Evoked response analysis
From NUTMEGwiki
Once you have computed the forward lead field, next compute the weights and source activation estimates. Click on the "Compute Activations" button on the main GUI, and a new GUI will appear.
There is a drop-down menu of at least 10 different types of source localization methods. The overall idea is to provide as many options for the user to decide which method makes most sense for their data/analysis types. (We plan to make a [tutorial exploratory problem set] for new users to become familiar with the options and hopefully gain insight about the different methods).
One general option is either a scalar or vector set of weights. A vector weight computes the weight for all (either 2 or 3) components available in the lead field. Then the orientation of the dipole is computed using the power in each orientation or can be viewed individually. Alternatively, a scalar weight computes the optimal data-dependent orientation and outputs only one power per voxel.
The weight can be computed for an active period only, or relative to a control period.
For comparing stimulus conditions, the weight should be computed over an average of all stimulus types first, then appropriate contrasts between conditions can be made.
The data can be averaged either before or after data covariance computation.
To avoid the bias towards the center of the head, options exist: lead field (column) normalization, weight/noise normalization, something like a neural activity index (with denominator either scalar*identity or full-rank).
You may remove the DC offset of the prestimulus period. Also in that pulldown menu you may linearly detrend, however highpass filtering at 1Hz is also recommended instead of linear detrending. (CTF 3rd order gradient may also be used insetad). Notch filters for 60Hz or 50Hz power line noise can be applied, but they should only be used if absolutely necessary (i.e., the power supply is still clearly visible in your averaged dataset.) The upper limit of the bandpass filter is set to half the sampling rate.
The condition number of your selected interval will appear in the Matlab command window. Lower conditions numbers are better since they indicate that the covariance matrix is less singular (and therefore can be inverted more reliably). A condition number as high as 10^12 is okay, however if the condition number is higher, some method of regularization will occur (Tikhonov by default). This regularizaton usually leads to more spatially blurred images, and therefore is not ideal. Thus, you should alter the number of sensors, the time window, and the filtering, to reduce the condition number. The general rule of thumb: more sensors -> higher condition number; smaller time window -> higher condition number; and smaller bandpass window -> higher condition number.
Once you have selected the weight method, time window, filtering, and number of eigenvalues, then you may edit the name of the file that the activations will be saved to. A default name will appear as s_beam_datasetname.mat. You may use this name, or also append additional information, such as what weight type, time window chosen, etc.
Hit 'Proceed' button.
With Matlab 7, the results will be automatically displayed, otherwise you will be prompted for the file you just created to load the results and view them.
Specific to Eigenspace beamformer
This type of beamformer is data-dependent, and thus you must select which aspect of your data to use to calculate the weights, which use the sensor data covariance. The weights will then be applied to your sensor data to reconstruct a time series at every voxel point within the head within your VOI.
You want to select the portion of your data that has the signal/peak of interest in it (e.g. the peak at 50ms shown above) but also a large enough time window around this peak so that the sensor covariance matrix is not too singular (in other words, so that the data from each sensor aren't too similar to the other sensors over the selected time window.) A very singular matrix cannot be reliably inverted, resulting in poor source reconstructions. For the shown SEF dataset above, a time window of 0-235ms works well. Note that the source activations will be computed for the entire time window (e.g. here -150ms to 240ms), but only 0-235ms will be used for calculating the sensor covariance used in weight computation.
The right plot is the magnitude of eigenvalues, ordered from largest to smallest. Each represents a component of your chosen dataset. The beamformer of this toolbox uses just the "signal" subspace of the sensor data. Presumably, the largest eigenvalues correspond to the "signal" space whereas the smaller ones correspond to the "noise" space. Between 2-4 eigenvalues is generally recommended, however you should examine your dataset and choose the appropriate number based on each individual dataset. The first three eigenvalues of the shown dataset appear to stand out well above the rest, so three would be a good choice here.
You can view the eigenvector for each eigenvalue in the inset plot within the righthand window, as shown above. You can view several at a time by entering a vector in the "view single eigenprojection" textbox, such as 1:4, which will show the first four eigenvectors at once.
The more eigenvalues chosen, the more spatial variations of your results will be seen. In the limit of only 1 eigenvalue, the time series at every grid point will have an identical shape, only at different scales. Two eigenvalues will show a little spatial variation in the timeseries, but also is less likely to have "noisy" spatial blobs; therefore, if you are interested in one or two major peaks, then two eigenvalues is best. For several areas of interest, 3-4 eigenvalues are recommended, and maybe up to 5-6.
