利用深度学习可解释性图谱揭示与晕动症不适持续相关的 EEG 模式
Uncovering EEG Patterns Consistently Associated with Cybersickness Discomfort Using Deep Learning Interpretability Maps
研究团队用神经网络与可解释性图谱构建框架，从两项听觉 ERP 晕动症用户研究（n=29 与 n=36）的 EEG 数据中提取对不适分类最关键的特征。在第一个数据集上跨三种网络、多个随机种子共 120 次运行中，模型一致指向左额、中线中央和右顶叶头皮位置，第二数据集标记出相似位置，显示跨数据集泛化性。两个数据集均将刺激后约 80 至 260 毫秒的早期 ERP 时段标记为分类重要区间。
Authors:Jacqueline Yau, Evan G. Center, Pawel Augustynowicz, Steven M. LaValle, Timo Ojala, Wenzhen Yuan, Nancy M. Amato, Pawel Strozak, Minje Kim, Kara D. Federmeier, Katherine J. Mimnaugh
Abstract:Uncomfortable sensations similar to motion sickness, called cybersickness, can develop when using Virtual Reality (VR) head-mounted displays. Cybersickness poses a hindrance to greater use of VR technology. Brain activity recorded using electroencephalogram (EEG) can be used to unintrusively detect cybersickness and the discomfort it causes. To intervene for mitigation, machine learning algorithms that can extract meaningful signals related to cybersickness discomfort from the rest of the brain data will be required. In this work, we determined which features in EEG data were most helpful for classifying cybersickness-related discomfort by building a framework with neural networks and interpretability maps. Using brain data from two separate auditory event-related potential (ERP) cybersickness user studies, we extracted which spatio-temporal EEG features (from sensor locations and time steps) were most important for discomfort classification. For the first dataset (n=29), across 120 runs of our framework with three different neural networks over multiple random seeds, the models consistently pointed to scalp locations in left frontal, midline central, and right parietal areas as most helpful in determining if EEG data belonged to someone who was experiencing cybersickness discomfort. Similar locations were also tagged in a second dataset (n=36), showing cross-dataset generalizability of our findings across time, VR stimulus, and participant sample. Early periods in the extracted ERP time windows, approximately between 80 and 260 milliseconds post-stimulus, were marked as important for model classification in both datasets. These results help clarify which tagged features can be used for cybersickness-related (and potentially other types of) discomfort classification with EEG in the future. Project webpage at: this https URL.
| Comments: | 30 pages, 6 figures. 19 pages, 18 figures in appendix |
| Subjects: | Human-Computer Interaction (cs.HC); Machine Learning (cs.LG) |
| Cite as: | arXiv:2512.20620 [cs.HC] |
| (or arXiv:2512.20620v3 [cs.HC] for this version) | |
| https://doi.org/10.48550/arXiv.2512.20620 arXiv-issued DOI via DataCite |
Submission history
From: Jacqueline Yau [view email]
[v1]
Mon, 3 Nov 2025 04:27:58 UTC (2,379 KB)
[v2]
Fri, 27 Mar 2026 04:24:50 UTC (10,968 KB)
[v3]
Thu, 1 Oct 2026 21:04:58 UTC (14,661 KB)
来源：arXiv：cs.LG · arxiv.org