The General Linear Model Dialog
The General Linear Model Dialog (GLM) contains important parameters needed for the GLM calculation. Please take a look at the GLM section of this guide to get more insights into the GLM calculations.
Adding confound predictors
There are several different confound predictors that can be added to the design matrix of the GLM. All of these confounds are added to one common design matrix, used for all channels in the later real-time analysis. To account for low frequency drifts, one can add linear confound predictor or high-pass confound predictors to the GLM. Just select the “Linear confound predictor” check box to add a linear trend predictor, or the “High-Pass confound predictor (sine + cosine) check box to add a sine and a cosine function of the respective frequency selected using the”Cutoff frequency" spin box. The cutoff frequency defines the frequency of the sine and cosine function with respect to the sampling rate of the current analysis.
Heartbeat confound predictor
It is also possible to add a confound predictor for the heartbeat, which is clearly visible in the fNIRS signal. The heartbeat predictor is using the resulting channels of a heartbeat detection algorithm to create a global heartbeat confound signal. The heartbeat detection algorithm consists of multiple steps applied for each channel individually. First, each channel’s signal is transformed into frequency space using a fast Fourier transformation (FFT). Only the last 256 data points are used for the FFT, resulting in a constant calculation time throughout the whole experiment. To calculate the heartbeat, the frequencies between 0.4 and 2.0 Hz are investigated using a peak detection. If the peak frequency is significantly different compared to the mean frequency in the range of 0.4 to 2.0 Hz using a P<0.001 (uncorrected), then the channel is added to the list of channels used for further heartbeat calculation. A little heartbeat icon is added to the channel selection table if a heartbeat is visible.
This display gives a direct indication of the overall signal quality. Ensure that you define a mask set of only the channels of interest to get a reasonable output incorporating only channels having a proper distance and that are expected to have a good signal.
The average oxy or deoxy signal of all channels containing a significant heartbeat is the source of the GLM heartbeat confound (individual predictors for oxy and deoxy, respectively).
Since the average of these channels can still contain global low frequency drifts, it is possible to additionally filter the predictor using the “Apply High-Pass filter” checkbox and the cutoff frequency next to it. Internally, a Butterworth high pass filter of order 3 is used.
Channel confound predictor
Depending on the layout and design of the experiment, it can be useful to add specific channel(s) as confound predictors to the GLM. These channels are either based on the fully preprocessed data or the raw oxy / deoxy signal, depending on whether the “Use fully preprocessed signal for GLM” option is selected. A channel can be added by adding the number of the source and detector to the line edit next to the “Channel confound predictor(s)” check box. It is possible to add up to seven channels.
If confound predictors for channels are added, a warning symbol is indicated in the top left of the channel selection table and the respective channel is marked with an (C) behind the name.
Note: To inspect the added confound predictors, you can navigate to the Analysis menu and select the GLM Diagnostics entry, as shown below.
Here you can select to show the confound predictors using the Show Confound Predictors option. These will be visualized in the selected channel plot (top left).
Short-distance channels confound predictors
To add all short-distance channels as confound predictors to the GLM just select the Add short-distance channels as confound predictors check box. Turbo-Satori recognizes short-channels as the channels having a higher detector number than normal detectors are available. If for example 8 normal detectors are connected the detector 8 to 15 are recognized as short-distance detectors (if only one short-detector bundle is used, the 8th detector is sacrificed for the short-detector bundle). The short-distance signal is either added as raw HbO/Hb signal or fully processed dependent on the Input for GLM option. Use the GLM diagnostic features to inspect the short distance time course or select the short-distance channels directly.
Note: Every short-detector must be connected to the last detector available.
To enable the correction for serial correlations, select the “Correct serial correlations” check box in the GLM Parameters dialog. It will perform a comprehensive autoregressive model (AR) based serial correlation correction. More details about this procedure can be found in the General Linear Model section. The residuals time course can be displayed using the menu shown above.
Input for GLM
Furthermore, one can decide to either use the raw HbO / Hb signal as an input for the GLM, or the fully preprocessed data. To use the raw HbO / Hb signal, uncheck the Use fully preprocessed signal for GLM checkbox. As described in recent articles by Huppert et.al. 2014, it is advised to use the not low pass filtered data but add respective confound predictors in the GLM. This can be achieved by turning off the low pass filter in the filtering options, while Use fully preprocessed signal for GLM is checked. This will apply different filters to the data used for the GLM.
GLM HRF Parameters
The two gamma HRF used in standard fMRI research might not result in the best fit for all fNIRS measurements and participants. Therefore, one can adjust the standard two gamma HRF by using the parameters shown in the top of the GLM Parameters dialog. In its standard settings, the HRF will result in the standard SPM HRF:
- Onset: 0
- Time to response peak: 6
- Response dispersion: 1
- Response Undershoot ratio: 6
- Time to undershoot peak: 16
- Undershoot dispersion: 1
Measured mean responses
This feature will be available in TSI Version 2.0.
Ten different conditions are currently supported for real-time application. To have a cleaner look on the contrast, it is possible to hide the unused conditions, meaning that only conditions, that are currently used in the experiment, are shown. This is done using triggers, received from NIRStar. if the “Hide unused conditions” check box is marked, and the trigger is received, the condition will be shown.
The contrasts can also be automatically balanced, which is recommended for any contrast. This means that the sum of all positive entries is equalized with the sum of all negative values.
Copyright © Brain Innovation B.V. 2019. All rights reserved.