Detecting positional vertigo using an ensemble of 2D convolutional neural networks.

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Positional vertigo, specifically Benign Paroxysmal Positional Vertigo (BPPV), is a common vestibular disorder characterized by brief, intense dizziness episodes triggered by specific head movements.
• Accurate and objective identification of these vertigo attacks is crucial for diagnosis and effective management.
• Current diagnostic methods often rely on subjective patient reporting and clinical maneuvers, which can be challenging to capture during an actual event.

Purpose of the Study:

• To develop and validate an automated system for detecting dizziness attacks in patients with BPPV.
• To leverage a novel medical device (CAVA) for continuous monitoring of eye and head movements.
• To improve the accuracy and efficiency of BPPV attack identification.

Main Methods:

• Utilized a novel medical device, CAVA, to continuously record eye- and head-movement data for up to 30 days in BPPV patients.
• Developed a novel ensemble of five 2D Convolutional Neural Networks (CNNs).
• Employed composite recognition features, integrating eye-movement data and three-channel accelerometer data.

Main Results:

• Achieved an F1 score of 0.63 in an 11-fold cross-validation experiment.
• Demonstrated the system's ability to detect motion-provoked dizziness episodes within extensive periods of normal movement data.
• The ensemble classifier outperformed individual networks and previous 1D Neural Network approaches.

Conclusions:

• The developed automated system shows significant promise for objective detection of dizziness attacks in BPPV patients.
• The use of composite recognition features combining eye and head movement data enhances detection performance.
• This AI-driven approach offers a potential advancement in the diagnosis and monitoring of vestibular disorders.