Coherence and Connectivity Training

Part 3 of this series ended on a deliberate handoff. Bipolar protocols, I argued, are the simplest member of a larger family — a two-point, single-band measure of how two sites relate. Coherence and connectivity training is that same conceptual move scaled all the way up: instead of one difference between two electrodes, we are now training relationships across many sites, in many bands, sometimes dozens of relationships at once. The promise is real, and so are the interpretive hazards. Both get bigger with scale.

This is the post where the series steps off the amplitude map entirely. Classic amplitude training (Part 2) and bipolar montages (Part 3) are, at bottom, about how much of a given rhythm is present — at a site, or as a difference between sites. Connectivity training asks a different question: are these regions working together, and how? That is not a heavier version of the amplitude question. It is a question about the architecture of communication in the brain rather than the volume of any single voice in it.

For readers who need the one-sentence orientation: neurofeedback is a learning-based method in which clients learn to modulate aspects of their own brain activity through real-time feedback, supported by reinforcement, attention, and a coaching relationship with a trained practitioner. (Biofeedback applies the same principle to peripheral physiology — heart rate, respiration, skin conductance, muscle tension.) Connectivity training keeps that learning chassis intact. What changes is the target: not amplitude in a band, but a measure of the statistical relationship between two or more signals.

Why should a clinician care? Because a lot of the conditions practitioners actually struggle with are not well described as "too much theta here" or "too little SMR there." Traumatic brain injury disrupts white-matter tracts and the coordination they support. Autism-spectrum presentations have been characterized — contestably, but seriously — in terms of atypical local and long-range connectivity. Complex anxiety, post-stroke reorganization, certain learning profiles: these are, at least plausibly, network problems (and compensations). If the clinical problem is a coordination problem, then a method that targets coordination directly is not a gimmick. It is, conceptually, the right tool.

The trouble is the word plausibly, and the gap between what we can measure at the scalp and what is actually happening in the tissue. This entry is going to spend real time on that gap, because it is where connectivity training is most often oversold and most often misunderstood — by enthusiasts and skeptics alike. My aim, as in every entry, is to give the method a fair turn: to name what it does that nothing else does, and to be honest about the interpretive discipline it demands before it can be trusted.

A brief history

Connectivity did not start as a training target. It started as a way of reading the EEG.

The idea that two EEG signals can be more or less related is old — coherence as a mathematical measure was borrowed from signal processing and applied to the EEG decades before anyone tried to train it. What turned it into a clinical concept was the maturation of quantitative EEG. Once you could compute, from a multi-channel recording, the degree to which two sites shared a stable phase-and-amplitude relationship in a given band, you had a number — and once you had a number, you could compare it against a normative reference and ask whether it was unusual.

Robert Thatcher's work sits near the center of this lineage. His developmental studies in the 1980s — including the much-cited finding that the cerebral hemispheres mature at different rates, read through patterns of EEG coherence across the lifespan — helped establish coherence as a meaningful index of cortical organization rather than a statistical curiosity. The normative-database tradition that grew out of this work (the comparison of an individual's connectivity metrics against an age-matched reference sample) is the backbone of most connectivity-guided neurofeedback practiced today.

The move from reading to training came later, and it came from several directions at once. Jonathan Walker and others reported coherence-training protocols for traumatic brain injury and headache, selecting deviant coherence values from a qEEG and reinforcing them toward the normative range. Robert Coben developed assessment-guided connectivity protocols for autism-spectrum disorder, working from the premise that the relevant abnormality was relational rather than local. And around 2008–2010, the approach that has defined the modern era of this method arrived: live z-score training, developed by Thatcher and implemented commercially by Thatcher, Tom Collura and others. Instead of training one coherence value, live z-score training compares many metrics — power, ratios, asymmetry, coherence, phase — against a normative database in real time, and reinforces the brain for moving any or all of them toward the reference range simultaneously.

That last development is worth pausing on, because it changed the character of the method. Earlier connectivity work was hypothesis-driven: the practitioner picked a specific relationship, on a specific clinical rationale, and trained it. Live z-score training is, by design, closer to "let the brain decide." It hands the optimization problem to the client's nervous system and rewards normalization wherever it occurs. There is something genuinely appealing in that — and, as I will argue later, something that needs careful handling. The history of this method is, in part, the history of a steady move from a single, interpretable target toward many simultaneous, harder-to-interpret ones.

