Residual Neural Network Architecture for Identifying Vestibular Disease Based on Head Kinematic Characteristics (Velocity)

In this paper, we assess human navigation in a virtual reality neuropsychological test named the VR magic carpet; the task involves repeating sequences shown by the test administrator via a VR environment built in Unity. The objective of this assessment is to identify patients with vestibular dysfunction. In earlier work, we used unsupervised classical algorithms on an aggregated participant analysis based on kinematic characteristics, specifically head velocity global average. However, experts mentioned the necessity to analyze further and consider other parameters related to head velocity time series extracted during the clinical trials. Knowing that machine learning techniques are based on hand-crafted features may lead to forgetting essential features. The intuitive solution is embedding deep learning in the analysis process. For that reason, Experts launched clinical trials and labeled the extracted files containing raw 3D spatial data to approach the problem as a classification problem. Furthermore, we used residual neural network topologies, with layers used for feature extraction and classification, to establish an identification model for vestibular dysfunction. The experimental results of the model in this study demonstrate an improvement since they more accurately distinguish between healthy and patients.