Contextual Motifs: Increasing the Utility of Motifs using Contextual Data
Ian Fox (CSE, University of Michigan);Lynn Ang (Department of Internal Medicine, Division of Metabolism, Endocrinology and Diabetes, University of Michigan);Mamta Jaiswal (Department of Internal Medicine, Division of Metabolism, Endocrinology and Diabetes, University of Michigan);Rodica Pop-Busui (Department of Internal Medicine, Division of Metabolism, Endocrinology and Diabetes, University of Michigan);Jenna Wiens (CSE, University of Michigan)
Abstract
Motifs are a powerful tool for analyzing long physiological signals. Standard motif methods, however, ignore important contextual information (e.g., what the patient was doing at the time the data were collected). We hypothesize that these additional contextual data could increase the utility of motifs. Thus, we propose an extension to motifs, \textit{contextual motifs}, that incorporates context. We present methods to discover contextual motifs, both with observed and inferred contextual information. Oftentimes, we may not observe context, or collecting context may simply be too burdensome. In this setting, we present methods to jointly infer motifs and context. Through experiments on simulated data, we illustrate the potential discriminative power of contextual motifs across a range of settings, improving discriminative performance, measured using AUROC, by up to 11 percentage points over a contextless baseline. We further validate the proposed approach on a dataset of continuous glucose monitor data collected from type 1 diabetics. Applied to the task of predicting hypo- and hyper-glycemic events, use of contextual motifs led to a 7.2 percentage point improvement in AUROC compared with a state-of-the-art baseline.