Alternate names

The vocabulary here is even muddier than it was for bipolar, partly because "connectivity" is an umbrella term covering several mathematically distinct measures, and partly because vendors and lineages use the words loosely. A short orientation, because the first practical skill in this method is knowing exactly which quantity a protocol is actually training:

• Coherence training. The most common name, and often used as a catch-all. Strictly, coherence is a frequency-domain measure of the consistency of the phase-and-amplitude relationship between two signals, normalized to a 0–1 scale. High coherence means the two sites maintain a stable relationship in that band over time. It is frequently described as a kind of linear correlation in the frequency domain, which is close enough for orientation but hides important detail (coherence conflates phase consistency and amplitude covariation).

• Connectivity training. The broader umbrella. Includes coherence but also phase-based measures, amplitude-covariation measures, and network metrics. When you see "connectivity," your first job is to find out which underlying measure is meant.

• Phase training, phase synchronization, phase-lag training. Names for protocols that target the timing relationship between sites rather than the shared power. This family includes measures designed to be more robust to volume conduction — the phase-lag index being the best known — precisely because zero-lag relationships are the ones most likely to be artifactual. If a protocol claims to train "phase," ask whether it is using a volume-conduction-resistant measure or a naive one.

• Comodulation training. The Sterman-Kaiser lineage. Comodulation indexes the correlation of amplitude fluctuations between two sites over time. It is a connectivity-adjacent measure, but it is not coherence — it is about whether the envelopes rise and fall together, not whether the rhythms are phase-locked. The distinction is significant and gets blurred constantly.

• Live z-score training (LZT / LZS), z-score neurofeedback. The real-time-normative-comparison approach described above. The "z-score" refers to how far each metric sits from the normative mean, in standard-deviation units; training rewards movement toward zero. Often run on two channels, four channels, or more, with several metrics trained at once.

• Surface connectivity vs source connectivity (LORETA connectivity, swLORETA / eLORETA connectivity). Whether the relationships are computed between scalp electrodes or between estimated cortical sources. This is one of the most important distinctions in the whole method, and Part 5 will take it up directly — source-space connectivity partially addresses the volume-conduction problem that haunts surface connectivity.

• Network training, graph-based neurofeedback. Names that invoke graph-theoretic metrics — hubs, small-worldness, nodal efficiency. These are mostly analysis constructs at present; training them directly in real time is uncommon and interpretively very demanding. Treat the term with caution when it appears in marketing.

When you read a connectivity protocol, translate the label into the measure. "We trained coherence between F3 and F4" tells you almost nothing on its own — coherence in which band, computed against which reference, at what montage, and is it the naive zero-lag version or a phase-corrected one? The protocol name does not tell you. You have to ask.

How the method works

The hardware floor is higher than for amplitude or bipolar work, and it rises with ambition. A single coherence pair can be trained with two channels. Connectivity-guided, qEEG-informed work — the kind worth doing — generally presupposes a full multi-channel assessment, which in practice means a 19-channel recording referenced to a normative database. Live z-score training runs on two, four, or more channels depending on the system and the protocol. The same professional-grade amplifiers and software that run amplitude and bipolar work (Thought Technology's ProComp lines with BioGraph Infiniti, Mitsar with WinEEG, and the dedicated normative-database environments) handle connectivity, with the appropriate analytic modules added.

The assessment comes first, and it is not optional. A connectivity protocol that is not anchored in a qEEG is, in my view, not a connectivity protocol — it is a guess with extra channels. The practitioner records a resting (and often task) EEG, processes it carefully (artifacting is more consequential here than anywhere else in the series, for reasons the next two sections will make clear), and compares the individual's connectivity metrics against an age-matched normative reference. The output is a map of which relationships, in which bands, deviate from the reference range — and by how much.

From there, protocol selection takes one of two broad forms.

The hypothesis-driven form is closer to the older tradition. The practitioner reads the qEEG, forms a clinical hypothesis — say, that a hypocoherent relationship between two frontal sites in a particular band is relevant to the client's presentation — and trains that specific relationship, reinforcing movement toward the normative range. Few targets, explicit rationale, interpretable.

