Sign in to Peak Brain Path
Sign in to access your courses, books, and progress tracker. New here? Signing in creates your account automatically.
Want to explore courses first?
Browse courses and booksSign in to access your courses, books, and progress tracker. New here? Signing in creates your account automatically.
Want to explore courses first?
Browse courses and booksModule 16
Chapter 16 · 1.5 h · 8 quiz items · pass at 80%
BCIA Domain VII includes the QEEG-guided approaches because they are now standard in many practices. This module explains surface and source-space z-score training, the workflow on representative software, and the limitations a practitioner must hold in mind, including the normative assumption that z-score convergence is not the same as clinical improvement. The quiz proves the learner can describe a z-score protocol and state its boundaries.
The protocols in the previous chapters start from a category. Theta/beta is the inattention protocol. SMR is the regulatory protocol. Alpha-theta is the trauma protocol. You match the client's presentation to a protocol and run it. QEEG-guided training starts somewhere else: from this client's brain map, what is statistically deviant, and can I move it toward the normative range while watching the symptoms? Instead of asking "what protocol fits this diagnosis," you ask "what does this map say is off, and can I train it back."
That shift sounds like an upgrade, and sometimes it is. A brain map can show you a frontal theta excess, a posterior alpha deficit, or a coherence abnormality that you would not have predicted from the intake. Training those findings directly, with the deviation itself as the target and a real-time z-score as the feedback signal, is a clean idea: let the database define the target, reward the brain for moving toward normal. But the idea carries every assumption baked into the database underneath it, and a practitioner who runs z-score training without understanding those assumptions is being driven by software they cannot interrogate. This chapter is about doing it with your eyes open: how surface and source z-score training work, what the feedback signal actually is, the filter choices that change the numbers, what convergence does and does not mean, the software workflows, and when guiding from the map is the right call versus when the symptom in front of you should drive the protocol.
For the database foundations underneath all of this, the normative samples, the z-score math, the multiple-comparison problem, and the limits of comparing a person to a database, this chapter assumes Chapter 12 and leans on The QEEG Field Guide, Chapter 5. We will use those foundations here rather than rebuild them.
QEEG-guided neurofeedback uses the findings from a quantitative brain map to personalize two things: where you train (the site or sites) and what you train (the frequency targets and metrics). Instead of placing the electrode where the protocol category says, you place it where this client's map shows the deviation. Instead of rewarding a band the category prescribes, you target the band, the coherence pair, or the source region that the map flagged.
The chain runs map to hypothesis to protocol. The map shows, say, elevated absolute theta at Fz with a z-score of +2.4 in a client who presents with inattention. The symptom-based pathway would have sent you to a theta/beta protocol at Cz or Fz anyway, so here the map confirms the category. The interesting cases are where the map and the category diverge: the inattentive client whose map shows not theta excess but a posterior alpha deficit, or the anxious client whose map shows a focal coherence abnormality rather than diffuse beta excess. QEEG-guided training follows the map into those cases.
The key distinction inside QEEG-guided work is between guiding and z-score training proper. You can use a map to guide an otherwise conventional protocol: read the map, pick the site and band, then run a standard amplitude protocol at that target. That is QEEG-guided in the broad sense, and it is what Chapter 13's "QEEG-guided pathway" mostly describes. z-Score training is narrower and more specific: the feedback signal itself is a real-time z-score computed against a normative database, and the client is rewarded for moving that z-score toward zero. The rest of this chapter is mostly about the second kind, because it is the one with its own mechanics, its own software, and its own failure modes.
In conventional amplitude training, the feedback signal is microvolts: the client is rewarded when power in a band crosses a threshold you set. In surface z-score training, the software does something different. It takes the live signal, computes the same metrics a brain map computes (absolute power, relative power, asymmetry, sometimes coherence and phase), compares each one to an age-matched normative database in real time, and expresses each as a z-score. The feedback rewards the client for pulling deviant z-scores toward zero, toward the database mean.
