Towards an Adaptive Algorithm for Online Substance Use Episode Detection

• Activity Number: 356 - Statistical Learning: Methods and Applications
• Type: Contributed
• Date/Time: Wednesday, August 5, 2020 : 10:00 AM to 2:00 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Joshua Rumbut* and Hua Fang
• Companies: University of Massachusetts Dartmouth, University of Massachusetts Medical School and University of Massachusetts Dartmouth, University of Massachusetts Medical School
• Keywords: online learning; anomaly detection; data streams; wearable biosensors; substance use disorder; machine learning

Abstract

With the growth of wearable biosensors as medical devices and in consumer products such as smart watches, data streams are being generated that can enable continuous health monitoring and new research into the individualized circumstances of health events such as relapse during treatment of Substance Use Disorder (SUD). Our recently proposed intelligent substance episode detection method, based on denoising and inference of episode type based on reconstruction error, is an online anomaly detection algorithm capable of identifying episodes of varying length. Automated parameter selection is achieved based on the distribution of reconstruction error on data without anomalies. Episode length and type are determined by a sliding window algorithm which looks for changes in the data stream consistent with previously observed usage episodes. We use simulated data streams based on those collected from clinical trials in order to characterize the performance of our method. We present preliminary results showing the relationship between the threshold for abnormal sensor readings and the ability to detect episodes in real and simulated data streams in the SUD context.