The live z-score form hands more of the work to the algorithm. The system computes many metrics in real time — power, ratios, asymmetry, coherence, phase relationships — and rewards the client when some specified proportion of them sits within the normative range. The brain is, in effect, invited to find its own route toward the reference profile. The practitioner sets which metrics are in play, how many must be in-range for reinforcement, and how tight the criterion is, but the moment-to-moment "which metric improves now" is left to the nervous system.

Within a session, the surface experience looks familiar. Impedance check. Twenty to thirty minutes of active training, two to three times a week in most clinical practices. Feedback delivered through the usual channels — a game element, a video that brightens and clarifies, an audio tone. Threshold logic of the same general shape as amplitude training: set the criterion so the client succeeds often enough early on (the 60–80% reinforcement-rate heuristic still applies), then tighten as performance consolidates.

What changes — and it changes a great deal — is what success means. In amplitude training, success is "more SMR at this site." In connectivity training, success is "these relationships moved toward the reference profile." The client cannot see that happening in any intuitive way, the practitioner is watching a more abstract quantity, and — crucially — the number of things that can change to produce a given reward is much larger. That is the seam where the interesting problems live.

Mechanistic specifics

What is being trained, mechanistically, in a connectivity protocol? The intended answer is: the degree and timing of coordination between cortical regions. The honest answer is: a statistical relationship between scalp signals that may or may not track the coordination we think it does. Holding both of those at once is the whole skill.

Start with the intended target. The brain does its work through coordinated activity across distributed regions. Oscillatory synchronization is one of the leading candidate mechanisms for how that coordination is achieved — regions that oscillate in a consistent relationship can exchange information more effectively than regions that do not. If that picture is right, then a measure of how consistently two regions maintain a phase relationship in a given band is, at least in principle, a window onto functional communication. Training that measure toward a normative profile is, at least in principle, training the brain's coordination toward a more typical pattern. That is the mechanistic story, and it is a reasonable one.

Now the complications, in order of how much they keep me up at night.

Volume conduction. This is the ghost in the machine, and it is the single most important thing to understand about surface connectivity. The skull and scalp smear electrical activity. A single cortical (or deeper) source projects to multiple electrodes at once, essentially instantaneously. Two electrodes picking up the same source will show a high coherence — a beautiful, stable, zero-lag relationship — that reflects no communication whatsoever between two regions. It reflects one source, two microphones. Naive coherence cannot distinguish this from genuine connectivity. A great deal of the high zero-lag coherence in any scalp recording is partly or wholly an artifact of volume conduction, and reinforcing it trains the client to do more of something that may not be a real network property at all.

How the amplifier samples its channels. There is a quieter version of this problem that is easy to miss, because it never shows up in the output — it lives inside the amplifier itself. Connectivity measures assume that every channel was recorded at the same instant, so that any timing difference between two sites must be coming from the brain. Some amplifiers honour that assumption and sample all channels simultaneously. Others, to keep the hardware cheaper, read the channels one after another through a single shared converter — a design usually described as multiplexed sampling — which builds a tiny, fixed delay into the recording between one channel and the next. That delay is small, but the connectivity calculation has no way of knowing it is there: it reads the amplifier's delay as if it were the brain's timing. At best this adds error. At worst it manufactures a relationship that is not real — and it can even fake the kind of time-lag that practitioners rely on to tell genuine connectivity apart from a volume-conduction artifact, quietly defeating the very safeguard that is supposed to protect them. The important point for a clinician is that this is not something the analysis can repair; it is a property of the equipment. It is worth knowing whether your system samples all channels simultaneously, or whether your software corrects for any inter-channel delay — the spec sheet will usually call the good version simultaneous sampling. It is also one more reason, as Part 1 argued, that the choice of hardware is not method-neutral.

Reference and montage dependence. Coherence values change — sometimes dramatically — depending on the reference. A linked-ear reference, an average reference, a Laplacian montage will each yield different connectivity numbers from the same underlying data, because the reference itself carries activity that is shared across channels and inflates apparent coherence. There is no neutral choice here, only a set of trade-offs the practitioner has to understand. Two practitioners training "F3–F4 coherence" on different references are not necessarily training the same thing.

