A systematic review of machine learning techniques to address diagnosis and treatment of autism: challenges and opportunities

Autism spectrum disorder (ASD) is a complex developmental disability marked by challenges in social communication, interaction, and behavioral patterns. Because its exact causes remain unclear, early diagnosis and personalized treatment rely heavily on identifying subtle features and correlations within patient data. Traditional knowledge-based approaches have increasingly been superseded by computational psychiatry frameworks leveraging machine learning (ML). This paper presents a systematic review evaluating 55 peer-reviewed studies published from 2017 to 2023. The primary objective is to examine how recent ML applications address ASD diagnosis and treatment, identifying dominant techniques, trends, dataset characteristics, and current methodological limitations.

The authors conducted a systematic literature search following PRISMA guidelines across major academic databases, specifically ScienceDirect, Scopus, and MDPI, using keywords related to autism and machine learning. Out of 163 initial records, the selection process filtered out duplicates, non-peer-reviewed materials, non-English articles, and studies outside the ASD scope, resulting in 55 finalized studies. These works were evaluated across eight structured research questions focusing on publication geography, dominant ML approaches, specific algorithms, data types, sample sizes, and database utilization.

The review demonstrates that supervised learning methods are the most frequently employed approach due to their alignment with diagnostic classification needs, offering high accuracy when labeled data are available. However, deep learning is experiencing rapid growth, driven by larger and more complex datasets. Additionally, hybrid algorithms—which combine supervised or unsupervised techniques with knowledge-based systems or fuzzy logic—are emerging as powerful tools for capturing nuanced behavioral patterns. Regarding data types, brain imaging data dominates the literature at 45 percent, followed by clinical assessments at 25 percent and eye-tracking metrics at 18 percent. Despite these technological advancements, the field faces significant challenges, including a heavy reliance on labeled data, potential overfitting, and a lack of large, standardized, and multimodal datasets that reflect the true diversity of the autistic spectrum.