Abstract
OBJECTIVE: To explore the potential of using artificial intelligence (AI)-based eye tracking technology on a tablet for screening Attention-deficit/hyperactivity disorder (ADHD) symptoms in children. METHODS: We recruited 112 children diagnosed with ADHD (ADHD group; mean age: 9.40 ± 1.70 years old) and 325 typically developing children (TD group; mean age: 9.45 ± 1.59 years old). We designed a data-driven end-to-end convolutional neural network appearance-based model to predict eye gaze to permit eye-tracking under low resolution and sampling rates. The participants then completed the eye tracking task on a tablet, which consisted of a simple fixation task as well as 14 prosaccade (looking toward target) and 14 antisaccade (looking away from target) trials, measuring attention and inhibition, respectively. RESULTS: Two-way MANOVA analyses demonstrated that diagnosis and age had significant effects on performance on the fixation task [diagnosis: F((2, 432)) = 8.231, (***)p < 0.001; Wilks' Λ = 0.963; age: F((2, 432)) = 3.999, *p < 0.019; Wilks' Λ = 0.982], prosaccade task [age: F((16, 418)) = 3.847, (***)p < 0.001; Wilks' Λ = 0.872], and antisaccade task [diagnosis: F((16, 418)) = 1.738, *p = 0.038; Wilks' Λ = 0.938; age: F((16, 418)) = 4.508, (***)p < 0.001; Wilks' Λ = 0.853]. Correlational analyses revealed that participants with higher SNAP-IV score were more likely to have shorter fixation duration and more fixation intervals (r = -0.160, 95% CI [0.250, 0.067], (***)p < 0.001), poorer scores on adjusted prosaccade accuracy, and poorer scores on antisaccade accuracy (Accuracy: r = -0.105, 95% CI [-0.197, -0.011], *p = 0.029; Adjusted accuracy: r = -0.108, 95% CI [-0.200, -0.015], *p = 0.024). CONCLUSION: Our AI-based eye tracking technology implemented on a tablet could reliably discriminate eye movements of the TD group and the ADHD group, providing a potential solution for ADHD screening outside of clinical settings.