Phase versus power. Coherence conflates two things — the consistency of the phase relationship and the covariation of amplitude. Two sites can be tightly phase-locked with unrelated amplitudes, or co-vary in amplitude with sloppy phase. These are different physiological phenomena. The phase-based measures (phase-lag index and relatives) were developed in part to isolate timing relationships and, by construction, to suppress the zero-lag relationships most likely to be volume-conduction artifacts. When a protocol trains "coherence" without specifying, you do not know which physiological phenomenon is in play.

Cross-frequency relationships. Some of the most interesting connectivity in the brain is between frequencies, not within them — the phase of a slow rhythm modulating the amplitude of a faster one (phase-amplitude coupling). This is mostly a research construct at present and is rarely trained clinically, but it is worth knowing it exists, because it is a reminder that "connectivity in a band" is only one slice of how regions relate.

Indeterminacy, again — and larger. Part 3's bipolar problem was that one differential value is consistent with many underlying-site configurations. Connectivity training inherits that problem and multiplies it. When live z-score training rewards the brain for bringing some proportion of many metrics into range, an enormous family of underlying states can produce the same reward. Worse, the practitioner often cannot reconstruct which metric improved on a given trial, let alone whether the improvement was a real network change or a shift in an artifact-prone surface statistic. The system can report that the z-scores normalized. It usually cannot tell you why, and neither can you.

Operantly, the conditioning logic is unchanged. The brain learns to spend more time in states that satisfy the reinforced condition. But the reinforced condition is now a relationship — or a bundle of relationships — rather than an amplitude at a site, and the relationship is measured through a glass that volume conduction, reference choice, and limited reliability have all fogged. The conceptual handle for this entry is the shift from training activity (Part 2), to training a difference (Part 3), to training a web of relationships — each step more powerful in what it can in principle address, and each step harder to read honestly.

Overview of the science base

The science base for connectivity training is best described as real clinical signal embedded in serious interpretive uncertainty. There is a meaningful body of clinical work, some of it reporting substantial effects; there is much thinner ground when you ask the harder questions about mechanism and controlled comparison.

The clearest clinical literature sits in two areas. In autism-spectrum disorder, Coben and colleagues developed assessment-guided connectivity protocols on an explicit network-dysfunction rationale, and reported improvements across several uncontrolled and partially controlled studies; their review of the autism neurofeedback literature is a reasonable entry point for the state of that evidence, which is promising in pattern but limited by design. In traumatic brain injury, coherence-guided protocols have been reported — often in case series and clinic data — with sometimes striking improvements, on the rationale that TBI is a connectivity problem and connectivity training addresses it at the level of the disruption. Live z-score training has accumulated a case-study literature across mixed presentations, with Collura and colleagues' published case series among the more cited examples.

That is the encouraging side. Here is the discipline the brief for this entry asked me to hold, and that I would hold anyway: be careful to separate "this method was associated with clinical improvement" from "this method worked because it changed connectivity." Connectivity training is one of the areas where this conflation happens most readily, for a specific reason. To establish the mechanistic claim, you would need to show (a) that the trained metric actually moved in the trained direction, (b) that the movement reflected a real change in neural coordination rather than in an artifact-prone surface statistic, and (c) that the clinical improvement was attributable to that change rather than to the many non-specific ingredients every neurofeedback session contains — attention, expectancy, the therapeutic relationship, regression to the mean, maturation in developmental populations. Very few studies in this literature close all three gaps. Many do not close any of them. The clinical improvement can be entirely real and the connectivity story can still be unproven.

Two further cautions are specific to this method. First, the foundational measurement problems — volume conduction, reference dependence, the modest test-retest reliability of connectivity metrics relative to amplitude — sit underneath the entire literature. A finding built on naive zero-lag surface coherence carries an interpretive asterisk that a finding built on phase-corrected or source-space connectivity does not. The methodological literature on this is mature and worth reading directly: the canonical treatments of coherence, reference, and volume conduction, and the development of phase-lag measures expressly to reduce common-source bias, are not obscure — they are foundational, and they should inform how any practitioner reads a connectivity claim.

