Artificial Intelligence Methods in Early Detection of Autism Spectrum Disorder: A DSM-5 Criterion-Based Systematic Review.

Can Artificial Intelligence (AI) revolutionize early detection of Autism Spectrum Disorder (ASD) by offering a more objective alternative to the subjective behavioral assessments? This systematic review evaluates AI-based methods for early ASD detection in children, assessing the distribution, performance, and DSM-5 alignment of AI techniques, and the methodological quality of existing studies. Following PRISMA guidelines, seven databases (PubMed, Web of Science, Scopus, IEEE Xplore, PsycNET, CINAHL, and Cochrane Library) were searched from January 2015 to March 2025; no backward or forward citation tracking was conducted. Studies were included if they focused on pediatric populations (ages 0-18) with clinically diagnosed ASD, employed AI detection/classification methods, utilized observable behavioral data, and reported diagnostic performance metrics. Studies using exclusively neurobiological data, adult populations, non-English publications, or lacking clear diagnostic reference standards were excluded. Two reviewers independently screened 1018 records; 43 met inclusion criteria. Methodological quality was assessed using QUADAS-2, adapted for machine learning considerations. Classical Machine Learning (34%), Deep Learning (34%), and Hybrid approaches (32%) were equally prevalent. Convolutional Neural Networks and Support Vector Machines dominated for unstructured and structured data, respectively. Most studies (74%) targeted both DSM-5 Criteria A and B; 23% focused solely on Criterion A. Direct DSM-5 alignment correlated with higher accuracy (median ~95%); overall accuracy ranged from 68.18% to 100%. Methodological concerns included 57% unclear patient selection and 43% unclear index test risk. AI demonstrates transformative potential for early ASD detection, particularly for social communication deficits. Gaps in Criterion B coverage, geographic skew toward high-income settings, and methodological inconsistencies limit clinical applicability. Future work should prioritize standardized protocols, multimodal integration, and diverse external validation. Trial Registration: PROSPERO registration ID: CRD420250656126.