Advancing Otolith Function Assessment: Integrating Artificial Intelligence with Video-Oculography (VOG) for Enhanced Vestibular Diagnosis
Krishna Mukunda1, Amir Kheradmand2, Kemar Green3
1Departments of Biomedical Engineering and Neurology, Johns Hopkins University, 2Department of Neurology, Johns Hopkins Medicine, 3Department of Neurology, John's Hopkins Medicine
Objective:
 To determine the accuracy with which machine and deep learning models can differentiate between normal and abnormal video Ocular Counter Roll (vOCR) videos.
Background:

Dizziness affects around 7.4% of adults, presenting diagnostic challenges, particularly in early differentiation of benign (inner ear disease) vs dangerous (stroke) causes. The ability to detect subtle changes in the otolith can provide clues regarding central and peripheral vestibular function. The degree and direction of dynamic ocular torsion obtained using the beside vOCR test is a quantitative measure of vestibular function. Recent advances in deep learning-based dynamic torsional detection may make it possible to automate the vOCR, increasing accessibility to non-expert providers on the frontline.

Design/Methods:

36 participants were enrolled, including 18 healthy controls and 18 with vestibular loss. Each participant underwent the vOCR test, involving 24 tilts consisting of both neck and trunk movements to stimulate otolith function. Torsional waveforms were recorded, and beats of torsion, each consisting of a fast and slow phase, were extracted. Clips of 500ms, containing at least one full beat of torsion, were isolated. Several Machine Learning Models (MLMs) were trained on the extracted waveform features. For image-based analysis, a filtered image was generated by summing the differences between consecutive video frames, creating a visualization that captured the dynamic motion in each clip. A simplified 2D ResNet18 model was then trained on these filtered images.

Results:
The simplified 2D ResNet18 image-based model outperformed the waveform-based MLMs, achieving an AUC of 81.07% on the validation set compared to the top waveform model's AUC of 66.00%.
Conclusions:
This study highlights the effectiveness of artificial intelligence in capturing and classifying subtle torsional movements associated with vestibular function. By leveraging a ResNet18 model, we demonstrate the potential to automate vOCR assessments, facilitating wider access to accurate vestibular diagnostics in clinical settings where specialist expertise may be limited.
10.1212/WNL.0000000000212709
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.