ADHD Behavior Monitoring Using Deep-Learning Models

Abstract

Attention-Deficit-Hyperactivity Disorder (ADHD) poses significant challenges in monitoring and understanding post-treatment behaviors in affected children. This paper compares post-treatment behaviors in children with Attention-Deficit/Hyperactivity Disorder (ADHD) using video-based monitoring and deep learning approaches. The study uses both skeleton-based and image-based categorization methods to identify actions including walking, sitting, and fidgeting. Several neural network designs, including CNN-LSTM, LSTM, GRU, VGG16, and ResNet50, are tested for their usefulness in behavior classification. The greatest accuracy achieved by the VGG16-LSTM model on the UCF dataset was 99.93%, whereas the HAR dataset achieved 99.82%, demonstrating the superiority of the image-based technique. Future approaches for boosting accuracy are highlighted, including the incorporation of facial expressions, real-time tracking improvements, and methodology diversification. This comparison study helps to the progression of understanding post-treatment behaviors in children with ADHD, offering valuable insights for healthcare professionals and caregivers.