso and ridge regression penalties. When used in combination with ensemble methods, it improves stability of the estimates and increases confidence in the results. We proposed and tested a version of this method that considers a measure of models' goodness of fit and gives estimates of importance for each feature weighted on this measure. The method was applied to an autism spectrum disorder (ASD) study to select a subset of biosensor-based features that can be used to predict clinical scores of study participants. In this study, the participants' responses to visual and audio stimuli were captured by the Janssen Autism Knowledge Engine (JAKE®) biosensors and used to construct approximately 50,000 features. We examined how well changes in these features mirrored changes in the Social Responsiveness Scale (SRS), a quantitative assessment of ASD individuals by clinicians. As a result, we isolated the top features changes which are most associated with changes in SRS, and built predictive models using these features.