The conceptual move is that you are no longer training a raw band amplitude up or down. You are training the brain toward the normative distribution on whatever metrics you have chosen. If a client's Fz theta sits at z = +2.4, the reward fires as that value moves toward +1, toward 0. The target is defined by the database, not by a microvolt threshold you picked by hand.
Single-channel surface z-score training puts one active electrode at the site of interest and trains the z-scores available from that one location. This is the simplest form and a reasonable place to start. You select which metrics to include in the reward (you rarely train all of them at once), set how many of the chosen metrics have to be in range for a reward, and the client works toward convergence at that site.
The appeal of z-score training is real: the target is individualized to the client's own deviations and age-normed automatically, you are training toward a defined reference rather than an arbitrary threshold, and the approach can address metrics (asymmetry, coherence) that are awkward to train with a single amplitude threshold. The catch is equally real and worth stating early: the feedback signal is only as trustworthy as the database it is computed against, and every limitation of database comparison from Chapter 12 (the normative-sample assumptions, the multiple-comparison inflation, the fat-tailed distributions that make extreme z-scores less rare than the percentile table claims) is now baked into the live reward signal. You cannot see the database doing its work; you only see the z-score it hands you. That is a reason to understand the database, not a reason to trust the number blindly.
Multichannel z-score training extends the same idea across many sites at once. Instead of one active electrode, you record from a montage (often four channels, sometimes the full nineteen) and the software computes z-scores across all of them simultaneously: power at each site, asymmetries between sites, coherence and phase across pairs. The reward is driven by some summary of how many of those many z-scores are in range, or by a weighted combination you specify.
The case for multichannel training is that brain dysregulation is often a network problem, not a single-site problem. A coherence abnormality lives in the relationship between two sites; an asymmetry lives in the difference between hemispheres. Training one electrode at a time cannot directly address a relational metric. Multichannel z-score training can target the network features the map shows, training the brain toward a normative pattern across the whole montage rather than at one point.
The case for caution scales with the number of channels. The more z-scores you fold into the reward, the more you run into the multiple-comparison problem from the live side: with nineteen channels and many metrics, you are computing hundreds to thousands of z-scores, some of which are deviant by chance, and a reward driven by "get most of these in range" can be chasing noise as much as signal. The practical answer is the same principle Chapter 12 applies to reading maps: target patterns, not isolated extremes. Choose the metrics and sites that the map shows as a coherent, symptom-relevant pattern, not every value that happens to exceed two standard deviations. A common middle path is the four-channel approach, enough channels to capture a network pattern, few enough to keep the reward interpretable. Whole-head nineteen-channel z-score training exists and has advocates, but the more channels and metrics you include, the more you are trusting the database and the summary statistic to separate signal from noise, and the less you can say about what specifically the client learned.
Surface z-score training works on scalp electrodes, and scalp electrodes have a resolution problem. The signal at Oz is a blurred sum of activity from a region of cortex, smeared by the volume conductor of skull and scalp. A deviation you see at a surface site does not tell you precisely where in the brain it is generated. Source localization addresses this.
LORETA (low-resolution electromagnetic tomography) and its standardized variant sLORETA estimate the three-dimensional distribution of current sources inside the brain that best account for the scalp recording. Instead of nineteen surface values, you get an estimate of activity in thousands of voxels across cortical space, which can be compared to a normative database the same way surface metrics are. sLORETA is low-resolution: it does not give you millimeter precision or reach deep subcortical structures cleanly, and the source estimate is a mathematical solution to an inverse problem that has no unique answer. But it does let you ask where, in source space, a deviation lives, which a surface map cannot answer. The QEEG Field Guide, Chapter 1, covers the principle and its limits in more depth; the relevant point here is that sLORETA gives you a source-space target.
