Towards defining biomarkers to evaluate concussions using virtual reality and a moving platform (BioVRSea)

Figures

• 

  Figure 1. BioVRSea experimental Setup. BioVRSea is a novel VR-motion platform coupled system designed to mimic the sensation of being at sea in a small boat. Participants in this study stood on the platform while wearing VR goggles, EEG, and EMG measurement equipment while simultaneously having CoP measurements taken by two force platforms under each foot. The platform motion was coupled to the sea scene, as observed through the VR goggles. The protocol from the experiment included a baseline segment where the platform was not moving, and the participants stood with their hands by their side and viewed a VR scene of mountains. After two minutes the VR scene changed to that of a small boat at sea, the platform does not move for the first 40 s of the VR sea scene. The participant remains standing quietly with their hands by their side. This is the PRE part of the experiment. After 40 s, the participants are told to hold onto the safety bars for the following part of the experiment, where the platforms moves synchronously with the waves in the VR scene. There are three segments where the platform moves for 40 s at 25%, 50% and 75% of the maximum amplitude. The platform then stops moving and…

• 

  Figure 2. Symptoms (SCAT5) reported by all participants before and after VR acquisition. Overview of concussion symptoms reported by all participants before VR acquisition.

• 

  Figure 3. The most statistically significant features computed for the CoP parameters. All groups (Concussion (CONC—blue), Non-concussion (NON CONC—grey) and Balance Problems (BP—orange) showed a statistically significant change from the PRE to the POST stage for TOTEX (Total Excursion), MDIST_AP (Mean distance from centre in antero-posterior direction), MVELO (Mean velocity), ELLIPSE_MAIN_AXIS—The length of the main axis of 95% confidence ellipse, SD_AP (Standard Deviation in antero-posterior direction), POSTERO_MAGNITUDE (Posterior extreme point distance). One spectral feature PSD_EXP_AP—Power Spectral Density (PSD) Power Law is also shown.

• 

  Figure 4. A topological plot for delta and theta frequency band of the concussion group. It displays the difference of power spectral density, between POST and PRE, only for the statistically significant electrodes (0.032 ≤ p ≤ 0.047, represented by a blue star in the figure). The delta band presented a few significant electrodes (6 out of 64) on the frontal and occipital cortex, whereas the theta band showed more distributed significant electrodes (36 out of 64) over the scalp.

• 

  Figure 5. A topological plot of theta frequency band for three concussion symptoms subgroup. (A) Difficulty concentrating, (B) Pressure in the head, (C) Fatigue and low energy). It displays the difference of power spectral density, between POST and PRE, only for the statistically significant electrodes (0.008 ≤ p ≤ 0.049, represented by a blue star in the figure). Each subgroup presented significant electrodes on the frontal lobe (16 significant electrodes for difficulty concentrating, 15 for pressure in the head, and 14 for fatigue and low energy) for all of them, as well as an increase in the occipital lobe for the Pressure in the head (8 significant electrodes) and Fatigue and low energy (10 significant electrodes) subgroup.

• 

  Figure 6. The POST–PRE results for median frequency in the (a) right and (b) left soleus muscles for the concussion (CONC—blue), concussion with balance problems (CONC BP—orange) and non-concussion groups (NON CONC—grey). Figures (c) and (d) show the POST–PRE results for median frequency in the left and right soleus muscles for the concussion balance problems vs concussion no balance problems group. When these groups were compared in the POST phase the right soleus showed a significant discrimination between the BP and No BP groups.

• 

  Figure 7. The results of 5 different algorithms (RF- Random Forest, SVM—Support Vector Machine, ADA-B—Adaptive Boosting , MLP- Multilayer Perceptron, GB-Gradient Boosted) and combinations of features. Highest accuracy in all algorithms is achieved using both the PCA and SCAT5 features, but highest accuracy overall (95.5%) when computed with an SVM algorithm.

• 

  Figure 8. This figure shows the position of the feet as they are placed on the force platform during the experiment. The anterior–posterior displacement is calculated using the formula: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{AP - Displacement}}\left( {{\text{cm}}} \right) = Y\left( {{\text{cm}}} \right) * {\text{StaticVR}}\_{\text{AP/}}2$$\end{document}AP - Displacementcm=Ycm∗StaticVR_AP/2. Where StaticVR AP is the percentage of the vertical size of the board between −1 and 1. The medio-lateral displacement is calculated using the formula: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{ML\_Displacement }}\left( {{\text{cm}}} \right) \, = \, X\left( {{\text{cm}}} \right)*{\text{StaticVR}}\_{\text{ML}}/2$$\end{document}ML\_Displacementcm=Xcm∗StaticVR_ML/2 Where StaticVR ML is the percentage of the horizontal…

• 

  Figure 9. A graphical description of the features used in the CoP analysis. A number of geometrical and spectral features were calculated for the PRE and POST stages of the experiment. The figure shows an example of a stabilogram on the left and how some features are extracted. Similarly, an example calculation of entropy features is shown on the right to give some indication of what was calculated for the entropy measures using the stabilogram data.

• 

  Figure 10. Feature Selection and Machine Learning Workflow. 51 features are selected from Brain, Muscles and Heart signal and on them, the PCA is performed to obtain 9 PCA features used to the binary classification of the concussion assessment using different ML algorithms; 22 SCAT5 features are then used as well as features for the concussion classification; the 22 SCAT5 and 9 PCA are then combined and used together to classify concussed and non-concussed subjects.