M-Health and Autism: Recognizing Stress and Anxiety with Machine Learning and Wearables Data

Abstract

Consumer-grade wearables provide physiological measurements which may inform m-health applications that predict adverse outcomes. Autism Spectrum Disorder (ASD) represents a compelling example. Many individuals with ASD present with challenging behaviors that are preceded by physiological changes. Physiological measures could, therefore, support real-time interventions to avert challenging behaviors in various social settings. However, no prior research has demonstrated a methodological approach to detect these changes using wearable device data. We sought to demonstrate a machine learning approach that uses wearables data to differentiate physiological states associated with stressful and non-stressful scenarios in children with ASD. In a controlled laboratory setting, we collected heart rate and RR interval measurements during rest and during activities designed to mimic stress using a consumer-grade wearable device. Our analysis included 38 participants (22 ASD, 16 non-ASD). Following outlier removal, we extracted 20 statistical features from data collected during each patient's rest and stressful periods. Using nested leave-one-out cross-validation over 76 sample periods (38 rest / 38 stress), we trained and evaluated logistic regression (LR) and support vector machine (SVM) classifiers to label each validation sample as a rest or stressful period. The SVM and LR models achieved 93% and 87% accuracy, respectively. These results suggest that machine learning models combined with wearables data may support real-time m-health intervention applications.