Enhancing Autism Detection Through Gaze Analysis Using Eye Tracking Sensors and Data Attribution with Distillation in...

Area of Science:

• Neurodevelopmental Disorders
• Artificial Intelligence
• Computer Vision

Background:

• Autism Spectrum Disorder (ASD) is a neurodevelopmental condition with characteristic social communication differences and repetitive behaviors.
• Atypical visual attention patterns are frequently observed in individuals with ASD.
• Current diagnostic methods can benefit from enhanced accuracy and explainability.

Purpose of the Study:

• To propose the Gaze-Based Autism Classifier (GBAC), a Deep Neural Network model for improved ASD classification.
• To enhance ASD classification accuracy and model explainability using data distillation and attribution techniques.
• To investigate the correlation between gaze patterns and ASD-specific characteristics.

Main Methods:

• Utilized eye-tracking sensor data to capture gaze behaviors.
• Developed a Deep Neural Network model (GBAC) incorporating data distillation.
• Applied TracIn, a data attribution technique, to compute self-influence scores and filter training data.
• Evaluated model performance against full dataset training and benchmark models.

Main Results:

• Achieved a test accuracy of 94.35% for ASD classification.
• Demonstrated improved accuracy and computational efficiency by using only 77% of the dataset.
• GBAC outperformed models trained on the full dataset and random sample reductions.
• Data attribution analysis provided insights into influential training examples and ASD-related gaze patterns.

Conclusions:

• The proposed GBAC model effectively enhances ASD classification accuracy and explainability.
• Integrating explainable AI with eye-tracking data offers a promising approach for neurodevelopmental disorder diagnostics.
• The study highlights the potential of GBAC to advance clinical research by revealing insights into visual attention in ASD.