VEOS: Vision-based vertical electrooculography inference from monocular periocular video for ocular artefact suppression in EEG

OBJECTIVE: Ocular artefacts from blinks and eye movements remain a persistent obstacle
to reliable electroencephalography (EEG), particularly when dedicated electrooculography
(EOG) electrodes are unavailable or undesirable. We investigate whether monocular
periocular video can provide a vertical EOG (VEOG)-like surrogate for ocular artefact
suppression in EEG.
Approach. We present VEOS, a video-to-VEOG inference pipeline based on monocular
periocular tracking, canthus-defined geometric normalisation, engineered eyelid and iris
features, and temporal modelling with a temporal convolutional network (TCN). The
inferred VEOG is used as an auxiliary reference in blink-window lagged ridge-regression
subtraction of ocular artefacts from EEG. Evaluation used leave-one-subject-out
validation on two multimodal datasets: an in-house development dataset (VEOS-Dev; 5
participants with synchronised EEG, EOG and 120 Hz video) and the public Eye-BCI
dataset (31 subjects, 63 sessions; high-speed video, EEG and ocular measurements). For
centred-window models, a ±ℓ boundary margin prevented temporal leakage between
training, validation and test data.
Main results. On VEOS-Dev, VEOS achieves median Pearson correlation r = 0.81 to
ground-truth VEOG and median blink-onset timing error of 18 ms. On Eye-BCI, the
video-only model achieves median r = 0.74, outperforming an eye-tracker baseline. For
blink-window cleaning, VEOS suppresses approximately 50% of blink peak-to-peak
amplitude on frontal EEG channels, compared with 66% for true VEOG. Cleaning
preserves task-relevant EEG structure, including motor imagery mu/beta rhythms,
SSVEP spectral signal-to-noise ratio, and P300 morphology. At the participant level,
VEOS yields a modest improvement in motor imagery, leaves SSVEP unchanged, and
approaches true-VEOG cleaning for P300 spelling (90% vs 91%).
Significance. These results show that camera-derived VEOG can act as a useful ocular
reference for EEG artefact suppression without additional periocular electrodes. The
validation was conducted on controlled recordings and should be interpreted as a best-case
demonstration rather than a full validation in unconstrained wearable settings.