• No results found

Chapter 5 Experiment I: Pilot study: Testing a potential of combining Functional Near-

2. Methods

3.1. fNIRS data analysis

Signal-To-Noise (SNR) estimation

Data was first visually inspected in order to identify noisy channels and trials during a calibration process. Due to the motion artifacts throughout the whole time series, four participants were excluded from the data analysis. Another two were removed due to a software crash. The remaining data from 5 participants was assessed for a period of time containing motion artifacts prior to the experimental session which was manually removed. The signal quality was calculated in Octave in order to assess the signal interference between NIRSport and Vicon tracking system. The procedure was performed with both tracking system ON and OFF for each participant. The Quality Scale tool in NIRStar was used to analyse signal quality after the calibration procedure for each participant. The quality scale provides information about a variety of parameters that affect the signal quality such as gains, signal level, noise level and hemodynamic index. These parameters are calculated based using standard deviations of the extrapolated HbO signal due to a higher amplitude. The quality scale presents the quality of the signal using a colour bar (Figure 13). The example signal quality test is shown in Figure 14.

88

Figure 13The signal quality scale indicator (taken from NIRStar Manual)

Figure 14 Example fNIRS signal quality test in Octave with Vicon tracking on (left) and off (right).

The next step involved calculating the coefficient of variation (CV) offline in order to quantify the signal-to-noise for the raw time series for each participant and each channel for both NIRS wavelengths (Schmitz et al., 2005). CV is mathematically defined as 100 times the standard deviation divided by the mean value, where the standard deviation and mean are computed from all the raw-data values in the measurement time series. NIRSLab software provides a Raw Data Checking feature (Figure 15) which allows identifying the noisy channels. fNIRS channels were removed when respective channels exceeded a variation coefficient of 15% (Piper et al., 2014).

89

90

Figure 16 Example of removed noisy channels (sudden spikes in time series) contaminated by motion artifacts

Preprocessing

The remaining fNIRS data were preprocessed and analysed using NIRSLab (NIRSLab version 2014 NIRX Medical Technologies). Optical densities were converted to the average haemoglobin concentration changes using the modified Beer-Lambert law (Cope & Delpy, 1988) for each channel and each subject. Oxy- (HbO) and deoxy – (HbR) and total (HbT) haemoglobin time series were band-pass filtered with low cutoff frequency 0.01 Hz and high cutoff frequency 0.2 Hz to remove drifts, respiratory and cardiac signals from a raw NIRS data. A differential pathlength factor of 7.25 for 760nm and 6.38 for 850nm was applied (Essenpreis et al., 1993). Molar extinction coefficients ε for HbO at 760 nm = 1486.5865 cm−1/M and 850 nm = 2526.391 cm−1/M, were applied from W. B.Gratzer, Med.Res. Council Labs, Holly Hill, London and N. Kollias, Wellman Laboratories, Harvard Medical School, Boston. Raw data was converted to the average haemoglobin concentration changes using the modified Beer-Lambert law for each channel, each trial, each block and each subject

GLM

Statistical data analysis was performed using NIRS-SPM (SPM 8) analysis tool implemented in NIRSLab in order to identify regions of the brain activated during the virtual height exposure. Data were modelled with GLM.

91

1st level analysis

The first level design matrix containing two regressors modelling two conditions (training and pit) was generated by convolving with the canonical hemodynamic response function provided by SPM8 (Friston et al., 1996). Discrete cosine transform basis function was used for temporal filtering and precoloring HRF was used for the serial correlations. t-contrasts were then created for HbO concentration changes to generate statistical parametric maps of activation for two regressors: training and pit, for each channel, each condition and each subject. SPM t-maps were generated by using two contrasts: training versus pit and pit versus training and thresholded at p < 0.05 (corrected). To control the family-wise error rate, NIRS-SPM implements an algorithm for Sun's tube formula and Lipschitz-Killing curvature based expected Euler characteristics for p-value correction (Tak, 2009; Ye, Tak, Jang, Jung, & Jang, 2009).

2nd level analysis

At the group analysis SPM group HbO t–statistics were calculated to identify the channel significantly activated by exposure to virtual heights with the significance level threshold set at p < 0.5 (corrected) according to the false discovery rate method FDR used in fMRI studies (Singh & Dan, 2006). The estimated anatomical location of each channel was determined using anatomical locations of international 10-10 system cortical projections of EEG sensors (Koessler et al., 2009; Okamoto & Dan, 2005).

ROI analysis

At the 3rd level analysis, a Region Of Interest analysis (ROI) was performed, which is a common practice in neuroimaging studies after channel/voxel-wise analysis to look further into specific regions of interest in the brain (Poldrack, 2007). Because there was no significant difference in total HbO between blocks, ROIs were averaged across all the blocks to investigate which brain area was mostly activated during the task and whether the activity was correlated with other explanatory variables. Using the probabilistic assessment of cortical projection sensors underlying 10-20 system anatomical surface locations (Koessler et al., 2009; Okamoto et al., 2004). Four ROIs - left DLPFC, right DLPFC, left MPFC and right MPFC were defined on the basis of BA atlas (Brodmann, 1909) to be represented by channels: 1, 3, 5 and 7 (L DLPFC); 14, 15, 18 and 20 (R DLPFC); 6 and 11 (L MPFC); 13, and 16 (R MPFC) respectively. The HbO beta-

92

estimates from those channels were extracted for each subject and each condition and then averaged within each of the ROIs across the selected channels and blocks (Poldrack, 2007). Next, a student t-test was performed on each ROI separately. This approach was taken because putting ROIs as a factor in ANOVA analysis could cause a bias in the statistical analysis due to the different optical properties in different ROIs (Kroczek et al., 2015; Yanagisawa et al., 2010).