Research on the severity of symptoms in children with ASD based on integrated machine learning and structural equation modeling: age-specific predictive features and mediation effect path analysis.

Background: The correlation between age at first diagnosis and symptom severity in autism spectrum disorder (ASD) remains inconsistent in existing research, and the age-specific mechanisms through which developmental level and physical development mediate this relationship remain to be elucidated. Methods: A total of 608 children with ASD aged 24∼71 months at first diagnosis were recruited and stratified into a low-age group (24∼47 months) and a high-age group (48∼71 months). The Gesell Developmental Schedule (GDS) was utilized to assess developmental level, the Childhood Autism Rating Scale (CARS) to evaluate symptom severity, and height-for-age Z-score (HAZ), weight-for-age Z-score (WAZ), and body mass index-for-age Z-score (BAZ) were calculated to assess physical development. Integrated machine learning was applied to construct classification models for distinguishing severe from non-severe ASD and to identify age-specific predictive features, and structural equation modeling (SEM) was used to analyze the mediation effects and path mechanisms. Results: Machine learning revealed language as the core predictive feature for the low-age group; whereas in the high-age group, the predictive importance of personal-social, fine motor, gross motor, HAZ, and WAZ increased significantly, with a relative decline in the predictive importance of language. SEM showed a masking effect in the total sample: age at first diagnosis had a positive total indirect effect on symptom severity via developmental level and physical development, but a negative direct effect that offset the total indirect effect, resulting in a non-significant total effect. The mediating role of developmental level was significant, and the effect of developmental level on symptom severity was more pronounced in the low-age group. Conclusion: This study identifies age-specific predictive features-language as the core predictive feature in the low-age group, whereas in the high-age group, the importance of personal-social, motor, and physical development increases while that of language declines. It also reveals a masking effect of age at first diagnosis on ASD symptom severity: a negative direct effect offset by a positive indirect effect via developmental level and physical development, resulting in a non-significant total effect. These findings provide a preliminary reference for age-stratified intervention in children with ASD.