LORETA z-score training takes the source-localization idea live. Rather than training surface z-scores, the software computes z-scores on sLORETA source estimates in real time and rewards the client for normalizing activity in a defined region of interest (or in a set of voxels, or in the current density of a specific Brodmann area). You are training the estimated source rather than the scalp signal.
The argument for source training is anatomical specificity. If a client's map localizes a deviation to, say, the anterior cingulate or a specific cortical region, surface training at the nearest electrode is training a blurred approximation of that target, while LORETA z-score training aims at the source estimate directly. For network and connectivity work, source-space training can target the regions and connections the map implicates rather than the scalp sites that happen to sit over them.
The argument for caution is that you are now training a feedback signal that is two inferences deep: an inverse-problem source estimate, then a z-score of that estimate against a normative database of source estimates. Each layer adds assumptions. The source solution is non-unique and low-resolution; the normative source database carries all the sampling and distributional issues of any database, plus the added uncertainty of being built on estimated sources rather than measured signals. None of this makes LORETA z-score training wrong, and in trained hands aiming at a clear, symptom-relevant source finding it is a legitimate and sometimes powerful approach. It does mean the practitioner has to hold the uncertainty honestly. The evidence base for LORETA z-score training is younger and thinner than for surface protocols (promising, resting on case series and clinical reports more than controlled trials), and the temptation to treat a colorful three-dimensional source map as more certain than it is should be resisted. The resolution of the picture is not the resolution of the knowledge.
This is the part of z-score training that practitioners skip and should not, because the filter sitting between the raw signal and the z-score changes the z-score. The question of how the software extracts band power before comparing it to the database is not a back-end detail; it is part of what the reward means.
Any frequency-band measurement requires a filter to isolate the band. Filters differ in two properties that matter here: how sharply they separate the target band from neighboring frequencies (their amplitude response), and how faithfully they preserve the timing and shape of the waveform (their phase response). These two goals trade off, and different filter designs make different trades.
A Bessel filter is designed to preserve phase: it introduces a nearly constant time delay across frequencies, which keeps the shape of the filtered waveform faithful at the cost of a gentler, less sharp cutoff between bands. A Butterworth filter (the classic "bandpass" in many descriptions) is designed for a flat amplitude response in the passband and a sharper cutoff, at the cost of phase distortion near the band edges. For real-time z-score neurofeedback, phase fidelity and low latency matter, because the feedback has to track the client's state moment to moment and reward the right thing at the right time. A filter that distorts phase or adds latency can shift when the reward fires relative to what the brain is actually doing. This is why several z-score platforms favor minimal-phase or phase-preserving designs for live training, and why "Bessel versus Butterworth" is not a trivia question but a description of two different feedback signals computed from the same raw EEG.
The practical consequence: the same recording, run through different filters, produces different band-power estimates and therefore different z-scores. A normative database was built with a particular filtering and processing pipeline, and a valid z-score requires that your live processing match the database's assumptions closely enough that the comparison means something. When two platforms report different z-scores for the same client, filter and processing differences are a common reason, alongside the database differences covered in Chapter 12. You do not need to design filters. You do need to know that the filter is part of the measurement, that phase and latency matter for live feedback specifically, and that a z-score is a number produced by a pipeline, not a fact read off the brain.
In z-score training, "convergence" is the visible goal: over a session and over a course, the trained z-scores move toward zero, deviant metrics pull back toward the normative range, and the proportion of values in range climbs. On the screen, convergence looks like success. The client is doing exactly what the protocol asks: moving their brain toward the database mean.
Two questions have to stay separate, and conflating them is the central interpretive error in z-score training.
The first question is whether the z-scores are converging. That is a measurement question, and the software answers it: are the trained metrics moving toward zero, within and across sessions. You can read this directly and you should track it.
