Exploring Gaze Patterns for Autism Diagnosis with LSTM and Attention Mechanisms

Background: Eye-tracking technology objectively measures eye fixations and movements to capture attention to stimuli. Observing gaze behavior and eye movements in different regions provides insights into how people inter- act visually with particular elements in a given context. Re- searchers can better understand visual attention allocation, and professionals can improve user experiences in advertis- ing, web design, psychology, usability testing, and human- computer interaction. In eye-tracking technology, special- ized hardware and software record and analyze gaze data within these predefined regions to shed light on cognitive processes, preferences, and behaviors Objectives: The main objective of our study is to classify children into three categories: Typically Developed (TD), High IQ Autism, and Low IQ Autism, using the gazing points. We are studying children with and without ASD to achieve this goal. This study investigates the potential of using eye-gaze patterns to stratify ASD and TD groups based on spatial and spatiotemporal scan paths. Methods: In this study, we will employ eye-tracking technology to conduct a comprehensive analysis of the scan path of the eyes.We will leverage deep learning techniques to process and interpret eye-tracking data, specifically Long Short-Term Memory (LSTM) networks. LSTM networks are well-suited for sequential data analysis, making them an ideal choice for modeling the temporal aspects of eye-tracking sequences. We aim to use LSTM to develop a reliable and precise classification model for early intervention and diagnosis of autism spectrum disorders. Eye gaze data was collected from 86 participants, of which 54 had ASD, and 32 had typical development. The participants ranged from 3 to 12 years old, and all were part of a study focused on children with ASD and TD. The study used several eye-tracking video scenarios to collect data, including Activity Monitoring (AM), Social Referencing (SR), Theory of Mind (ToM), and Dyadic Bid (DB) tasks. These tasks were explicitly designed to evaluate different aspects of social cognition, such as the ability to track the gaze of others, understand intentions, and respond to so cial cues. The dataset comprises 1,001,352 data points, providing a rich source of information for analyzing and understanding the participants’ eye gaze behavior. Results: The high accuracy of 91% demonstrates the effectiveness of using LSTM networks for ASD classification, outperforming traditional and otherdeep learning models Conclusions: The paper concludes that combining eye-tracking technology with deep learning techniques has excellent potential for classifying ASD. The LSTM model have shown promising results, allow- ing further refinement and optimization. This study highlights the significance of visual processing in individuals with ASD and contributes to the ongoing efforts to leverage advanced technologies for ASD diagnosis