Diagnosing Autism Spectrum Disorder Based on Eye Tracking Technology Using Deep Learning Models

Mosleh Hmoud Al-Adhaileh^1*Saleh N. M. Alsubari²Abdullah H. Al-Nefaie¹Sultan Ahmad^3*Asma Abdulmana Alhamadi⁴

• ¹King Salman Center for Disability Research, Riyadh, Saudi Arabia
• ²Department of computer science , college of technology and business, Riyadh Elem University, Saudi Arabia, Riyadh, Saudi Arabia
• ³Department of Computer Science, College of Computer Engineering and Sciences, Prince Sattam Bin Abdulaziz University, Alkharj, Saudi Arabia
• ⁴Department of Basice Sciences College of Science & Theoretical Studies Saudi Electronic University, Riyadh, Saudi Arabia

Introduction: Children with Autism Spectrum Disorder (ASD) often find it difficult to maintain eye contact, which is vital for social communication. Eye tracking (ET) technology helps determine how long children with ASD focus on someone, how frequently they do so, and in which direction their gaze moves. ET provides insights into social attention by enabling precise, real-time tracking of gaze patterns as individuals process social information visually. It is a dependable method for identifying and developing social attentional biomarkers, particularly in challenging conditions like ASD. Objective: This study aims to implement deep learning (DL) algorithms using eye-tracking data from social attention tasks involving children with ASD. Methods: The approach was tested using standard datasets collected from individuals with and without ASD through eye-tracking technology. Convolutional neural networks (CNNs) and long short-term memory (LSTM) models were used to analyze data from children with ASD. Data preprocessing techniques addressed missing data and converted categorical features into numerical values. Mutual information-based feature selection was employed to reduce the feature set by identifying the most relevant features, thereby improving system performance. These features were then analyzed using LSTM and CNN-LSTM models to evaluate their potential for diagnosing ASD. Results: The experimental results showed that the highest accuracy achieved was 99.78% with the CNN-LSTM model. Furthermore, the findings indicated that the proposed method outperformed previous studies. Conclusions: The system successfully diagnosed ASD using the ET dataset. This approach shows promise for clinical application, assisting healthcare professionals in diagnosing ASD more accurately through advanced artificial intelligence technology.