Mapping the Landscape of ASD-AI: Multimodal Gains, XAI Adoption, and Fairness Gaps - A Systematic Review

Artificial intelligence (AI) is increasingly used to automate the detection of autism spectrum disorder (ASD), yet most research still focuses on maximizing classification accuracy. This narrow emphasis leaves a critical gap in the systematic synthesis of model explain-ability and algorithmic fairness, both essential for clinical adoption. Following PRISMA guidelines, this systematic review examines AI-based ASD detection from three perspectives: (i) the use of explain-ability techniques, (ii) performance gains from multi-modal fusion, and (iii) equity across demographic subgroups. A comprehensive search of major databases (including PubMed, IEEE Xplore, Scopus and the ACM Digital Library) yielded 10,117 records; after screening, 45 primary studies met the inclusion criteria. Three findings emerge: (1) when data are reasonably harmonized, multi-modal fusion (e.g., combining neuroimaging with phenotypic data) delivers accuracy gains of roughly 10-25\% over uni-modal approaches, but multi-site heterogeneity often diminishes these benefits; (2) Explainable AI (XAI) appears in 71.1\% of studies, yet methods such as SHAP and Integrated Gradients are largely used to validate group-level biomarkers rather than to provide actionable, patient level explanations required for clinical decision-making; and (3) despite growing awareness of algorithmic bias, only 40\% of studies conduct formal fairness audits, leaving underrepresented groups, especially women and racially diverse populations, at continued risk of healthcare disparities. Overall, despite accuracy improvements, integration of explain-ability and fairness remains nascent. We synthesize these insights into a framework that advocates a shift toward human-centered AI, jointly addressing accuracy, explain-ability, and fairness to enable clinically trustworthy and equitable translation.