DASD- diagnosing autism spectrum disorder based on stereotypical hand-flapping movements using multi-stream neural...

The early detection and diagnosis of autism spectrum disorder (ASD) remain critical challenges in developmental healthcare, with traditional diagnostic methods relying heavily on subjective clinical observations.

In this paper, we introduce an innovative multi-stream framework that seamlessly integrates three state-of-the-art convolutional neural networks, namely, EfficientNetV2B0, ResNet50V2, DenseNet121, and Multi-Stream models to analyze stereotypical movements, particularly hand-flapping behaviors automatically. Our architecture incorporates sophisticated spatial and temporal attention mechanisms enhanced by hierarchical feature fusion and adaptive temporal sampling techniques designed to extract characteristics of ASD related movements across multiple scales. The system includes a custom designed temporal attention module that effectively captures the rhythmic nature of hand-flapping behaviors. The spatial attention mechanisms method was used to enhance the proposed models by focusing on the movement characteristics of the patients in the video. The experimental validation was conducted using the Self-Stimulatory Behavior Dataset (SSBD), which includes 66 videos.

The Multi-Stream framework demonstrated exceptional performance, with 96.55% overall accuracy, 100% specificity, and 94.12% sensitivity in terms of hand-flapping detection and an impressive F1 score of 97%.

This research can provide healthcare professionals with a reliable, automated tool for early ASD screening that offers objective, quantifiable metrics that complement traditional diagnostic methods.