Visual Cues to Autism

Abstract

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition characterised by deficits in social interaction and restrictive and repetitive behaviours. ASD is often accompanied by anxiety, elevated rates of depression, and reduced quality of life. Current diagnostic tools are time-consuming, subjective, and less effective in certain groups, particularly young children. Eye tracking has emerged as a promising biomarker, holding the potential to improve diagnostic accuracy and developmental outcomes, thereby reducing the lifetime socio-economic costs of autism. This thesis aims to advance the understanding of gaze behaviour as a biomarker for ASD, thereby driving the development of more accurate, accessible, and scalable diagnostic tools.

A literature review identified fifteen eye tracking features that are associated with ASD diagnosis or Social Responsiveness Scale (SRS) scores. These features were extracted from raw eye tracking data in the Generation R database. Statistical analyses were performed using Generalised Linear Models to evaluate their relationship with ASD diagnosis and SRS scores. Additional analyses addressed feature distributions, gender differences, and the effects of video content and participant age and gender. Features with significant relationships with ASD diagnosis or SRS scores were used in a nested cross-validation framework with predictive machine learning models. The Area Under the Curve (AUC) was the primary performance metric, supported by F1 score, precision, and recall.

Classification proved challenging due to limited discriminative power of individual features. The CatBoost Gradient Boosting Decision Tree achieved the highest performance for ASD prediction with an AUC of 0.71, indicating that gaze-derived features hold promise when used in complex non-linear models. In contrast, the model predicting SRS scores performed worse, with an AUC of 0.57, suggesting that social responsiveness is a part of ASD that may be more effectively masked in adolescents. An alternative explanation is that the features reflect aspects of the autism phenotype that are unrelated to social responsiveness.

Key limitations include the small number of 33 participants with an ASD diagnosis, which reduced statistical power. The reliance on pre-existing clinical diagnoses and SRS scores, which does not capture the full complexity of ASD, led to a lack of ground truth. The adolescent age of the participants posed limitations, as masking behaviours can obscure gaze-based markers. Co-occurring factors such as attention difficulties or cognitive ability were not controlled for. Finally, choices in feature engineering, such as gaze data aggregation, reduced temporal detail; and exclusion of participants with very low screentime may have caused individuals with low social engagement to be underrepresented.

In summary, gaze analysis is a promising tool for understanding and identifying ASD. The findings in this thesis suggest that while individual gaze features offer limited diagnostic power on their own, their integration within more advanced models holds potential to improve the diagnostic process and provide deeper insights into the mechanisms underlying ASD. As datasets grow in size and quality, and stimuli continue to evolve, there are substantial opportunities to uncover more nuanced relationships between gaze behaviour and ASD traits. These developments ultimately support the creation of low-cost, inclusive and scalable ASD assessment tools, thereby enhancing both individual quality of life and broader socio-economic outcomes.