Web Tabanlı Göz İzleme Verilerini Kullanarak Yetişkinlerde Otizmi Tespit Etmek İçin Derin Öğrenme

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition characterized by social-communication difficulties, repetitive behaviors, and restricted interests. Detection supports timely intervention and improved outcomes. Eye-tracking combined with Artificial Intelligence offers a promising solution. While various Machine Learning (ML) approaches have been explored for ASD diagnosis, Deep Learning (DL) has not previously been applied to web-page interaction eye-tracking datasets. Web page interactions offer insights into cognitive and attentional patterns through naturalistic tasks reflecting real-world scenarios, making them crucial for studying ASD-related traits. This study explores DL methods, including STA (Scanpath Trend Analysis)-enhanced models, for ASD prediction using two web eye-tracking datasets (Dataset 1 and Dataset 2) and a publicly available children’s dataset (Dataset 3). We re-implemented a DL baseline from Dataset 3 and optimized it via hyperparameter tuning, architectural variations, and comparison with a Gated Recurrent Unit (GRU) model. On Dataset 2, the enhanced GRU achieved accuracies of 74.42% and 77.67% for the browsing and synthesis tasks. On Dataset 1, our STA-enhanced GRU reached 78.80% for the browse task, outperforming all literature baselines, and ranked second on the remaining tasks. Overall, GRU/Long Short-Term Memory models perform best on the more complex Dataset 2, while STA-enhanced models boost performance on Dataset 1. This research demonstrates that DL, can effectively be applied to web-based eye-tracking datasets.