Assessing the Importance of Eye Position, Pupil Position, and Point of Regard as Relevant Biomarkers for Autism Spectrum Disorder Detection Using Deep Neural Networks (DNNs) and Long Short-Term Memory (LSTM) Models

More efficient and precise diagnoses, enabled by identifying important biomarkers, can significantly improve the standard of living of those with ASD, or Autism Spectrum Disorder. This study investigates the comparative performance of Deep Neural Networks (DNNs) and Long Short-Term Memory (LSTM) models by utilizing different hyperparameter configuration methods across ten trials to determine the relevance of three primary factors that contribute to autism detection: eye position, pupil position, and point of regard. On a dataset of 59 participants, with 547 time-window rows, both architectures achieved accuracy of 88%. However, through subsequent feature ablation experiments, it was observed that the accuracy of the models remained consistent despite the removal of the pupil position and point of regard parameters for both the left and right eye, whereas the accuracy of the model decreased when the eye position variable was removed from the code, ultimately postulating that solely the eye position of the subject is pertinent for effective autism detection.