ADHD identification from real-life recorded video

Attention Deficit Hyperactivity Disorder (ADHD) is a neurodevelopmental disorder that affects more than 10% of the population. Accurate and early diagnosis of ADHD is crucial for effective treatment and management of the disorder. Today, ADHD identification is mostly performed at clinical settings and is based on diagnostic interviews, and self-report questionnaires. This evaluation suffers from rater-bias, requires visiting the medical facility, is dynamic and depends on environmental stimulations, which cuases the ADHD identification process to fail to capture the sub-threshold clinical nuances of the ADHD spectrum. To monitor ADHD level continuously at home settings, we utilize real-life video clip recording dataset consisting of 78 video clips (40 females, 38 males). A pre-trained deep learning framework was used to extract upper body skeleton joint coordinates over time, from which dynamic spatial- temporal features were computed. For ADHD recognition and scoring evaluation we used deep learning architecture applied directly on the skeletal joints (TabNet, TCC, and LSTM), and general-purpose classifiers (XGBoost, Naive Bayes, and SVM) with recognition accuracy between 0.66 and 0.88, and 0.65-0.82 for female, and male, respectively. We derived real-life ADHD and biomarkers based on contextual clustering method for each gender. The suggested technology has a potential to be used for continuous assessment of ADHD symptoms in real-life home settings, can assist as an additional evaluation tool for clinicians in addition to the commonly used questionnaires, can contribute new real-life ADHD biomarkers that may assist in understanding the course of the disorder, and assist the clinician in treatment planning and in optimizing medication usage in ADHD and other neurological diseases.