Multi-classification of autism spectrum disorderbehavior for children using explainable artificialintelligence techniques
Wisal Hashim Abdulsalam, Rasha H. Ali
Precise and interpretable classification of autism-related behaviors is importantfor initial diagnosis, personalized intervention, and support arrangements. This studyproposes an interpretable machine learning (ML) model using Light GradientBoosting Machine (LightGBM) and Categorical Boosting (CatBoost) to classifybehavioral patterns into four categories (normal, mild, moderate, and severe)associated with Autism Spectrum Disorder (ASD) based on a custom 377-instancesurvey dataset from Iraqi parents and teachers of children aged 6-12. The modelobserves 16 key features across communication and social interaction, repetitivebehaviors, 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, andPermutation Feature Importance (PFI) quantifies global feature importance. CatBoosthad better accuracy (99. 53%) precision recall, and F1-scores (even reaching 1. 00for some classes), completely outshining LightGBM (97. 65%). This combination oftwo XAI tools improves clinicians' trust and the practicality of insights, taking ASDassessment that is both accessible and transparent a step further beyond the use ofblack-box sensor-based models.