Mojan Izadkhah¹ (), Cathy S Zhong¹, Jason J S Barton¹, Ipek Oruc¹; ¹University of British Columbia

We investigated spatiotemporal patterns of neural processing underlying saccadic planning using explainable AI techniques. We recorded scalp EEG using a 64-channel BioSemi setup while participants (N = 20) completed pro-saccades to the left or right, randomly selected on each trial. A 3D adaptation of EEGNet trained with a novel data augmentation technique we developed to address the relatively small number of EEG trials (mean: 436.45 per participant, range: 273-492), successfully predicted saccade direction prior to onset with performance significantly above chance (mean AUC=0.77, p<.001, 95% CI: [0.67, 0.88]; mean accuracy=0.70, p<.001, 95% CI: [0.61, 0.79]) based on the held-out validation data. Model performance correlated strongly with dataset size for both AUC and accuracy (r=0.68, p=0.001), suggesting that sufficiency of training data was the primary limiting factor for model generalization performance. We applied a modified GradCAM algorithm to identify spatial and temporal features informing CNN predictions. Right frontal electrodes (FP2, AF4, F2, FC2, C2, CP2) were critical for predicting left saccades, while left electrodes (F1, FC1, C1, CP1) were important for right saccades. Of note, right frontal electrodes (FP2, AF4) remained critical for both directions, possibly reflecting a right-lateralized 'cognitive motor planning' signal. Temporal analysis of this signal revealed harmonic oscillations, peaking at 30Hz for both left and right saccades, consistent with low gamma band activity with a ~10ms lag for left saccades compared to right. A lateralized 'motor preparation signal' peaked during the final 32-16ms for right saccades at the left frontal electrodes. This motor preparation signal was more broadly distributed temporally for left saccades, appearing 36-24ms and 72-56ms prior to the onset of the saccade at the right frontal electrodes. We conclude that our explainable AI analysis can identify saccadic preparation signals and reveals significant asymmetries in the generation of left versus right saccades that require exploring.

Acknowledgements: This work was supported by an NSERC Discovery Grant RGPIN-2019-05554 (IO) and an NSERC Accelerator Supplement RGPAS-2019-00026 (IO). JB was supported by Canada Research Chair 950-232752. CZ was supported by the Neurological Foundation of New Zealand Grant 2246 CHF.