Real-Time Source Imaging with hdEEG

This paper by Guarnieri and colleagues offers a thoughtful methodological step for anyone interested in where neurofeedback may be heading next. Rather than asking whether scalp EEG can be processed in real time, the authors asked a more ambitious question: can high-density EEG be reconstructed online in source space, with enough fidelity to approximate the kinds of analyses we usually trust only after offline processing? Their answer, at least in this proof-of-concept study, is cautiously encouraging.

Because this article was published more than a year ago, it fits squarely in the From the archives category. Even so, it remains highly relevant. One of the enduring tensions in EEG-based neurofeedback is the tradeoff between immediacy and anatomical specificity. Conventional neurofeedback is often delivered at the sensor level, which is practical and fast but can blur together multiple underlying neural generators. Neurofeedback refers to training in which individuals receive real-time information about brain activity and learn, through practice, to shift that activity in a desired direction. Biofeedback uses the same general learning principle for bodily signals such as heart rate, respiration, skin conductance, or muscle activity. This paper sits at the intersection of those ideas because it uses sophisticated signal processing to move EEG feedback closer to specific neural sources while still preserving real-time operation.

That matters. If source-space reconstruction can be made sufficiently fast and reliable, clinicians and researchers may be able to design feedback protocols around functionally meaningful cortical targets rather than relying only on electrode-level approximations. The present study does not yet demonstrate a therapeutic intervention, but it does address a major bottleneck: whether individualized, high-density source reconstruction can happen online with low enough latency to support closed-loop applications, including source-based neurofeedback.

Methods

RT-NET is a MATLAB-based toolbox designed to guide users through real-time source reconstruction with a graphical interface rather than a coding workflow. The pipeline is organized into five modules: study overview, head segmentation, leadfield matrix creation, spatial filter creation, and real-time activity estimation. In practical terms, the workflow begins with a participant-specific T1-weighted MRI, followed by electrode-position digitization, a calibration EEG recording, and then the real-time session itself.

For validation, the authors tested the toolbox in 10 healthy right-handed adults aged 23 to 39 years. Each participant first underwent structural MRI. In a separate EEG session, the authors digitized electrode positions and recorded two high-density EEG datasets using a 128-channel actiCHamp system sampled at 1 kHz. The first recording consisted of 4 minutes of eyes-open fixation and served as the calibration dataset. The second recording lasted 6 minutes and involved repeated right wrist flexion-extension, alternating 6 seconds of self-paced movement with 6 seconds of fixation. Horizontal and vertical EOG were recorded, along with multiple EMG channels. Some EMG channels were used to support artifact detection, while a forearm EMG signal was used to identify movement onset.

The nonstandard methodological contribution is the way RT-NET front-loads the computationally heavy steps. The participant’s MRI was segmented into brain, skull, and skin compartments using SPM12, and a realistic three-layer head model was created. Electrode positions were co-registered to the MRI space, and the leadfield matrix was generated using a symmetric boundary element method implemented in OpenMEEG. During calibration, RT-NET identified bad channels, applied average re-referencing, filtered the data in a selected frequency band, and ran independent component analysis. Artifactual components were identified using auxiliary physiological channels and spectral/statistical criteria, allowing the system to build an artifact attenuation filter.

Source localization was then implemented with eLORETA by default, although other inverse methods were available. Crucially, the resulting spatial filters were applied to incoming EEG data online. For real-time tests, the authors used a 500 ms buffer and quantified motor-related event-related desynchronization in the beta band from 13 to 30 Hz, comparing RT-NET output against a conventional offline source-analysis workflow.

Results

The headline result is feasibility with surprisingly low online delay. Average processing times for the preparatory modules were 1938 seconds for head segmentation, 302 seconds for leadfield matrix creation, and 735 seconds for spatial filter creation. In other words, the MRI segmentation remained the slowest step by far, while leadfield estimation and filter creation were short enough to be incorporated into the same experimental session. During real-time acquisition, using 500 ms data windows, the maximum delay introduced by artifact attenuation and source localization was 4 ms per buffer.

That speed would not matter much if the reconstructed neural activity looked unconvincing, so the authors compared RT-NET with their offline pipeline. In single participants, the real-time beta-band event-related desynchronization maps during right-hand movement consistently peaked in the expected motor regions. Across trials, the average spatial correlation of online ERD maps was 0.78. When online and offline processing were compared directly, the difference in across-trial ERD-map correlations did not reach statistical significance in the reported Wilcoxon signed-rank test (p = 0.06), suggesting reasonably similar stability.

The group-level comparison was also encouraging. The correlation between the beta-band ERD maps derived from RT-NET and from the offline workflow with artifact removal was 0.76, whereas the correlation between RT-NET and data processed without artifact removal was lower at 0.56. That pattern is important because it suggests that the online artifact attenuation was not merely fast, but meaningfully improved the physiological plausibility of the reconstructed maps.