Second, the normative-database logic carries its own assumption. Training a metric toward the population mean presumes that "normal for the reference sample" equals "healthy for this individual." Usually a reasonable working assumption; not always a safe one. There are individuals whose atypical connectivity is compensatory, or simply their baseline, and for whom regression toward the database is not obviously the therapeutic goal. The database is a powerful tool and a strong prior. It is not a substitute for clinical reasoning about whether this deviation, in this person, is the thing to move.

The summary I would give a colleague: connectivity training has a genuine clinical literature, strongest in network-characterized conditions like TBI and ASD, weakened throughout by under-controlled designs and by foundational measurement problems that make the mechanistic claim hard to establish. It is a method where "it worked" and "we understand why it worked" are unusually far apart.

Strengths and weaknesses

Set out fairly, connectivity training has the following profile:

Strengths

• It targets something amplitude training cannot directly reach. If the clinical problem genuinely is a coordination problem — disrupted communication between regions rather than too much or too little activity at one of them — then connectivity training is, conceptually, the right tool rather than a workaround. For some TBI and some ASD formulations, that conceptual fit is the whole argument, and it is a good one.

• It is intrinsically assessment-driven and individualized. Done properly, it cannot be run from a generic menu — it requires a qEEG, a normative comparison, and a clinical hypothesis about which relationships matter. That discipline is congruent with the way I think neurofeedback should be practiced.

• Live z-score training can reduce single-metric tunnel vision. By letting the brain normalize whichever deviant metric it can reach, it avoids the trap of fixating on one number that may not be the operative one — and in skilled hands, that flexibility is a real asset.

• Multi-site training can capture distributed patterns that no single-site protocol would represent. When the formulation is genuinely about a network, training the network has a face-validity that training one node does not.

Weaknesses

• Volume conduction — the headline weakness, and the one that distinguishes this method from everything before it. Surface coherence can reflect one source feeding two electrodes rather than two regions communicating. Naive, zero-lag connectivity is especially suspect, and reinforcing it can train a statistical artifact. This is not a fringe caveat; it is foundational, and it is the first thing to interrogate about any surface-connectivity protocol.

• Reference and montage dependence. Connectivity values shift with the reference. Without an explicit, defensible montage choice, "training coherence" is underspecified.

• Hidden hardware timing errors. Connectivity assumes every channel was sampled at the same instant. Amplifiers that read channels sequentially rather than simultaneously — a multiplexed, cost-saving design — build a small fixed delay between channels that connectivity measures misread as brain timing. At best it adds error; at worst it fabricates a relationship or fakes the time-lag used to screen out volume conduction. A spec to verify before you buy ("simultaneous sampling"), not something the analysis can fix afterward.

• Indeterminacy at scale. The bipolar indeterminacy problem grows here. With many metrics trained at once, a large family of underlying states produces the same reward, and the practitioner frequently cannot say which relationship changed or whether the change was real.

• Modest reliability. Connectivity metrics are generally noisier and less test-retest reliable than amplitude measures. Training a noisy target risks over-fitting to qEEG noise — chasing deviations that would not reappear on a second recording.

• High interpretive demand. This method asks more qEEG literacy of the practitioner than any other in the series so far. Run without that literacy, it is not safer for being automated — it is more dangerous, because the automation hides the very assumptions the practitioner most needs to examine.

• Mechanism-versus-outcome conflation in the literature. The published base makes it easy to believe the mechanistic story is better established than it is. A practitioner reading uncritically will overestimate how well the field understands what connectivity training does.

• No clean felt correlate. As with bipolar — and more so — a coherence value has no accessible interoceptive handle. "Make these regions more coherent in alpha" is not a state a client can learn to recognize and re-enter. This tilts the method toward operant conditioning alone and leaves the second active ingredient, conscious and voluntary self-regulation, largely out of reach.

• Normative-database assumption. Training toward the reference mean assumes the mean is the right destination for this individual. Usually fine; not always; and the method does not flag the exceptions for you.

