Deep Learning Approaches for Autism Spectrum Disorder: A Comprehensive Review

Abstract

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental disorder with a complex etiology that presents obstacles to an early and reliable diagnosis. Recent innovations in deep learning (DL) have revolutionized ASD research by allowing the automatic identification of subtle, non-linear signatures from noisy and data-rich heterogeneous and multimodal sources. In this work, we provide the first systematic comparative study of 23 recent DL models for ASD detection, spanning multiple multimodal datasets from facial imaging, neuroimaging, functional magnetic resonance imaging (fMRI)/electroencephalogram (EEG), eye-tracking, kinematic profiles, and electronic health records (EHRs) via novel fusion architectures. Modern architectural designs such as Vision Transformers (ViTs), hybrid DL frameworks, attention-augmented models, and ensemble strategies achieve consistently high diagnostic accuracies that notably surpass the capabilities of classical machine learning (ML) approaches. Secondly, the improvement in transfer learning and multimodal fusion, together with interpretability methods (e.g., gradient-weighted class activation mapping (GradCAM) and attention heatmaps), makes features representable and strengthens the generalization of models. Although formidable progress has been made in this area, the adoption of real-world applications is still hampered by the heterogeneity of datasets, the lack of demographic representation and diversity among validation cohorts, and variations in clinical environments. This will necessitate the adoption of standardized pipelines, strict cross-population validation, and explanatory AI frameworks designed to ensure transparency and clinical utility. In sum, realizing the transformative capacity of DL for ASD Dx relies on ongoing interdisciplinary partnerships among computer scientists, clinicians, and neuroscientists to deliver new solutions that are both accurate and inclusive in nature (i.e., they need to be clinically relevant), as well as principled from an ethical point of view.