Smartphone video nystagmography using convolutional neural networks: ConVNG.

Area of Science:

• Ophthalmology and Neurology
• Biomedical Engineering
• Computer Vision

Background:

• Eye movement abnormalities are common in neurological disorders, but current assessment methods lack detail.
• Videooculography (VOG) enhances diagnostic accuracy but is resource-intensive, limiting widespread use.
• A novel framework for smartphone video-based nystagmography is validated to address this care gap.

Purpose of the Study:

• To validate a smartphone video-based nystagmography framework (ConVNG) using computer vision.
• To assess the accuracy and precision of ConVNG for calculating slow-phase velocity (SPV) of optokinetic nystagmus.
• To compare ConVNG performance against established computer vision benchmarks (OpenFace, MediaPipe) and VOG.

Main Methods:

• A convolutional neural network (ConVNG) was fine-tuned for pupil tracking using over 550 annotated frames.
• Slow-phase velocity (SPV) was calculated in 10 subjects using ConVNG and VOG.
• Equivalence testing (TOST, Bayesian interval-null) and direct comparisons were performed against benchmark CV algorithms.

Main Results:

• ConVNG achieved high pupil tracking accuracy (9-15% of average pupil diameter) and robust keypoint detection (median confidence 0.85).
• SPV measurement accuracy with ConVNG was equivalent to VOG (TOST p < 0.017; BF > 24).
• ConVNG demonstrated significantly higher precision (0.30°/s) than MediaPipe (0.7°/s) and comparable accuracy to VOG (0.12°/s).

Conclusions:

• ConVNG enables accurate offline smartphone video nystagmography, comparable to VOG.
• ConVNG offers significantly higher precision than MediaPipe for gaze estimation.
• This framework provides a blueprint for accessible tools to advance precise and personalized medicine in neurology.