A note on the automation trade-off

It is worth naming directly, because the live z-score approach is where this method is growing. Automation that lowers the practitioner's moment-to-moment decision load is potentially useful when the practitioner understands what is being automated. The risk specific to connectivity work is that the automation sits on top of the field's least intuitive measurements — the ones most vulnerable to volume conduction, reference choice, and noise — and presents a clean normalizing z-score that looks like understanding. The cleaner the dashboard, the more important it is to know what the dashboard is computing. This is not an argument against live z-score training (I do have a bit of a list of those, however, we’ll get to those at a later date). It is an argument for never running it without being able to reconstruct, in principle, what the numbers on the screen actually represent.

Brendan's perspective

Three thoughts anchor this section. The first is about what connectivity training inherits from bipolar, and how the inheritance scales. The second is about volume conduction, which I think is genuinely under-discussed in clinical training relative to how much it should shape practice. The third is the clinical-honesty point: the gap between "it worked" and "connectivity changed," and how a working practitioner should hold that gap without either dismissing the method or overselling it.

The relational target, scaled up — and the indeterminacy that scales with it

In Part 3 I argued that a strict bipolar protocol asks the brain to learn changes in a quantity whose underlying physiology no one in the room can read — the same differential value is consistent with many different site configurations. Connectivity training is the same move at network scale, and the indeterminacy comes along for the ride, larger than before.

When you train a single, hypothesis-driven coherence relationship, the indeterminacy is bounded and at least examinable — you picked the relationship, you know what you are reinforcing, and you can watch the constituent signals. When you run live z-score training across many metrics, the indeterminacy expands to fill the space. The brain is being rewarded for normalizing some proportion of a large set of relationships, and the path it takes is, by design, not specified in advance. That is sold as a feature — let the brain choose — and in skilled hands it can be one. But "let the brain choose" is only a virtue if you can see what it chose. Most of the time, with surface metrics and many simultaneous targets, you cannot. You can confirm that the z-scores moved toward zero. You usually cannot say which underlying relationship did the moving, or whether it was a real network change or a shift in something the montage and the volume-conduction physics conspired to make look like one.

I am not against the marketing punchline of “handing optimization to the nervous system”; that is, in a sense, what all of neurofeedback does. I am against handing it over to an algorithm (no matter how “smart” it might think it is) and then narrating the result as if we know more than we do. The discipline I would ask for is modesty proportional to the number of simultaneous targets. The more metrics you train at once, the less you can claim about the mechanism, and the more your confidence should rest on the clinical outcome rather than the normalizing dashboard.

Volume conduction is the ghost in the machine

If I could put one idea into every connectivity-curious practitioner's head, it would be this: at the scalp, you cannot reliably tell a conversation between two regions from the echo of one region in two microphones.

That sentence is the whole problem. A single source — cortical or deeper — projects to multiple electrodes nearly instantaneously, producing high, stable, zero-lag coherence that represents exactly zero communication between two regions. Naive coherence cannot tell that apart from genuine functional connectivity. So a practitioner who reinforces high zero-lag surface coherence may be training the client to produce more of an artifact of skull physics. The feedback works — the number goes where you want it — and the thing the number represents may not be a network property at all.

This is not a reason to abandon the method. It is a reason to practice it with specific safeguards. Prefer measures built to resist common-source bias — phase-based metrics that suppress the zero-lag relationships most likely to be artifactual — over naive coherence whenever the system offers them. Be especially skeptical of any protocol whose headline finding is high zero-lag coherence between nearby sites, which is exactly where volume conduction is strongest. And where it matters clinically, corroborate surface findings in source space, which is precisely the upgrade the next entry in this series takes up. Source-localized connectivity does not make the problem vanish — the inverse problem brings its own caveats, and we will get to those — but it changes the terms of the measurement in a way that addresses the worst of the volume-conduction confound. The methodological literature on all of this is mature and not hidden. It is, in my opinion, part of the minimum competency for anyone training connectivity, and it is under-taught relative to its importance.

When "it worked" does not mean "connectivity changed"

Here is the clinical-honesty thread, and it is the one I care most about leaving you with.

