Technical Foundations of Neurofeedback Principles and Processes for an Emerging Clinical Science of Brain and Mind

Chapter 9 – LORETA Neurofeedback

There has long been interest and progress in determining the brain activity that underlies the EEG, without using invasive sensors. Various solutions to this problem have been developed, and all of them incorporate a solution to the “inverse” problem described previously. As was pointed out, there are an infinite number of possible source configurations that can lead to given measurement, so this has been a difficult problem to solve. Practical solutions require additional assumptions to be made, and may also depend on empirical data such as actual measurements, or trial solutions, in order to limit the possible range of solutions. Generally, the simplest, most likely, or otherwise “best” estimates are what are produced by particular methods. While these approaches have evolved for decades, they have not been relevant to neurofeedback until real-time implementation was possible, allowing calculation and feedback of source information without excessive delay.

Low Resolution Electromagnetic Tomographic Analysis (LORETA) was developed and has been described by Pasqual-Marqui (2002). It is a mathematical technique that uses scalp data from a whole-head EEG, and computes an estimate of the cortical electrical activity that could have given rise to that surface potential. Note that this estimate is only one of an infinite number of possible solution, but it is based upon simplifying principles, and has been subjected to independent validation.

sLORETA is based upon the same basic scientific principles as LORETA, but includes three additional important features. The first is that 6,239 voxels are used (the amygdala is not included in sLORETA. The second is that it incorporates an assumption of smoothness into the solution. This determines a unique solution that is intuitively the simplest brain dipole configuration that can lead to the observed surface potentials. The third is that the algorithm has zero localization error.

The introduction of LORETA-based techniques to neurofeedback introduces several important and unique capabilities. It also places certain new demands on the practical aspects of training. The most important and obvious factor introduced by these approaches is that it is possible to provide feedback related to the activity of a particular region or regions of the brain, rather than basing training on scalp activity. The ability to train regionally opens the possibility of reinforcing activity in, for example, the anterior cingulated gyrus, or the insula, in contrast to simply Fz or Cz. This has some impact on typical downtraining, such as theta inhibit, in that the feedback will be more directly related to specific activity in the chosen area. Selected brain regions are referred to as “regions of interest” or “ROI’s” in practice. More importantly, it is also possible to train particular higher frequencies, such as beta, in a very specific manner. Because of the more localized specificity of higher-frequency activity in the cortex, localized training becomes more important in this regard. It might be harder to justify uptraining beta on Fz, for example, than it would be to do the same training specific to a site known to be exhibiting an deficit, such as the anterior cingulate.

In order to achieve the benefits of localized training, it is necessary to provide sufficient leads to compute the internal “solution.” Generally, all 19 10-20 sites are required to do training based on the inverse solution. The solution consists of values computed for thousands of spatial regions, known as “voxels.” In the same way that a two-dimensional image can be broken into “pixels,” a three-dimensional solid can be broken into these voxels. A key aspect of such localization is the size and number of voxels. LORETA uses approximately 2,300 7-millimeter voxels, while sLORETA and eLORETA use 6,239 5-millimeter voxels.

A second important requirement of practical regionalized training is that the EEG must be free of artifacts. Any contamination in the scalp signals will be included in the computation of the inverse solution. If activity is coming from the eyes or muscles, for example, the assumptions of the computation are violated, and the results will be unreliable.

The LORETA (or sLORETA or eLORETA) computation is based upon a matrix operation that converts the scalp data (“lead field”) into the internal brain representation. For example, an sLORETA solution will involve a matrix that has 19 columns and 6,239 rows, and contains three elements in each cell. These elements represent the x, y, and z axes of the dipole solution for each voxel. Therefore, a single solution for one time-point will require 19x6,239x3 = 355623 computations. At 256 samples per second, and for 10 frequency bands, this equates to over 91 million floating-point computations per second, in addition to the data acquisition, signal processing, and display demands for all 19 channels, plus any images being shown.

