Objective: We aimed to classify nystagmus videos into central and peripheral nystagmus using a deep learning model.

Background: Diverse causes underlie vertigo, and while clinically excluding central nervous system disorders is critical, diagnosing central vertigo is often difficult, requiring specialized knowledge and experience. Nystagmus detection using deep learning models has recently been reported[1], but models for diagnosing nystagmus have not been sufficiently developed.

Method: We retrospectively collected data from patients who underwent video recording using an infrared CCD camera for nystagmus evaluation. Patients with uncertain diagnoses or unclear nystagmus were excluded. We applied the Segment Anything Model (SAM)[2] to perform pupil segmentation in nystagmus videos by analyzing pupil movement as a point motion. The captured point motion characteristics were input into the convolutional neural network model to predict whether the patient had central or peripheral nystagmus. The performance of the model in diagnosing nystagmus was calculated using Precision-Recall AUC (PR-AUC), F1 score, and accuracy.

Results: In total, 129 patients were included in analyses. Thirty-five patients had central nervous system disorders and 94 had peripheral vestibular disorders. Respective PR-AUC, F1, and accuracy scores were 0.48, 0.62, and 0.71.

Conclusion: Our deep learning model requires performance improvements to achieve clinical applicability.

References: 1. N. Wagle, J. Morkos, J. Y. Liu, H. Reith, J. Greenstein, K. R. Y. Gong, et al. aEYE: A deep learning system for video nystagmus detection. Frontiers in Neurology 2022 Vol. 13
2. A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, et al. Segment Anything. Arxiv 2023.

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/a-deep-learning-model-based-screening-system-for-nystagmus-in-central-vertigo/