Deep Learning Model to Evaluate Sensorimotor System Ability in Patients With Dizziness for Postural Control.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Balanced posture relies on integrating visual, vestibular, and somatosensory inputs.
• Center of pressure (COP) signal frequencies correlate with sensorimotor system function.
• Current balance assessments can be time-consuming and require specialized equipment.

Purpose of the Study:

• To develop a deep learning protocol for estimating equilibrium scores (ES) using COP signal frequencies.
• To quantitatively assess sensory system contributions in individuals with dizziness.
• To explore a more efficient and accessible method for balance assessment.

Main Methods:

• Utilized COP signals from quiet standing in normal controls and patients with Meniere's disease or vestibular neuritis.
• Applied Short-time Fourier Transform to convert COP signals to the frequency domain.
• Trained and tested Convolutional Neural Network (CNN) models (GoogleNet, ResNet-18, SqueezeNet, VGG16) with encoded ES sub-classes.

Main Results:

• Achieved an absolute difference of approximately 1.7 between measured and predicted ES using ResNet-18 with 20 sub-classes.
• Demonstrated an average error of ~1.0% in sensory analysis using predicted ES.
• Validated the potential of COP signal frequencies for quantitative balance assessment.

Conclusions:

• Deep learning analysis of COP signal frequencies can effectively estimate sensory system contributions to balance.
• This approach offers a promising, time-efficient alternative to traditional balance testing methods.
• The findings support the widespread application of simplified balance assessments for dizziness evaluation.