Machine Learning Algorithms for Detection of Autism Spectrum Disorders in Early Childhood: A Scoping Review
Xiaoling Gu1,2, Jinghan Jia1,2,3, Xinrui Zhang1,2,3
1First Teaching Hospital of Tianjin University of Traditional Chinese Medicine, Tianjin, China.
Objective:
This scoping review assesses machine learning (ML)-based prediction models for autism spectrum disorder (ASD) in early childhood, with the aim of providing a technical and conceptual foundation for improving early ASD detection.
Methods:
Relevant studies on ML-driven ASD prediction models were systematically retrieved from eight databases: PubMed, Embase, Web of Science Core Collection, Cochrane Library, China National Knowledge Infrastructure (CNKI), China Biomedical Database (CBM), Wanfang Data Knowledge Service Platform (WF), and VIP Chinese Science and Technology Journal Database. The scoping review methodology was strictly followed for data extraction and analysis.
Results:
A total of 16 studies focusing on the application of diverse machine learning algorithms for ASD identification and prediction were included. Among these, 4 studies (25%) employed multiple algorithms for predictive modeling. The most frequently utilized algorithms were tree-based methods (7 studies, 44%), neural networks (NNs) (7 studies, 44%), support vector machines (SVMs) (5 studies, 31%), and regularized logistic regression (3 studies, 19%). Twelve studies (75%) reported Area Under the Curve (AUC) values, all exceeding the 0.7 threshold. Notably, 7 studies (44%) achieved excellent predictive performance with AUC values surpassing 0.9.
Conclusions:
ML-based models hold substantial promise for the early identification of ASD, which is critical for improving patient outcomes. Future research should focus on standardizing ML model frameworks, refining theoretical underpinnings to enhance practical applicability, and promoting clinical implementation following rigorous validation. These efforts will further enhance the accuracy and utility of such predictive models.
