Using Explainable Machine Learning and Eye-Tracking for Diagnosing Autism Spectrum and Developmental Language Disorders in Social Attention Tasks

Adoración Antolí^1,2*Francisco Javier Rodríguez-Lozano³José Juan Cañas⁴Julia Vacas^1,2Fatima Cuadrado^1,2Araceli Sánchez-Raya^1,2Carolina Pérez-Dueñas^1,2Juan Carlos Gámez Granados³

• ¹Department of Psychology, University of Córdoba, Córdoba, Spain
• ²Maimónides Biomedical Research Institute of Córdoba, Córdoba, Spain
• ³Department of Electronic and Computer Engineering, Higher Polytechnic School, University of Córdoba., Córdoba, Spain
• ⁴Mind, Brain and Behavioral Research Center, University of Granada, Granada, Spain

Background: Eye-tracking technology has proven to be a valuable tool in detecting visual scanning patterns associated with autism spectrum disorder (ASD). Its advantages in easily obtaining reliable measures of social attention could help overcome many of the current challenges in the assessment of neurodevelopmental disorders.However, the clinical use of this technology has not yet been established. Two key challenges must be addressed: the difficulty in reliably distinguishing between disorders with overlapping features, and the efficient management of eye-tracking data to yield clinically meaningful outcomes. Purpose: The aim of this study is to apply explainable machine learning (XML) algorithms to eye-tracking data from social attention tasks involving children with ASD, developmental language disorder (DLD), and typical development (TD), in order to assess classification accuracy and identify the variables that best differentiate between groups. Methods: Ninety-three children participated in a visual preference task that paired social and non-social stimuli, specifically designed to capture features characteristic of ASD. Participants were distributed across three groups: ASD (n = 24), DLD (n = 25), and TD (n = 44). Eye-tracking data were used to generate four datasets, which were then analyzed using XML algorithms to evaluate the accuracy of group classification across all possible combinations. Results: The model achieved an F1-score of 0.912 in distinguishing DLD from TD, 0.86 for ASD vs. TD, and 0.88 for the combined ASD+DLD group vs. TD. Performance was moderate for ASD vs. DLD, with an F1-score of 0.63. The most informative areas of interest were those broadly grouping social and non-social stimuli, while more specific variables did not improve classification accuracy. Naive Bayes and Logistic Model Trees (LMT) emerged as the most effective algorithms in this study. The resulting model enabled the identification of potential disorder-specific markers, such as the mean duration of visits to objects. Conclusion: These findings highlight the potential of applying XML techniques to eye-tracking data collected through tasks designed to capture features characteristic of neurodevelopmental conditions. They also underscore the clinical relevance of such approaches for identifying the variables and parameters that differentiate between disorders.