Predicting autism traits using Q-CHAT behavioral features: a machine learning approach

Abstract Autism Spectrum Disorder (ASD) is a complex neuropsychiatric developmental condition that significantly affects social interaction, communication, and learning skills. Early and precise diagnosis is essential for timely intervention but remains challenging due to the variability in symptoms and reliance on time consuming clinical assessments. This study proposes machine learning (ML) classifiers for autism trait screening using a toddler dataset with 19 attributes, emphasizing optimized feature selection and hyperparameter tuning. Feature selection techniques, including Chi-squared tests, ML based feature importance, and pairwise correlation analysis, identify key predictive features. Multiple subsets of features are systematically evaluated to optimize classifier performance. Notably, considering only the Q1–Q10 behavioural features resulted in a 41% reduction in features, enhancing classification performance while reducing computational complexity. The Logistic Regression and k-nearest neighbors (LR-KNN) ensemble achieved an accuracy of 99%, followed by the Artificial Neural Network (ANN) at 98%, Logistic Regression (LR) at 97%, and k-nearest neighbors (KNN) at 94%. Through a systematic analysis of feature selection, an optimal feature subset is identified that balances model complexity and predictive power. These findings highlight the role of feature selection in improving ML based screening, offering a scalable and cost effective approach for early detection and personalized intervention strategies. This approach not only streamlines the screening process but also facilitates the way for integrating ML techniques into clinical settings.