The second question is whether the client is getting better. That is a clinical question, and the software does not answer it. A z-score moving from +2.4 to +0.8 is statistical movement toward the norm; it is not, by itself, evidence that attention improved, anxiety dropped, or sleep consolidated. Normalizing a deviation that was never functionally relevant produces a tidier brain map and no clinical change. The deviation might have been a benign variant, a genetic trait, or a finding with no bearing on the client's symptoms, in which case you can train it to zero and the client feels nothing. Chapter 12 makes this point about reading maps; it applies with double force when the map's deviations are driving a live reward, because the protocol can manufacture convergence on metrics that did not matter.
The discipline that follows is simple to state and easy to neglect: track the z-scores and track the symptoms, and treat clinical outcome, not convergence, as the measure of success. Convergence with symptom improvement is the result you want. Convergence without symptom improvement means you normalized something that was not the problem, and it is a signal to re-examine the map, the targets, and whether QEEG-guided training is the right approach for this client at all. The brain map getting prettier is not the goal. The client getting better is.
NeuroGuide, built on Robert Thatcher's normative database, offers a live z-score training mode that is among the most established surface and LORETA z-score platforms. The workflow gives a concrete picture of how z-score training runs in practice.
You begin from an assessment. A brain map identifies the deviant metrics and the sites or source regions that match the client's presentation. From that map you choose your training targets: which metrics, at which sites (or which sLORETA regions for LORETA z-score mode), and how many of them have to be in range to drive a reward. The live screen displays the running z-scores for the selected targets, the feedback signal the client experiences (a visual display, a game, audio, or media that responds to convergence), and the raw signal so you can watch for artifact. Threshold configuration in z-score mode is less about picking a microvolt level and more about deciding the z-score window you are rewarding and how many targets must sit inside it, with the database supplying the reference. You watch the raw EEG throughout, because an artifact that inflates a z-score will drive a false reward exactly as it does in amplitude training.
The general shape, brain map to target selection to live z-score feedback to artifact-watched session, holds across z-score platforms even though the screens differ. NeuroGuide is one specific, well-validated implementation; the workflow logic transfers.
BrainMaster integrates hardware (the Discovery and related amplifiers) with software that supports surface and multichannel z-score training, including z-score modes licensed against normative databases. The relevant practitioner point is the hardware-software integration: the amplifier, the acquisition software, and the z-score engine have to work as one system, with the database, the filtering pipeline, and the feedback display matched so the z-scores mean what they claim. BrainMaster's four-channel and full-cap z-score options are a common route into multichannel and whole-head z-score work.
Other platforms occupy the same space with their own database licenses and processing pipelines. The qEEG-Pro database (the largest resting-state normative sample of any commercial system, covered in Chapter 12) is offered for z-score applications through partner software. Several acquisition systems and clinical EEG platforms support z-score or normative live training in some form. The specific menus and module names differ from vendor to vendor and change with versions, so the durable knowledge is structural rather than brand-specific: a z-score platform is an amplifier plus an acquisition pipeline plus a normative database plus a filtering and processing chain plus a feedback display, and a valid session requires all of those to be matched and the practitioner to know which database and which processing pipeline are producing the numbers. When you evaluate a platform, ask which database it uses, what the normative sample is, how it filters, and whether it supports surface, multichannel, and LORETA modes. Those answers, not the brand, tell you what the z-scores will mean.
z-Score training is built on a map, and a map is a snapshot. Over a course of training the brain changes, which means the map that defined your targets goes stale. Re-assessment is how you keep the targets honest.
The logic for when to re-map follows from the protocol. An initial brain map defines the targets and the baseline. A mid-course re-assessment (commonly somewhere around session 10 to 20, adjusted to the client and the pace of change) tells you whether the trained metrics are actually converging on a full map rather than only on the live training display, whether new deviations have emerged or old ones resolved, and whether the targets should change. An end-of-course map documents where the client landed and supports the discharge decision. The exact session numbers are clinical judgment, not a fixed rule; the principle is that you re-map when you need to know whether the target still describes the client.
