Attention based rewards in LSTM for detection of stimming behaviour in autistic children

Autism Spectrum Disorder (ASD) is a neurological condition characterized by challenges in social communication and sensory processing. Children with ASD often exhibit repetitive or stimming behaviours such as arm flapping, head banging, spinning and repeating words. This study focuses on reward based deep learning techniques to analyse and differentiate stimming behaviours in children with ASD and typically developing children. The objective is to develop an advanced recognition model to improve action detection in ASD populations. This research proposes a Long Short-Term Memory (LSTM) network enhanced with a novel attention-based reward mechanism to detect stimming behaviours from video data. The LSTM network captures long-term dependencies, while the attention mechanism highlights the most relevant features in each frame. The model incorporates three distinct reward functions designed to improve prediction performance by optimizing correctness, diversity, class weightage and attention effectiveness. The reward-enhanced LSTM model achieves improved classification performance, attaining accuracies of 81.01%, 83.54% and 88.84%, compared to 79.26% in a baseline setting without the reward mechanism. This study highlights the effectiveness of a reward-integrated LSTM framework for recognizing stimming behaviours in children with ASD. By leveraging both temporal dependencies and attention-driven feature selection, the proposed approach outperforms conventional models, providing a robust solution for action recognition in ASD diagnosis.