Suppose you run a connectivity protocol with a client who has post-concussive difficulties, and the client improves. Genuinely improves — better sleep, steadier attention, fewer of the symptoms that brought them in. What can you conclude? You can conclude that the client improved during a course of treatment that included connectivity training. You cannot conclude, from that alone, that the connectivity training worked because it changed connectivity. The session contained a dozen other active ingredients — the structured attention, the expectancy, the therapeutic relationship, the natural recovery trajectory of a concussion, regression to the mean if they came in at their worst. To pin the result on the connectivity mechanism, you would need to show the metric moved as trained, that the movement was a real network change rather than an artifact, and that the change drove the outcome. In a single clinical case, you almost never have that.

This is not a counsel of despair. It is a counsel of accurate bookkeeping. I would rather a practitioner say "my client improved on a connectivity protocol, and I am not certain why" than "connectivity training repaired their network." The first statement is true and keeps the clinician honest about the next case. The second is a story we tell ourselves that quietly degrades our judgment over time. The field's enthusiasm for connectivity has, historically, outrun the field's ability to demonstrate its mechanism — and the cost of that overrun is paid in over-claiming, in mis-set expectations, and in protocols chosen because the network story sounds elegant rather than because the formulation demanded them.

So, the practitioner-workflow handles I would actually take into a session.

Sequencing and prerequisites. Do not run connectivity training without a real qEEG and the literacy to read it. If a practitioner cannot articulate which relationship they are training, on what montage, against what reference, and why that relationship matters for this client, the protocol is not ready to run. In most of my own decision-making, the methodological core — qEEG-informed amplitude work — comes first, and connectivity is reserved for formulations that are genuinely network-level and that amplitude work has not adequately addressed.

Number-of-targets management. Fewer targets, not more. The temptation of live z-score training is to throw many metrics at the problem and let the brain sort it out. I would push the other way — train the smallest set of relationships that the formulation justifies, so that you retain some ability to interpret what changed. Treat each added simultaneous target as a subtraction from your interpretive clarity, and spend those subtractions deliberately.

Learning-tracking, decoupled from the dashboard. Track the trained metric and the clinical outcome as two separate things, and never let the first stand in for the second. Watch for the failure mode where the z-scores normalize beautifully and the client does not change — that pattern is the method telling you the surface number and the clinical reality have come apart, and it should prompt a return to formulation, not a tightening of thresholds.

Montage and measure discipline. Prefer volume-conduction-resistant measures where the system offers them. Corroborate surface findings with referential power and, where it matters, with source-space estimates. Be explicit, every time, about the reference — because the connectivity numbers are not reference-neutral and your interpretation has to account for that. And know your amplifier: confirm that it samples all channels simultaneously, or that the software corrects for inter-channel timing, because a multiplexed amp can corrupt every connectivity number before the client does anything at all — a hardware problem no amount of careful analysis will undo.

Adaptive logic. If a client is not responding to a connectivity protocol, the right next move is often not a different connectivity target but a step back toward a more interpretable method — amplitude work at a well-chosen site, or a state-regulation foundation (HRV biofeedback, respiration) if the formulation suggests the autonomic substrate is in the way. Connectivity is not the top of a quality ladder that everything else is climbing toward. It is one specialized tool, and non-response is often a signal to return to the methods whose mechanisms you can actually read.

The crystallization I would leave you with: amplitude training asks how loud a region is; connectivity training asks whether two regions are talking — and the hardest part of the method is that, at the scalp, you cannot always tell talking from echo. Everything else in this entry is a consequence of taking that sentence seriously.

Would I do this method myself? In what context?

My honest answer here is more restrictive than the one I gave for the last few methods. I do not do connectivity training in its standard, surface form — not coherence training, not live z-score connectivity. Everything this article has laid out — volume conduction, reference dependence, the indeterminacy that grows with every added target, the hidden hardware skew, and the absence of any felt correlate the client can learn from — adds up, for me, to a surface measure I cannot read cleanly enough to trust as a primary clinical tool. When I cannot tell talking from echo, I would rather not build a protocol on the distinction.

That does not mean I ignore the relational question the method is trying to answer. It means I answer it, where I can, with tools I can actually read.

