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

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Vertigo is a common symptom often associated with nystagmus, an involuntary eye movement.
• Accurate diagnosis of vertigo subtypes, like benign paroxysmal positional vertigo (BPPV), is crucial for effective treatment.

Purpose of the Study:

• To design and validate a three-dimensional nystagmus recognition model.
• To develop an automated system for diagnosing benign paroxysmal positional vertigo using deep neural networks.

Main Methods:

• An object detection model was employed to track pupil center movement.
• Convolutional neural network-based models were trained for 3D nystagmus pattern detection.
• A deep neural network architecture was utilized for the BPPV diagnosis system.

Main Results:

• Nystagmus detection models achieved high areas under the curve: 0.982 (horizontal), 0.893 (vertical), and 0.957 (torsional).
• The automated BPPV diagnosis system demonstrated a sensitivity of 0.8848, specificity of 0.8841, accuracy of 0.8845, and an F1 score of 0.8914.
• The system provides a clinical reference and facilitates nystagmus detection and diagnosis.

Conclusions:

• The developed deep neural network system effectively recognizes nystagmus in three dimensions.
• The automated BPPV diagnosis system shows promising performance for clinical application.
• This technology can aid healthcare professionals in diagnosing vertigo and related conditions more efficiently.