Virtual Reality Based Adaptive Game for Enhancing Joint Attention in Children With Autism

Atypical brain development is a characteristic of autism spectrum disorder (ASD), a neurodevelopmental disorder that frequently shows up as deficits in joint attention, a crucial social-communication ability including gaze sharing and gaze following. Children with ASD frequently exhibit atypical gaze behaviours that hinder effective coordination of attention with others. Although recent studies have explored virtual reality (VR) to address these aspects, the existing systems rely primarily on performance-based learning and offer limited adaptability to individual behavioural patterns. To overcome these limitations, this study proposes a novel virtual reality and adaptive game-based system (VRAGS) designed to assess and enhance attention in children with ASD, thereby supporting improved learning capacity. The proposed framework begins with data acquisition through the 3D game environment. The collected data are preprocessed using an improved tanh normalization, which effectively suppresses noise while preserving informative variations for learning. Then, the feature extraction process takes place, in which the raw data features, statistical features and improved entropy features are derived. The improved entropy is computed using information gain, prioritizing features that maximally reduce uncertainty with respect to attention states, thereby enhancing discriminative power. Finally, the attention prediction is done via the hybrid classification (HC) model that integrates an Adam-optimized deep convolutional neural network (AODCNN) and a bidirectional long short-term memory (Bi-LSTM) network. This architecture collectively captures hierarchical spatial patterns and temporal dependencies within the data. The integration of the improved tanh activation and Adam optimization in the AODCNN model ensures stable convergence and robust feature learning. Moreover, the experimental evaluation demonstrates that the proposed model significantly outperforms existing methods, achieving an accuracy of 0.933, an F-measure of 0.925 and a negative predictive value (NPV) of 0.936. These findings demonstrate the potential of the suggested framework as a customized, adaptable tool for therapeutic intervention and support its efficacy in precisely predicting attention levels in children with ASD.