Abstract
Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder influenced by genetic, epigenetic, and environmental factors. ASD is characterized by a higher prevalence in males compared to females, highlighting the potential role of sex-specific risk factors in its development. This study aimed to develop sex-specific prenatal and perinatal prediction models for ASD using machine learning and a national population database. A retrospective cohort design was employed, utilizing data from the Korea National Health Insurance Service claims database. The study included 75,105 children born as singletons in 2007 and their mothers, with follow-up data from 2007 to 2021. Twenty prenatal and perinatal risk factors from 2002 to 2007 were analyzed. Random forest models were used to predict ASD, with performance metrics including accuracy and area under the curve (AUC). Random forest variable importance and SHapley Additive exPlanation (SHAP) values were used to identify major predictors and analyze associations. The random forest models achieved high accuracy (0.996) and AUC (0.997) for the total population as well as for the male and female groups. Major predictors included pregestational body mass index (BMI) (0.3679), socioeconomic status (0.2164), maternal age at birth (0.1735), sex (0.0682), and delivery institution (0.0549). SHAP analysis showed that low maternal BMI increased ASD risk in both sexes, while high BMI was associated with greater risk in females. A U-shaped relationship between socioeconomic status and ASD risk was observed, with increased risk in males from lower socioeconomic backgrounds and females from higher ones. These findings highlight the importance of sex-specific risk factors, particularly pregestational BMI, and socioeconomic status, in predicting ASD risk.