Explaining Autism Detection by Deep Learning Through Eye Gaze Patterns and Integrated Gradients

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental con-dition marked by repetitive behaviors and challenges in communication and social interaction, with about 1 in 100 children affected. Diagnosis of ASD is challenging due to its reliance on subjective behavioral evaluations, but advanced technologies such as Deep Learning are being explored to improve early and accurate diagnoses. Among them, one promising application in ASD research is analyzing eye gaze patterns, as individuals with ASD often exhibit atypical behaviors, such as reduced eye contact.This paper introduces a bimodal neural network combining Convolu-tional Neural Networks and Long Short-Term Memory networks for high-accuracy ASD classification through eye gaze patterns, and an explainability component is also included to analyze the model’s decision-making process. The proposed neural network architecture achieves an average accuracy of 96.6% 96.6%, outperforming exist-ing benchmarks, furthermore, the integration of the Integrated Gradients method enhances the model’s clinical interpretability by identifying key visual features and fixation patterns. This explainability component aligns with well-documented ASD traits, such as reduced direct gaze and atypical fixation preferences: these findings underscore the potential of our approach to provide an objective and interpretable tool for ASD detection, offering valuable insights into attentional mechanisms in neurodivergent people.