Detecting Major Depressive Disorder Presence Using Passively-Collected Wearable Movement Data in a Nationally-Representative Sample

ImportanceMajor Depressive Disorder (MDD), is a widespread and often debilitating mental health condition characterized by significant heterogeneity in symptom profiles, which poses a challenge for early detection. Changes in sleep and movement patterns associated with MDD carry valuable diagnostic information. This study explored the utility of wrist-worn actigraphy data in combination with machine learning and deep learning techniques to detect MDD.ObjectiveThis study aimed to evaluate the efficacy of domain-knowledge driven, traditional machine learning and domain-knowledge agnostic deep learning techniques in using wrist-worn actigraphy data to evaluate the performance of these two commonly used approaches in detecting MDD, and to identify actigraphy-derived biomarkers indicative of individuals with MDD.DesignThis research leveraged minute-level, wrist-worn actigraphy data obtained over a consecutive week from a cross-sectional sample from two subsequent waves (2011-2012 & 2013-2014) of the National Health and Nutrition Examination Survey (NHANES).SettingNHANES is a nationally-representative population-based study that assesses the health and nutritional status of individuals in the United States.ParticipantsParticipants were individuals with a completed Patient Health Questionnaire-9 (PHQ-9) and one-week of actigraphy data available.ExposureThe primary exposure was the binary presence of MDD, defined by a PHQ-9 total score ≥ 10.Main Outcome(s) and Measure(s)The main outcome was movement intensity, as measured by a summative metric derived from a wrist-worn Actigraph GT3X+ over one week.ResultsThe analysis incorporated data from 8,378 participants (MDD Presence = 766 participants, 51.20% Female, AgeMean = 47.59 years. We employed two machine learning modeling strategies: (1) a traditional machine learning (ML) approach with theory-driven feature derivation (AUCTest = 0.61), and (2) a deep learning Convolutional Neural Network (CNN) approach, coupled with gramian angular field transformation (AUCTest = 0.68). Model introspection revealed movement-related features to be the most influential in the traditional ML approach and nighttime movement to be the most influential in the CNN model’s predictions of MDD presence.ConclusionsThis study highlights the potential of using passively-collected, actigraphy data for understanding and detecting MDD. It is the first known attempt to compare theory-driven machine learning and deep learning methodologies using minute-level wrist-worn actigraphy data from a large, nationally-representative U.S. sample.