Precise and interpretable classification of autism-related behaviors is important for initial diagnosis, personalized intervention, and support arrangements. This study proposes an interpretable machine learning (ML) model using Light Gradient Boosting Machine (LightGBM) and Categorical Boosting (CatBoost) to classify behavioral patterns into four categories (normal, mild, moderate, and severe) associated with Autism Spectrum Disorder (ASD) based on a custom 377-instance survey dataset from Iraqi parents and teachers of children aged 6-12. The model observes 16 key features across communication and social interaction, repetitive behaviors, language, and adaptive skills, preprocessed via interquartile range (IQR) outlier removal, mean imputation, and K-nearest neighbors (KNN) balancing. Shapley Additive Explanations (SHAP) provide instance-level explanations, and Permutation Feature Importance (PFI) quantifies global feature importance. CatBoost had better accuracy (99. 53%) precision recall, and F1-scores (even reaching 1. 00 for some classes), completely outshining LightGBM (97. 65%). This combination of two XAI tools improves clinicians' trust and the practicality of insights, taking ASD assessment that is both accessible and transparent a step further beyond the use of black-box sensor-based models.
