A Deep Learning Approach to Imputation of Dynamic Pupil Size Data and Prediction of ADHD

Abstract

Attention-deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental disorder in children and adolescents. Traditional diagnosis methods of ADHD focus on observed behavior and reported symptoms, which may lead to a misdiagnosis. Studies have focused on computer-aided systems to improve the objectivity and accuracy of ADHD diagnosis by utilizing psychophysiological data measured from devices such as EEG and MRI. Despite their performance, their low accessibility has prevented their widespread adoption. We propose a novel ADHD prediction method based on the pupil size dynamics measured using eye tracking. Such data typically contain missing values owing to anomalies including blinking or outliers, which negatively impact the classification. We therefore applied an end-to-end deep learning model designed to impute the dynamic pupil size data and predict ADHD simultaneously. We used the recorded dataset of an experiment involving 28 children with ADHD and 22 children as a control group. Each subject conducted an eight-second visuospatial working memory task 160 times. We treated each trial as an independent data sample. The proposed model effectively imputes missing values and outperforms other models in predicting ADHD (AUC of 0.863). Thus, given its high accessibility and low cost, the proposed approach is promising for objective ADHD diagnosis.