Comparative Deep Learning–Based Facial Image Analysis for Early Autism Prediction in School-Aged Children

Abstract

Abstract: Early identification of Autism Spectrum Disorder (ASD) in school-aged children is critical, as timely intervention has been shown to markedly enhance developmental trajectories. This study investigates the feasibility of facial image analysis for ASD screening by leveraging four pre-trained convolutional neural network (CNN) architectures—VGG-16, InceptionV3, EfficientNet-B0, and EfficientNet-B7—applied to a balanced dataset comprising 2,540 labeled facial images (1,327 autistic and 1,327 non-autistic), curated from a publicly available Kaggle repository. VGG-16 yielded the highest classification accuracy at 84.33%, followed closely by EfficientNet-B0 (83.67%), InceptionV3 (81.00%), and EfficientNet-B7 (80.00%). To assess the robustness of these findings, we conducted five independent training runs per model, followed by statistical significance testing using one-sample t-tests and one-way ANOVA. All models significantly outperformed the chance baseline (p < 0.05), though pairwise differences in accuracy did not reach statistical significance at the α = 0.05 level. Unlike many prior studies that employed limited or imbalanced datasets, or assessed only a single architecture, this work offers a systematic comparative evaluation under uniform training conditions with a specific focus on school-aged populations. The results suggest that CNN-based facial analysis holds promise as a non-invasive, scalable adjunct screening method, particularly suited for deployment in educational contexts where clinical resources may be constrained.