Introduction:
Autism spectrum disorder (ASD) is a developmental brain condition that causes problems with social interaction and communication, as well as repetitive behaviors. Early ASD screening is vital for prompt ac tions, as existing diagnostic methods suffer from scalability limitations in resource-constrained settings. Herein, a machine learning (ML)-based explain able predictive model was developed and evaluated for early screening of ASD in toddlers using Q-CHAT-10 behavioral, demographic, and clinical features.

Material and methods:
A total of 1054 toddlers’ records sourced from the Autism Screening for Toddlers dataset freely available at Kaggle were ret rospectively analyzed. Data preprocessing, statistical feature selection, and dimensionality reduction were performed. Multiple ensemble models were trained using Q-CHAT-10 behavioral features combined with demographic and clinical variables. Several algorithms were tested, including Logistic Re gression, Random Forest, Gradient Boosting, and Multilayer Perceptron, with k-fold cross-validation for model selection. SHAP analysis was employed to explore the reasons behind individual predictions.

Results:
Model performance was evaluated using ROC-AUC. Feature im portance was checked to identify the most predictive items. The Gradient Boosting classifier achieved the best performance, with an accuracy of 0.98 (95% CI: 0.85–0.93), sensitivity of 0.91, specificity of 0.87, and ROC-AUC of 0.94 on the held-out test set. SHAP analysis revealed total Q-CHAT-10 score, response to name, pointing to share interest, and pretend play as the most influential predictors.

Conclusions:
This ML framework accurately detects ASD traits in toddlers, highlighting the potential of a scalable, low-cost screening tool to enable early ASD detection and improve equitable access to pediatric care. How ever, external validation across diverse populations with larger samples is warranted before clinical application can be recommended.