Machine Learning-Assisted Autism Risk Stratification in Toddlers Using the Vietnamese M-CHAT-R and Perinatal Predictors: A Cross-Sectional Study in Vietnam.

introductionThe Modified Checklist for Autism in Toddlers, Revised (M-CHAT-R) is widely used for autism spectrum disorder (ASD) screening; however, evidence on the Vietnamese version and on scalable risk stratification approaches suitable for routine preschool settings remains limited. PURPOSE: This study aimed to evaluate the internal consistency of the Vietnamese M-CHAT-R and to develop machine learning (ML) models integrating M-CHAT-R–derived features with key obstetric–perinatal predictors to support ASD risk stratification in preschool-based screening.

methodsThis cross-sectional study conducted in Ca Mau province, Vietnam, 3,639 children aged 18–36 months were screened using the Vietnamese Ministry of Health–issued M-CHAT-R administered. A mobile clinical team performed onsite DSM-5 evaluations, and the M-CHAT-R/F follow-up was implemented for moderate-risk (scores 3–7) cases.

resultsSix ML algorithms, including Random Forest, were trained using a 75/25 train–test split with SMOTE applied to the training set. Model performance was assessed using accuracy, recall, precision, F1-score, and area under the receiver operating characteristic curve (AUC). 75 children met DSM-5 criteria for ASD (2.0%); 53.3% were male and 96.0% were aged 24–36 months. The Vietnamese M-CHAT-R showed good internal consistency (Cronbach’s alpha = 0.863). Random Forest with SMOTE achieved the best performance (AUC = 0.983; recall = 0.95; precision = 0.351; F1-score = 0.513) using a refined feature set including M-CHAT-R risk level and a composite biological risk index. High recall with moderate precision reflects a screening-oriented emphasis on sensitivity.

conclusionsML- assisted risk stratification may support prioritization for follow-up and specialist assessment in preschool-based pathways in low- and middle-income countries.