Eye Tracking Scanpaths for Classification of Autism Spectrum Disorder: Leveraging LSTM-Based Models

Abstract

The neuro-developmental condition known as autism spectrum disorder (ASD) is char-acterised by difficulties in social interaction and communication as well as the prevalence of restricted and repetitive behaviours. Early ASD diagnosis is essential for successful intervention and better results. Eye tracking technology has been a useful technique for investigating social cognition and identifying possible ASD signs in recent years. This study uses eye tracking scan pathways captured during a computer-based task to exam- ine the classification of children with ASD and children who are Typically Developed (TD). We suggest using a Long Short-Term Memory (LSTM) model as the core of our classification strategy. With 99.0% accuracy on the training data and 97.4% accuracy on the validation data, the model performs quite well. While participants complete a task requiring social cueing and gaze cueing on a computer screen, eye tracking data is being gathered. The inclusion of social cueing tasks enables us to evaluate the participants’ social cognitive skills, which are known to be impaired in people with ASD. We seek to uncover distinctive patterns and features within the eye tracking scan paths that distin-guish ASD and TD participants by utilising the temporal dynamics collected by LSTM models. The results of this study add to the expanding body of work on using machine learning and eye tracking approaches for diagnosing and evaluating ASD. A non-invasive and objective method for early identification of ASD is provided by the potential for reliably differentiating TD children from children with ASD based on eye tracking scan trajectories, allowing for prompt intervention and support. Furthering our understanding of the underlying mechanisms underlying this complex condition, our findings further offer light on the function of social cognition and gaze cueing in ASD.