Knowledge-Based Label Construction and Machine Learning Classification for ASD Technology Recommendation

Abstract

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder with heterogeneous behavioral profiles, including impairments in communication, social interaction, repetitive behaviors, and sensory sensitivities. Public ASD datasets only provide binary diagnostic labels without technology intervention recommendations, leaving behavioral profiles underutilized for recommendation purposes. This study proposes a knowledge-based labeling approach using soft scoring to construct five technology recommendation labels: AAC, Serious Game, Social Robot, Sensory App, and Video Modeling, from AQ-10 behavioral profiles derived from verified clinical literature. The dataset combines UCI ML Repository (ID=419) and Kaggle All Ages Combined, totaling 5,505 instances with 14 input features. Seven machine learning algorithms were compared using test-set evaluation, 10-Fold Stratified Cross-Validation, hyperparameter tuning, and SHapley Additive exPlanations (SHAP), with class imbalance handled via Synthetic Minority Oversampling Technique (SMOTE). Tuned Gradient Boosting achieved the best performance with accuracy 96.73%, F1-macro 96.54%, and CV accuracy 94.68% ± 0.79%. SHAP analysis confirmed prediction patterns consistent with the knowledge-based labeling design, while demographic features contributed minimally. The system is designed as a decision-support tool for practitioners, not a substitute for comprehensive clinical assessment.