Early Detection of Autism Spectrum Disorder in Toddlers: A Fast and Efficient Machine Learning Approach

Abstract

Autism Spectrum Disorder (ASD) significantly affects cognitive, social, and
behavioral development in early childhood, making timely and accurate diagnosis critical for
improving long-term outcomes. Early intervention has been shown to substantially mitigate
the impact of ASD on daily functioning. In this study, we propose a lightweight and highly
accurate machine learning framework for early ASD detection using the Toddlers dataset. To
enhance data quality and predictive performance, several preprocessing techniques were
employed, including SMOTETomek for class balancing, independent t-tests for statistical
feature analysis, and MinMaxScaler for data normalization. Feature selection was performed
using Chi-squared scoring to identify the most discriminative attribute for classification. We
evaluated eight machine learning classifiers—Decision Tree (DT), CatBoost, XGBoost,
Voting, Stacking, Extra Trees, Gaussian Naïve Bayes, and Bernoulli Naïve Bayes—to
determine the most effective model for ASD prediction. Experimental results demonstrate
that the Decision Tree classifier outperformed all other models as well as existing state-ofthe-
art approaches, achieving 100% accuracy in early ASD detection. The proposed
framework offers a computationally efficient and reliable solution that can support healthcare
professionals in automating and enhancing early ASD screening processes