VR discomfort shows up in the EEG: left frontal, central and right parietal sites keep surfacing

EEG classification of cybersickness-related discomfort now points to which electrodes and time steps drive a model decision, rather than resting on band power alone. Most earlier machine learning work extracted alpha and beta band power from continuous EEG, while event-related potential data remained a much less studied route with poorer comparability across samples. Across 120 ERPVR runs spanning three neural networks, multiple random seeds and permutations of trial type, sampling, outlier detection and electrode ordering, models consistently tagged left frontal, central and right parietal scalp locations as most helpful.

The models concentrate on an early stretch of the auditory response rather than the full 800 millisecond window. Prior cybersickness detection typically used windows of several seconds up to 30 seconds, whereas each sample here covers 800 milliseconds after the stimulus. Important time windows in both datasets fell roughly 80 to 260 milliseconds post-stimulus, with EEGNet's GradCAM rising near 100 milliseconds and peaking near 200 milliseconds in ERPVR, and starting near 80 milliseconds with a peak near 110 milliseconds in VectionVR.

The same scalp locations reappear in a second, fully independent VR dataset, so the pattern is not tied to one experimental setup. ERPVR and VectionVR differ in collection time, country, headset, EEG system, electrode montage and VR task. In the 36-participant VectionVR dataset, C3, T7 and CPz were among the strongest contributors to high-discomfort classification, echoing the left central region tagged for the discomfort group in ERPVR.

Per-subject calibration sharply improves discomfort discrimination, while direct transfer to unseen participants remains much harder. Under leave-one-subject-out validation, ERPVR's best balanced accuracy was about 76.9 percent, and VectionVR's overall baseline sat near chance. After calibration, ViT reached about 95.6 percent balanced accuracy for ERPVR discomfort subjects and about 91.5 percent overall on VectionVR, with roughly 90 percent for high discomfort and 92.6 percent for low discomfort.

Results come from two auditory oddball VR EEG datasets: ERPVR with 29 participants on a 16-channel approximately equidistant montage, and VectionVR with 36 participants on a 30-channel 10-20 layout, both classifying the general discomfort item of the SSQ. Feature interpretation covers only participants whose classification accuracy exceeded 75 percent, so it describes the patterns those models relied on in those participants. The authors propose testing next whether these spatial patterns can support discomfort detection in continuous EEG without an external auditory probe.

Scalp electrodes record mixtures of activity from many neuronal populations, distorted by the skull and meninges, so the tagged locations cannot be read directly as specific brain sources. Labels in both datasets are coarse: ERPVR collected an SSQ roughly every two minutes, VectionVR only one label across eight blocks, and discomfort is scored on a 0 to 3 scale. Only 14 ERPVR participants felt discomfort, and high-accuracy features come from a smaller subset of them; both montages had few occipital electrodes, leaving occipital usefulness unknown.