Using explainable machine learning and eye-tracking for diagnosing autism spectrum and developmental language...

Area of Science:

• Neuroscience
• Computer Science
• Developmental Psychology

Background:

• Eye-tracking offers objective measures of social attention, crucial for assessing neurodevelopmental disorders like autism spectrum disorder (ASD).
• Current challenges include distinguishing between overlapping disorders and efficiently managing complex eye-tracking data for clinical use.

Purpose of the Study:

• To apply explainable machine learning (XML) algorithms to eye-tracking data from children with ASD, developmental language disorder (DLD), and typical development (TD).
• To assess the accuracy of group classification and identify key variables differentiating these groups.

Main Methods:

• Ninety-three children (ASD, DLD, TD) participated in a visual preference task with social and non-social stimuli.
• Eye-tracking data were analyzed using XML algorithms (Naive Bayes, Logistic Model Trees) across four datasets.

Main Results:

• High classification accuracy was achieved: 0.912 (DLD vs. TD), 0.86 (ASD vs. TD), and 0.88 (ASD+DLD vs. TD).
• Moderate accuracy (0.63) for ASD vs. DLD classification.
• Broad social/non-social stimulus areas were most informative; specific variables did not improve accuracy. Mean duration of object visits emerged as a potential disorder-specific marker.

Conclusions:

• XML techniques show significant potential for analyzing eye-tracking data in neurodevelopmental research.
• This approach can identify clinically relevant variables for differentiating between disorders with overlapping features.