Hybrid deep learning architecture for autism spectrum disorder detection from gait kinematic data

Abstract

Early detection of autism spectrum disorder (ASD) is essential for enabling timely interventions that significantly impact developmental outcomes. Traditional diagnostic procedures often rely on clinical observations and subjective assessments, which can reduce objectivity and delay diagnosis. This study presents an automated classification approach for detecting ASD in children using gait kinematic data obtained through video-based skeletal tracking. We propose a hybrid deep learning architecture that combines one-dimensional convolutional layers (Conv1D), bidirectional long short-term memory (BiLSTM) networks, and a multi-head attention mechanism to capture complex spatiotemporal patterns in motion data. The dataset includes 100 participants—50 diagnosed with ASD and 50 typically developing (TD) peers. Preprocessing steps included Euclidean norm transformation, logarithmic scaling, Z-score normalization, and sliding window segmentation. The proposed model achieved 94.9% accuracy, 0.91 recall, 0.86 precision, and an area under the curve (AUC) of 0.97, outperforming a baseline long short-term memory (LSTM) architecture. These findings demonstrate the potential of gait-based kinematic features and hybrid neural networks for objective and reproducible ASD screening. The developed system can contribute to AI-assisted decision support tools in clinical and educational environments, enhancing diagnostic accuracy, and supporting inclusive developmental care.