Artificial Intelligence for Early Autism Detection: Current Advances, Challenges and Future Directions

Abstract

Autism spectrum disorder (ASD) is a developmental disorder that must be identified early but traditional diagnosis can be examiner-dependent, time-consuming and not accessible. This study included a dual bibliometric and systematic review on the evidence supporting the application of artificial intelligence (AI) tools in the detection of ASD according to PRISMA 2020 and SWiM guideline. A total of 41 empirical studies were identified after a search in the Scopus database. The studies were conducted between 2011 and 2026, in 11 data modalities, 7 algorithm families and 28 countries. The accuracy of the reports was between 68.18% and 100% with a median of 91%. Facial image research was the biggest modality, but all nine studies used the same public Kaggle dataset. Unstated validation designs occurred in 46% of studies, and 93% were at high risk of bias in the areas of patient selection and reference standards, and unclear flow and timing. Very high accuracy scores were massed in small, one cohort datasets, with population level studies obtaining significantly lower accuracy scores. But in general, there was only a small difference in median accuracy between the externally validated studies and the single-cohort studies (1.5 percentage points). The results suggest that the use of AI to detect ASD is not yet ready for clinical use. Transparent reporting of validation, independent external cohorts, diverse datasets, and prospective testing of positive predictive value in realistic population prevalence should be the focus of future research.