Classification Autism Severity Using a Multidimensional Behavioral Datasets

Abstract

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by significant heterogeneity, making the accurate classification of its severity levels crucial for effective intervention. This study develops and validates a comprehensive machine learning framework to classify ASD severity (mild, moderate, severe) by investigating the differential impact of feature engineering and selection. A dataset of 340 individuals was analyzed using a dual-framework approach, integrating supervised classification (SVM, Random Forest, XGBoost) and unsupervised clustering (K-Means, Hierarchical, GMM). The methodology centered on comparing a feature-centric approach, using a hybrid selection strategy on 60 engineered and original features, against a baseline approach using only 33 original features. The feature-centric approach yielded markedly superior results; a Random Forest classifier, trained on a minimal subset of just nine engineered features, achieved a test accuracy of 76.5%, significantly outperforming all other models, including an SVM trained on 22 original features (70.6% accuracy). This highlights that feature quality is more critical than quantity. The unsupervised analysis revealed a critical “evaluation paradox,” where radical, unguided feature reduction improved geometric cluster cohesion but degraded clinical accuracy. Conversely, a guided, domain-informed selection improved both internal and external metrics.