Generating and Detecting Autistic Facial Expression Patterns Using Generative AI and Deep Learning

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental conditioncharacterized by distinct behavioral traits and subtle differences in facial morphologyand expression patterns. Early and reliable detection of ASD is crucialfor enabling timely interventions and mitigating its effects. This study is a preliminaryexploration aimed at demonstrating the methodological feasibility ofleveraging Generative AI (Gen-AI) and Generative Adversarial Networks (GAN),specifically StyleGAN2-ADA, to synthesize realistic facial expressions of bothautistic and neurotypical individuals, with a focus on generating facial images ofchildren. This approach seeks to explore the monumental challenges in autismrelatedfacial analysis, including data scarcity, ethnic imbalance, and the ethicalconcerns surrounding the collection of facial data from minors. Continuing alongthis exploratory path, we develop a deep convolutional neural network classifierbased on the EfficientNetB3 architecture to learn and classify images based on subtle morphological and expressive features present in datasets labeled asbelonging to individuals diagnosed with ASD. The experimental results indicatethat the generated synthetic data captures statistical patterns present inthe training dataset, which highlights the potential of the proposed methodology—centered on the use of high-fidelity synthetic datasets—to enhance theperformance, diversity, and generalization of autism classification models. Oncevalidated on clinical datasets, this approach could serve as a valuable assistivetool for AI-driven screening and support systems, paving the way for morerobust, equitable, and clinically relevant applications in autism assessment andintervention.