Detecting ADHD from Eye Movements - Comparing Time-Series Imaging Methods

Abstract

Diagnosing mental illnesses is complex and often relies heavily on behavioral assessments, which can be subjective. The application of eye tracking for diagnostic purposes offers a promising alternative. However, complicated data processing techniques are required to analyze the time series data it produces. To address this challenge and effectively detect gaze patterns from individuals with Attention Deficit and Hyperactive Disorder (ADHD), we compare various time-series imaging algorithms. Using a vanilla Convolutional Neural Network (CNN), we rank these algorithms in their ability to encode the relevant information that distinguishes between the gaze patterns of ADHD individuals and Typically Developing (TD). This study utilizes the ’ecml-ADHD’ dataset by Deng et al. (2023) to develop and validate a diagnostic approach. Four imaging techniques, namely the Markov Transition Field (MTF), Gramian Angular Field (GADF), Gramian Angular Summation Field (GASF), and Recurrence Plot (RP), are used in this study to convert time series eye movement data to images. These images are then analyzed using a vanilla CNN to classify participants into two groups: ADHD and TD individuals. The proposed method achieved 89.9% with GADF, 92.3% with GASF, 91.2% with MTF, and 89.7% with RP imaging techniques as the highest accuracy when compared in combination with different eye movement variables. These results validate the efficiency of using eye movement data, processed through advanced imaging techniques and analyzed with CNN, for distinguishing between ADHD and TD subjects. Therefore, it can be said that the combination of eye tracking, imaging, and machine learning techniques can significantly assist in diagnosing ADHD. We also discuss how the methodology developed in the study has the potential to be adapted to assist in the diagnosis of other psychiatric disorders. Keywords: ADHD Detection, Eye Tracking Technology, Gaze Dynamics, ecml-ADHD Dataset, Convolutional Neural Network (CNN), Imaging Time Series