Abstract:Background: The assessment of eye movements is commonly considered a “window into the brain”. However, sole clinical observation is limited to semiquantitative findings. Recent advances in quantitative oculomotor phenotyping using video-oculography (VOG) promise additional diagnostic and prognostic value. However, methodological complexity and limited availability have hampered broad use of VOG in clinical and outpatient settings. Here, a novel approach for binocular video-nystagmography (VNG), capitalizing on recently deployed artificial intelligence frameworks for markerless pose estimation, is validated. Methods: Standardized 2D-optokinetic stimuli were presented on a smartphone screen and elicited nystagmus was recorded with two methods: first, using a conventional smartphone camera (1920x1080 px, 30 frames per second) mounted frontoparallely on a tripod and second, using gold standard infrared VOG goggles (EyeSeeCam, 188x120 px, 220 frames per second). Using the open-source software DeepLabCut, a recurrent neural network (RCNN) architecture was fine-tuned for tracking of pupils and facial anatomical landmarks. For subsequent comparisons, kinematic parameters (e. g. slow phase velocity (SPV)) of nystagmus were calculated from RCNN- and VOG-derived time series datasets using an instantaneous gradient algorithm. Precision of methods was compared using ANOVA and modified T-statistics, namely two one-sided T-testing (TOST) and one-sample T-tests. Results: In monocular analyses of RCNN-derived data, neither eye nor stimulus direction exerted significant influence on SPV (F (1, 5) = 1.33, p= [DZ1].30 and F (3, 15) = 0.66, p = .59). Furthermore, method did not significantly influence omnidirectional SPV measurements (F (1, 5) = 0.009, p = .78). Finally, TOST results demonstrated equivalence in both horizontal (upper bound T(5) = -3.45, p = .009; lower bound (T(5) = 5.0, p= .002) and vertical (T(5) = -3.06, p = .01; T(5) = 3.35, p = .01) stimulus planes. Conclusions: This proof-of-concept study shows that a deep learning based tool (DeepVNG) can extract characteristics of nystagmus from smartphone video recordings of eye movements elicited by standardized optokinetic stimuli with comparable precision to gold standard VOG. These results may lay the groundwork for broadly available “vestibular event monitors” assisting both caregivers and patients in detection and quantification of nystagmus and may open an avenue for more granular analyses of nystagmus expected to improve its diagnostic value.

Deep Learning-Based Nystagmus Detection for BPPV Diagnosis

Developing a Diagnostic Decision Support System for Benign Paroxysmal Positional Vertigo Using a Deep-Learning Model

Multimodal deep learning-based diagnostic model for BPPV

A Deep-Learning Model for 3d Nystagmus Detection and its Primary Application

Torsional nystagmus recognition based on deep learning for vertigo diagnosis

Diagnosing the benign paroxysmal positional vertigo via 1D and deep-learning composite model

A Deep Learning Model for Three-Dimensional Nystagmus Detection and Its Preliminary Application

Deep learning based torsional nystagmus detection for dizziness and vertigo diagnosis

A nystagmus extraction system using artificial intelligence for video-nystagmography

Nystagmus patterns classification framework based on deep learning and optical flow

aEYE: A deep learning system for video nystagmus detection

LAD: A Hybrid Deep Learning System for Benign Paroxysmal Positional Vertigo Disorders Diagnostic

Vertical Nystagmus Recognition Based on Deep Learning

1D Convolutional Neural Networks for Detecting Nystagmus

Gram-based Attentive Neural Ordinary Differential Equations Network for Video Nystagmography Classification

Feasibility of video-based real-time nystagmus tracking: a lightweight deep learning model approach using ocular object segmentation

P 56 Deep learning derived quantitative Video-NystagmoGraphy using smartphone cameras: DeepVNG

Deep Learning Detection of Subtle Torsional Eye Movements: Preliminary Results

Automated nystagmus detection: Accuracy of slow-phase and quick-phase algorithms to determine the presence of nystagmus

Nystagmus Signal Feature Extraction and Tracking for Diagnosis of the Vestibular System