Region-of-interest analyses told a similar story. The expected motor-related desynchronization was visible in left primary motor cortex, supplementary motor area, and left ventral premotor cortex, whereas the control region in left superior temporal gyrus showed little task-related modulation. Residual contamination from EOG and EMG appeared minimal after processing, with absolute correlations close to zero and no significant online-versus-offline differences for EOG (p = 0.5542) or EMG (p = 0.1923). The temporal correspondence between online and offline power modulations was stronger in the beta band than in the full 1 to 50 Hz band for left M1 (p = 0.0078), SMA (p = 0.0156), and left VPMC (p = 0.0078), but not for the control ROI in left STG (p = 0.061).

Results like these do not prove equivalence between online and offline source reconstruction, but they do suggest that RT-NET captured the expected motor physiology with useful fidelity.

Discussion

What this paper shows, above all, is that individualized source-space hdEEG processing can be brought into real time without collapsing under its own computational weight. That is not a trivial achievement. For years, the promise of EEG source imaging has been tempered by a practical reality: once you add MRI-based head modeling, dense montages, realistic forward models, artifact correction, and inverse solutions, the analysis tends to migrate firmly into the offline domain. RT-NET attempts to solve that bottleneck by estimating the expensive parts up front and then applying adaptive spatial filters to the incoming data stream. In this validation study, that strategy worked well enough to reproduce the expected beta-band motor signatures with low latency and reasonable agreement with offline results.

For neurofeedback, the conceptual importance is easy to see. Sensor-level EEG feedback is often clinically useful, but it remains an indirect proxy for cortical generators. A training target at C3, for example, is never only “left motor cortex.” It reflects a weighted mixture of neural and non-neural sources shaped by volume conduction, montage geometry, referencing choices, and artifact burden. Source-space reconstruction does not eliminate those problems, but it may reduce some of the ambiguity by anchoring feedback to anatomically constrained estimates. In that sense, this paper strengthens the technical plausibility of source-based neurofeedback.

Still, the study should not be overread. The validation sample was small, healthy, and limited to a robust motor task that predictably produces contralateral beta desynchronization. This is exactly the kind of paradigm one would choose for a first-pass validation because the signal is strong, well characterized, and spatially interpretable. That makes the results useful, but it also means they do not automatically generalize to the kinds of source targets that are often clinically attractive and methodologically harder: distributed affective networks, default-mode dynamics, frontal midline control systems, or low-amplitude oscillatory processes during resting states. In those contexts, the signal-to-noise ratio may be worse, the sources may be more bilateral or more diffuse, and the expected online-offline agreement may be lower.

The study also does not test neurofeedback learning itself. No participant was asked to use the reconstructed source activity as feedback, no operant-learning outcome was measured, and no comparison was made against standard sensor-level protocols. So the paper supports the feasibility of source reconstruction for possible neurofeedback applications, but it does not establish that source-based neurofeedback is more effective, more efficient, or more clinically meaningful than existing approaches. That distinction matters.

There are also practical barriers. The workflow depends on a structural MRI, electrode-position digitization, high-density EEG, calibration data, and substantial preprocessing infrastructure. For a research lab, that is feasible. For many clinics, it is still a heavy lift. The authors were appropriately transparent about this. Their three-layer boundary element head model was chosen as a compromise between realism and speed, and they explicitly note that richer models might improve localization but remain computationally impractical in real-time use. They also point out that artifact attenuation depends on an adequate calibration dataset; if the calibration period does not sufficiently sample ocular or myogenic artifacts, online cleaning may be less robust than the present results suggest.

Even so, the paper has real translational value. For referring professionals and scientifically minded clients, it offers a clear example of where the field is maturing: away from simplistic “brainwave training” narratives and toward individualized, model-based signal interpretation. For neurofeedback professionals, it highlights a useful design principle. The future of EEG feedback may not lie in ever more complicated reward screens or broader frequency menus, but in better target definition, better artifact control, and better alignment between the physiology being trained and the clinical question being asked.

A final interpretive point is worth emphasizing. This paper is not best understood as evidence that hdEEG will replace fMRI for all functional targeting, nor as proof that real-time source imaging is ready for routine therapeutic deployment. Rather, it is a demonstration that a longstanding methodological gap can be narrowed: high-density EEG can retain temporal immediacy while moving closer to anatomically informed source estimates. That is a meaningful advance. Whether it ultimately changes clinical outcomes will depend on subsequent work showing that better source specificity leads to better learning, better transfer, or better symptom change. This study makes that next generation of work more possible.

