A Comprehensive Survey of Explainable Multimodal deep Learning Methods for Video-Based Autism Spectrum Disorder Detection

Autism Spectrum Disorder (ASD) is a complex, neuro-developmental disorder that impacts communication, social interaction and behaviour. The early diagnosis is still a difficult process because of the subjective clinical judgments and insufficient observation of behaviors. Video-based behavioral analysis has been the most recent development in the field of automated ASD screening, with the aid of artificial intelligence (AI) technologies such as multimodal deep learning and explainable artificial intelligence (XAI). A general review of the recent work on the ASD detection frameworks developed from 2022-2026 on visual, temporal and audio modalities is provided in this survey. Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, Vision Transformers (ViTs), Graph Neural Networks (GNNs), self-supervised learning, federated learning and lightweight edge-AI architectures are explored. Additionally, the approaches of explain ability such as Grad-CAM, Saliency maps, SHAP and visualization of attention are analyzed to enhance transparency and reliability in clinical use. By comparing with unimodal systems, it is found that the multimodal fusion system provides superior detection accuracy, robustness and interpretability. This survey furtherly provides mathematical expressions, standard architectural modules, a flowchart of mathematical models using TikZ package, a comparison table of experiments, and future research directions for privacy-preserving and real-time ASD diagnosis systems. This presented work is designed to offer a systematic and up-to-date reference for researchers who are creating reliable AI-based ASD diagnostic regimes.