Comparing Machine Learning Models for Pupillary Light Reflex–Based Autism Spectrum Disorder Classification

Autism spectrum disorder (ASD) prevalence is rising globally, underscoring the critical need for early detection to improve intervention outcomes. Atypical pupillary light reflex (PLR) has emerged as a promising biomarker for this purpose. This study leverages machine learning to enhance the accuracy of early ASD screening in children aged 3–6. We developed and evaluated five classifiers—Naive Bayes, logistic regression, decision tree, linear discriminant analysis (LDA), and deep neural network (DNN)—using PLR data from 25 children with ASD and 50 typically developing peers. Model performance was assessed across 15 metrics, including accuracy, F1 score, ROC–AUC, and computational efficiency. The comprehensive evaluation demonstrated that the DNN outperformed all other models, achieving the highest test accuracy of 85% and excelling across multiple key metrics. These findings confirm the DNN’s superiority for automated ASD screening based on PLR, contributing significantly to the development of objective, efficient tools for early autism diagnosis.