Artificial intelligence for autism diagnosis: Algorithms, data and challenges

Abstract

Autism Spectrum Disorder (ASD) presents diverse and complex diagnostic challenges, traditionally reliant on subjective behavioral assessments. Artificial Intelligence (AI) offers a transformative approach to enhancing ASD diagnosis by enabling objective, data-driven insights from multimodal sources. This paper examines the breadth of available datasets, including neuroimaging, genetics, behavioral assessments, and electronic health records (EHRs), and provides a comprehensive review of current AI methodologies used in ASD diagnosis, including traditional machine learning, deep learning architectures, and ensemble techniques. The paper explores the parallel evolution of data modalities and AI techniques, as well as critical issues such as data heterogeneity, privacy concerns, and standardization gaps. Key technical challenges—such as model generalizability, explainability, and ethical considerations surrounding bias and transparency—are also addressed. To assess real-world applicability, we present a preliminary study using structured EHR data from Parc Sanitari Sant Joan de Déu. Our findings reveal limited predictive power from structured variables alone. Therefore, future work will focus on mining clinical insights by applying Natural Language Processing techniques to the unstructured EHR data available. This work offers a roadmap for advancing AI-driven ASD diagnosis toward more robust, interpretable, and clinically relevant systems.