Deep learning diagnosis plus kinematic severity assessments of neurodivergent disorders

Abstract

Early diagnostic assessments of neurodivergent disorders (NDD), remains a major clinical challenge. We address this problem by pursuing the hypothesis that there is important cognitive information about NDD conditions contained in the way individuals move, when viewed at millisecond time scales. We approach the NDD assessment problem in two complementary ways. First, we applied supervised deep learning (DL) techniques to identify participants with autism spectrum disorder (ASD), attention-deficit/hyperactivity disorder (ADHD), comorbid ASD + ADHD, and neurotypical (NT) development. We measured linear and angular kinematic variables, using high-definition kinematic Bluetooth sensors, while participants performed the reaching protocol to targets appearing on a touch screen monitor. The DL technique was carried out only on the raw kinematic data. The area under the receiver operator characteristics curve suggests that we can predict, with high accuracy, NDD participant's conditions. Second, we filtered the high frequency electronic sensor noise in the recorded kinematic data leaving the participants' physiological characteristic random fluctuations. We quantified these fluctuations by their biometric Fano Factor and Shannon Entropy from a histogram distribution built from the magnitude difference between consecutive extrema unique to each participant, suggesting a relationship to the severity of their condition. The DL may be used as complementary tools for early evaluation of new participants by providers and the new biometrics allow for quantitative subtyping of NDDs according to severity.