Open Collections

Abstract

We investigated the spatiotemporal patterns of neural activity underlying prosaccade and antisaccade planning using deep learning combined with explainable AI techniques applied to scalp electroencephalography (EEG). EEG was recorded from 20 participants using a 64-channel BioSemi system while they completed one block of randomized left- or rightward prosaccades, followed by a block of randomized left- or rightward antisaccades. Saccade onset timing was extracted using electrooculography (EOG). For prosaccades, the number of accepted trials per participant was M = 436.45 (SD = 61.21), range = 273–492; for antisaccades, M = 436.15 (SD = 75.25), range = 194–519. Mean reaction times were 172.51 ms (SD = 21.37) for prosaccades and 229.50 ms (SD = 18.19) for antisaccades. A 3D convolutional neural network was trained on approximately 100 ms of pre-saccadic EEG data. To address the relatively limited number of trials per participant, we developed a novel data augmentation pipeline. Classifiers were trained separately for each participant, and performance was evaluated on held-out data from the same individual. For prosaccade (rightward vs. leftward) classification, mean accuracy across participants was M = .84 (SD = .14, n = 20); for antisaccade (leftward vs. rightward) classification, M = .85 (SD = .14, n = 19). Both results were substantially and statistically significantly above chance, p < .001. To interpret the model’s predictions, we applied a saliency-based post-hoc interpretability method, which revealed contralateral saliency patterns in both prosaccade and antisaccade tasks, as well as an additional ipsilateral frontal contribution unique to antisaccades within the analyzed window preceding saccade onset. Overall, our findings demonstrate that explainable deep learning can predict saccade direction from brief EEG windows and uncover interpretable neural dynamics associated with motor planning and inhibition.