Improving Machine Learning based ASD Diagnosis with Effective Feature Selection

Autism spectrum disorder (ASD) poses major public health challenges due to its increasing prevalence and the critical need for early intervention to improve outcomes. This study investigates machine learning approaches for automated screening and diagnosis of ASD, leveraging questionnaire and demographic data from large cohorts of toddlers (n=1,054) and adults (n=704). After data cleaning and preprocessing, an efficient sequential feature selection technique identifies subsets of the most predictive features. Several classification algorithms, including neural networks, random forests, support vector machines, and logistic regression, are developed using these refined datasets and rigorously evaluated through 10-fold cross-validation. The models demonstrate excellent predictive performance, achieving 100% accuracy in identifying ASD in both toddlers and adults. This signifies that the selected feature sets are highly relevant for capturing autistic traits across age groups, although manifestations vary developmentally. Compared to existing methods, the models show substantially improved accuracy, highlighting machine learning’s potential for accessible and scalable assessment of ASD. Additional validation on bigger real-world datasets would further establish generalizability. Overall, this study demonstrates that machine learning can assist evidence-based clinical decision-making for ASD diagnosis. Promising future work should focus on enhancing model interpretability, translating these technologies responsibly to bolster early intervention, and evaluating impact on patient outcomes through rigorous trials. Machine learning shows immense promise in aiding the identification and timely treatment of this import yet often underdiagnosed public health issue.