Using Machine Learning and Sensing to Contextualize Hyperactivity Measurement on Wearable Devices

PROJECT SUMMARY/ABSTRACT ADHD is the most common early childhood diagnosis, affecting around 9% of US children. LemurDx is a software system that uses state-of-the-art sensor technology to measure hyperactivity. LemurDx passively collects data from smartwatch sensors which are transmitted to a HIPAA-compliant server. Machine learning (ML) algorithms are used to assess context (a combination of activity and environment) using multi-level classification to first identify the wearer's class of activity, then contextualize the level of motion, and evaluate activity level based on that context. Context (e.g., classroom versus playground) is critical to differentiate children with clinical vs normal levels of hyperactivity. LemurDx is superior to research-focused tools such as traditional actigraphy in a number of ways. Beyond the improved precision afforded by modern smartwatch sensors and ML algorithms, we are designing an integrated system to address clinical trial needs. LemurDx has the potential to improve the process of medication titration and outcome assessment in ADHD clinical trials by providing investigators with rapid feedback and an objective outcome measure. In this project, we will test the utility of LemurDx to measure changes in hyperactivity in children on days when they are on versus off medication in standardized and real- life settings. This pilot data will allow us to test the utility of LemurDx as an objective measure of hyperactivity to measure response to medication and help investigators assess medication efficacy in clinical trials. This SBIR (R44) project will accomplish two aims over two phases. Aim 1 is to refine the LemurDx product to enhance ease of use among clinical researchers, including the automated collection, cleaning, and feature extraction of wearable sensor data, the ML processing pipeline, and an updated smartwatch app. Aim 2 is to test LemurDx in 100 children (ages 6-11 years) with ADHD-Hyperactive Type or Combined Type who are prescribed stimulant medication on days when they are on versus off medication, both in real-world and standardized lab settings. We will refine the ML algorithms to detect the presence and severity of hyperactivity. We will iteratively test and refine LemurDx’s ML algorithms to improve its generalizability and to achieve high sensitivity and specificity in estimating hyperactivity severity. We will use these insights to design a scalable clinical delivery system. LemurDx innovations include a readily available and reliable measurement technology combined with accurate ML algorithms to objectively measure hyperactivity and its context in a system that is intentionally designed to meet the needs of clinical trial investigators. We will demonstrate the commercialization potential of LemurDx by validating the feasibility of using LemurDx to measure medication response in children with ADHD-Hyperactive Type or Combined Type and the acceptability of the system to ADHD clinical investigators.