Krishna Mukunda1, Tianyi Ye2, Yi Luo2, Asimina Zoitou2, Kyungmin (Esther) Kwon2, Richa Singh2, JiWon Woo2, Nikita Sivakumar3, Joseph L. Greenstein3, Casey Overby Taylor4, Amir Kheradmand5, Kemar Green5
1Department of Biomedical Engineering and Department of Neurology, 2Department of Biomedical Engineering, 3Institute for Computational Medicine, Department of Biomedical Engineering, 4Departments of Medicine and Biomedical Engineering, Institute for Computational Medicine, Johns Hopkins University, 5Department of Neurology, Johns Hopkins Medicine
Objective:
To determine the accuracy with which deep learning video-based models detect torsional nystagmus and explain their predictions.
Background:
Nystagmus is spontaneous repetitive ocular oscillations due to injury to vestibular and ocular motor neural circuitry. Of particular interest is torsional nystagmus – the result of damage to central and peripheral vertical-torsional eye muscle circuitry. The presence, degree or direction of torsional nystagmus can provide clues on central versus peripheral lesion localization, especially in the emergency setting. Timely diagnosis enables efficient interventions in the setting of central (stroke) causes of torsional nystagmus. Expert eye movement and vestibular specialists are scarce, and non-experts, like emergency room providers, may struggle with subtle eye movements. Current torsional detection relies on unreliable iris tracking due to eye closures and mobile device video quality.
Design/Methods:
To address this challenge, we utilized a small dataset of simulated torsional infrared video-oculography (VOG) to develop video-based deep learning models for remote torsional sensing without the need for traditional pupil or iris tracking. This dataset comprised torsion labels for clips of 500ms, 750ms, and 1s, with degrees of torsion ranging from 0.5-10.0 degrees. These models were trained using the ResNet architecture and incorporated 2D, 2.5D, and 3D convolutional layers. These models were compared to a small cohort of clinicians who classified 100 videos of the test set. We also applied explainable AI methods, including gradient-weighted class activation mapping (Grad-CAM), to enhance model interpretability.
Results:
Our models achieve 93.05% accuracy, using ocular features and neurophysiology-supported phenomena as evidenced by the Grad-CAM. Furthermore, when compared to the clinician, the model boasted a 90% overall accuracy compared to a 75% from the clinician.
Conclusions:
Our study illustrates that machine learning techniques can precisely identify small amplitude torsional nystagmus, offering a promising avenue for the early detection of pathological vestibular disorders that outperforms expert neuro-otologist.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.