Therefore, a significant issue with inverse solutions is the time required for the computation. Traditionally, this has been very long, and complete real-time solutions including graphic displays have not been possible until relatively recently. Solving the problem of delay can be handled in two basic ways. One is to reduce the work by selecting a subset, or representative voxels, for the computation. This method is described by Lubar (2003), Thatcher (2010) and by Kaiser (2010), and must incorporate assumptions or rationale regarding the specific voxels chosen, and their use for feedback. Another approach is to implement the entire mathematical transformation in real time, using specific computer techniques beyond the scope of conventional “programming.” Modern personal computer hardware has been increasingly oriented toward the needs of media, animation, and gaming, and graphics hardware has become a sophisticated resource in addition to main computer processing resources. It is possible to exploit the distributed multiprocessing resources designed for these applications in the solution of the inverse calculations, using techniques proprietary to this area. As an example, at the time of this writing, a typical sLORETA solution can be computed for all 6,239 voxels at about ½ real time. That means that a 20-second record would take about 40 seconds to compute values for all voxels, for all data points at 256 samples per second. Using advanced media and game-oriented techniques incorporating specialized hardware, it is possible to compute a complete sLORETA solution faster than 10 times real-time. For neurofeedback, it is thus possible to compute inverse solutions for all voxels and all frequency bands in real-time, and provide immediate feedback for operant training. Using this approach, rather than using a representative voxel for a region, the entire activity of that region can be imaged and fed back. Figure 9-1 shows an example of an sLORETA solution for the cingulate gyrus. The difference in activity from front to back is clearly evident. A single voxel approach would be limited in its ability to accurately represent the activity of this region, because the voxel chosen might not accurately represent the activity of the entire region.

LORETA-based techniques can be combined with z-score concepts, to provide assessment and training of voxels based on normative or other references. In order to achieve regionalized z-scores, a reference must be computed for every voxel to be estimated. This increases the size of the reference database considerably, as now the typical 19 scalp sites are now replaced by thousands of voxels, and normative data must be known for every voxel.

The voxel data are computed in three dimensions, producing a vector for each voxel. That means that each voxel has not only a magnitude, but also a direction, or “moment” associated with it. Theoretically, this dipole will reflect the direction of depolarization of the pyramidal cells in that voxel. Therefore, LORETA-based approaches can be used to estimate voxel activity both as a quantity across time, and also as a direction in space.

By combining brain activity images across time, it is possible to visualize and interpret complex sequences of operations, rather than the static representation of a state or condition. Fore example, the amount of beta activity may be normal in an individual, but the pattern of activation may reveal nuances that only appear in the combined space and time analysis. As one example, Figure 9-2 shows a snapshot of beta activity from a client who was depressed as well as anxious. While the static QEEG showed no particular abnormality in beta activity, a time review of the animation revealed a pattern of cyclic beta activity between the right dorsolateral frontal cortex, the cingulate gyrus, and the left parietal area. This activation can be interpreted as mediating a negative mood, and alternating rumination that involves the “individual as self.” By incorporating this type of information into therapeutic interventions, with our without neurofeedback, this dynamic imaging capability provides a valuable illumination of clinically relevant data.

It is possible to combine voxel data into collections that constitute relevant brain regions of interest (ROI’s). Examples of these ROI’s are shown in the accompanying text box. When selecting ROI’s, the system should compute a suitable representation of the activity of that ROI. The optimal method is to calculate the activity of all of the constituent voxels, and to compute an average or similar metric that adequately represents that region.

LORETA/sLORETA Regions of Interest (ROI’s)
Frontal Lobe

Limbic Lobe

Occipital Lobe

Parietal Lobe

Sub-lobar

Temporal Lobe

Angular Gyrus

Anterior Cingulate

Cingulate Gyrus

Cuneus

Fusiform Gyrus

Inferior Frontal Gyrus

Inferior Occipital Gyrus

Inferior Parietal Lobule

Inferior Temporal Gyrus

Insula

Lingual Gyrus

Middle Frontal Gyrus

Middle Occipital Gyrus

Middle Temporal Gyrus

Orbital Gyrus

Paracentral Lobule

Parahippocampal Gyrus

Postcentral Gyrus

Posterior Cingulate

Precentral Gyrus

Precuneus

Rectal Gyrus

Sub-Gyral

Subcallosal Gyrus

Superior Frontal Gyrus

Superior Occipital Gyrus

Superior Parietal Lobule

Superior Temporal Gyrus

Supramarginal Gyrus

Transverse Temporal Gyrus

Uncus

Brodmann areas 1-11, 13, 17-25, 27-47