Re-assessment also guards against the convergence trap from earlier in the chapter. If the live z-scores have converged but a fresh full map shows the deviations returning the moment training stops, or shows that you normalized one metric while another drifted, the re-map catches what the running display cannot. Phenotype stability over time is the rationale here, and The QEEG Field Guide, Chapter 7, develops it: some patterns are stable traits that resist training and some are state-dependent and shift, and re-mapping is how you tell which you are dealing with. One operational caution carries over from Chapter 12: a re-assessment is only comparable to baseline if you hold the conditions constant. Same reference, same montage, same eyes-open and eyes-closed protocol, same artifact standard, and as far as possible the same medication status and time of day. A re-map run under different conditions is measuring the conditions, not the training.
z-Score training generates more numbers than any other protocol in this book, and the documentation has to capture them or the course is not reconstructable. The record for a z-score course needs the targets (which metrics, which sites or source regions, which database), the baseline z-values from the initial map, the reward configuration (the z-score window, how many targets in range), and a session-by-session account of the trained z-scores alongside the clinical picture. The two have to sit side by side in the record, because the whole interpretive discipline of the chapter, convergence is not the same as improvement, depends on being able to read the z-score trajectory and the symptom trajectory together. The re-assessment maps go in the record as the objective markers of change across the course. Chapter 19 covers session-note requirements in general; z-score training adds the obligation to record the specific quantitative targets and their movement, because "trained z-scores at Fz and Pz" is not a record, and "Fz absolute theta from z = +2.4 to z = +1.1 over twelve sessions, PCL-5 down 14 points" is.
Every limitation of database comparison from Chapter 12 applies to z-score training, and a few sharpen when the database is driving a live reward rather than informing a written report.
The deepest limitation is the normative assumption itself. z-Score training treats movement toward the database mean as the goal, which assumes that normal, for this client, is good. That assumption holds often and not always. A high-performing outlier, a benign genetic variant, an adaptive specialization, and a medication-driven pattern can all read as deviant against the norms, and training them toward the mean treats normality as a target when it may be irrelevant or even counterproductive. The database does not know which of a client's deviations matter; it only knows which are statistically unusual. You supply the clinical judgment about which deviations are worth training, and the protocol cannot supply it for you.
The fat-tailed-distribution problem from Chapter 12 has a live edge here. Because EEG metrics have heavier-than-Gaussian tails, a z-score of +2.0 is less rare and less meaningful than the percentile table implies, and a reward driven by "pull this z toward zero" can be working hard on a deviation that was partly an artifact of the distribution's shape. Weighting the larger, pattern-consistent deviations over isolated extreme values is as much a z-score-training principle as a map-reading one.
The multiple-comparison problem reappears every time you add metrics and channels to the reward. More targets means more chances that some are deviant by noise, and a summary reward can chase those. The defense is the same throughout: target the coherent, symptom-relevant pattern, not every value past two standard deviations.
And the central thing z-score training does not tell you is whether the client is better. It tells you whether the brain moved toward the norms on the metrics you chose. That is worth knowing and it is not the same as clinical improvement, which is why this chapter keeps the two questions apart and why the symptom record, not the convergence display, is the measure that decides whether the course worked.
When QEEG-guided differs from symptom-based comes down to this. Use the map to guide when the client's presentation is complex, atypical, or treatment-resistant, when a coherence or source finding implicates a target you would not reach from symptoms alone, or when an objective marker of change is clinically or administratively valuable. Let the symptom drive when the presentation is a clean fit for a well-evidenced category protocol, when the map would only confirm what the symptom already tells you, or when the map's deviations do not match the clinical picture. The map is a powerful guide and a poor master. At the chair, the sequence that keeps you honest is the same every time: read the map, choose targets that match the symptom, watch the raw signal so you are not rewarding artifact, track the z-scores and the symptoms side by side, re-map to check that the target still describes the client, and judge the course by whether the client improved, not by whether the numbers converged.