The thing I do reach for is a two-channel, same-band amplitude protocol with a conjunctive threshold: two sites, the same frequency band at each, and feedback that fires only when both channels cross their threshold at the same moment. The client is rewarded for getting both sites "up" in the band together — not for a coherence value computed between them. This is, I will happily admit, a derivative of connectivity thinking. It nudges toward co-activation of two regions, which is a crude, indirect cousin of what connectivity training is after. But it is not a connectivity measure, and that is exactly the point. It is amplitude logic with an AND gate. I keep everything that makes amplitude training interpretable — I can see precisely what each site is doing, each site has a felt correlate the client can learn to recognize, and no volume-conduction artifact or inter-channel timing skew can masquerade as a relationship, because I am never computing a relationship in the first place. It is connectivity-flavoured work that stays inside the part of the field whose mechanisms I can read. For the co-activation questions I actually meet in practice, it does most of what I would have wanted surface connectivity to do, without the fragility.

When I do reach for genuine connectivity training — training an actual relationship between sites rather than a conjunction of amplitudes — it is on the rare occasions I would already be reaching for LORETA, in source space. That is not a coincidence. Source-localized connectivity estimates the relationship between cortical generators rather than between smeared scalp electrodes, which is the one setting where the volume-conduction objection is meaningfully addressed rather than merely acknowledged. Even there it is rare, always qEEG-anchored, and reserved for formulations where a source-level network target is doing real explanatory work. Surface connectivity, trained as a primary protocol, essentially never.

What I would say to a colleague is the same thing I say in NeuroLogic's trainings. If the formulation calls for a relational target, ask first whether a two-channel conjunctive-amplitude protocol can answer it — it usually can, and it keeps you in interpretable territory. Watch what happens with connectivity. Reserve true connectivity training for source space, for the rare case that genuinely needs it, and only with the qEEG literacy to read what you are doing. And whatever you train, learn what volume conduction and inter-channel skew do to a surface connectivity number, so that the next beautiful z-score someone shows you gets the skepticism it deserves. The discipline that runs through every entry in this series applies here with the least margin for error — the method has the most ways to go wrong, and its failures are the quietest and best-dressed in the field.

The next post — Part 5 — LORETA and source-localized neurofeedback — takes up the upgrade this entry kept gesturing toward. If volume conduction is the ghost in the surface-connectivity machine, source localization is the field's most serious attempt to exorcise it: estimating activity in source space rather than at the scalp, so that the relationships we train are between estimated cortical generators rather than between smeared electrode signals. It does not come for free — the inverse problem replaces one set of caveats with another, and we will be honest about that — but it changes the terms of the measurement in exactly the place this entry found them weakest. The reasoning that earns connectivity training a narrow, principled place is the reasoning the next entry will carry into source space.

References

• Bastos, A. M., & Schoffelen, J.-M. (2016). A tutorial review of functional connectivity analysis methods and their interpretational pitfalls. Frontiers in Systems Neuroscience, 9, 175.

• Coben, R., Linden, M., & Myers, T. E. (2010). Neurofeedback for autistic spectrum disorder: A review of the literature. Applied Psychophysiology and Biofeedback, 35(1), 83–105.

• Coben, R., & Padolsky, I. (2007). Assessment-guided neurofeedback for autistic spectrum disorder. Journal of Neurotherapy, 11(1), 5–23.

• Collura, T. F., Guan, J., Tarrant, J., Bailey, J., & Starr, F. (2010). EEG biofeedback case studies using live Z-score training and a normative database. Journal of Neurotherapy, 14(1), 22–46.

• Nunez, P. L., Srinivasan, R., Westdorp, A. F., Wijesinghe, R. S., Tucker, D. M., Silberstein, R. B., & Cadusch, P. J. (1997). EEG coherency I: Statistics, reference electrode, volume conduction, Laplacians, cortical imaging, and interpretation at multiple scales. Electroencephalography and Clinical Neurophysiology, 103(5), 499–515.

• Stam, C. J., Nolte, G., & Daffertshofer, A. (2007). Phase lag index: Assessment of functional connectivity from multi channel EEG and MEG with diminished bias from common sources. Human Brain Mapping, 28(11), 1178–1193.

• Thatcher, R. W. (2012). Coherence, phase differences, phase shift, and phase lock in EEG/ERP analyses. Developmental Neuropsychology, 37(6), 476–496.

• Thatcher, R. W., Walker, R. A., & Giudice, S. (1987). Human cerebral hemispheres develop at different rates and ages. Science, 236(4805), 1110–1113.