Brendan’s perspective

What I like about this paper is that it nudges us to think a little more precisely without becoming less practical. In clinic, most of us are not sitting with a 128-channel hdEEG cap, a structural MRI for every client, and a real-time source-imaging pipeline running in the next room. But that does not make this work irrelevant. Quite the opposite. It gives us a cleaner conceptual map for how to think about targets, individualization, and restraint. And frankly, those three things matter more than technological glamour.

From scalp sites to source-space goals

One of the most useful takeaways here is not that every clinician should suddenly abandon sensor-level training and move to source imaging. It is that we should become more honest about what scalp sites do and do not represent. When we train at C3, C4, Cz, FCz, Pz, or Fz, we are not training a neat little cortical postage stamp. We are sampling a mixture of underlying generators filtered through volume conduction, skull properties, reference choices, artifact burden, and state-dependent variability. Clinicians already know this intuitively. This paper simply gives the intuition better engineering.

So how would I apply that in practice? I would treat source-space thinking as a way of improving target selection, even when I am still delivering feedback from conventional scalp electrodes. If a client presents with motor inhibition problems, restlessness, poor timing, or performance inconsistency, the lesson is not merely “train C3” or “train central beta.” The better question is: which functional network am I really trying to influence, and is my chosen electrode configuration a reasonable proxy for that network? For some clients, sensorimotor rhythm training around C3/C4 or Cz may still be clinically useful. For others, especially when symptoms are mixed or bilateral, a simple scalp-site rule can become too blunt.

This matters most in the gray-zone cases. The easy cases are where symptoms, history, and physiology point in the same direction. The harder cases are where the qEEG suggests diffuse slowing, the client describes racing thoughts, sleep disruption, and sensory sensitivity, and the training response to a textbook protocol is mediocre. In those cases, source-informed thinking can save us from overcommitting to a single electrode narrative. Even without real-time source imaging, the clinician can ask more sophisticated questions: am I chasing local rhythm abnormalities, unstable state regulation, poor inhibitory control, or network-level inefficiency? That shift alone can improve protocol reasoning.

Protocol individualization is the real message

If I had to reduce this paper to one clinically relevant theme, it would be this: precision is not mainly about better machines. It is about better matching between the physiology you train and the person sitting in front of you. RT-NET is impressive because it individualizes the forward model, the artifact attenuation, and the source estimate. In other words, it does not assume that one participant’s anatomy or signal characteristics can stand in for another’s. That should sound very familiar to experienced neurofeedback clinicians, because our best work already depends on the same principle.

This is where I think the paper has real relevance for day-to-day EEG practice. Even if you never implement source-space neurofeedback, it reminds you not to confuse a protocol template with a formulation. Two clients with attentional problems may not need the same frequency target, the same reward band, the same inhibit strategy, or the same site selection. One may respond beautifully to SMR-oriented work over central regions, especially if sleep instability and motor overflow are prominent. Another may need a more cautious approach to fast-frequency work because hyperarousal, muscle tension, and anxiety are already driving the picture. Another may not tolerate central training very well at all until autonomic regulation and sleep are steadier.

That is also why I would rarely use neurofeedback in isolation when cases are complex. Source-informed thinking pairs naturally with multimodal regulation work. If a client is physiologically brittle, I often think in layers. Neurofeedback may help shape cortical stability, but respiration training and HRV biofeedback can improve the autonomic terrain in which that learning happens. Slow, comfortable breathing practice can reduce sympathetic overcoupling, improve interoceptive awareness, and often make neurofeedback sessions more stable and more tolerable. In trauma-spectrum or chronically dysregulated clients, that matters. Sometimes the issue is not that the brain cannot learn. It is that the whole organism is too mobilized, too defended, or too erratic for efficient learning to consolidate.

Psychotherapy belongs in this conversation too. A more specific neural target does not remove the meaning of symptoms, the person’s story, or the behavioral contingencies maintaining distress. In fact, when neurofeedback is integrated well with psychotherapy, the combination can be stronger than either alone. Better physiological regulation may widen the client’s window for reflective work, while psychotherapy helps the client use that increased flexibility in real life. That is much more interesting to me than treating source imaging as a standalone miracle.

A clinician’s reality check

Now for the part I think the field always needs to hear: elegant methods do not automatically become elegant care. Research pipelines are cleaner than clinics. Participants are selected, tasks are controlled, movement paradigms are simple, and noise is managed with unusual discipline. Real clients show up tired, activated, under-slept, over-caffeinated, hormonally variable, masked, guarded, skeptical, hopeful, and sometimes carrying three or four diagnostic stories that do not reduce nicely to one oscillatory target. That is the real-world laboratory.

So I am enthusiastic about papers like this, but I do not think enthusiasm should become overclaiming. This study tells us that real-time source reconstruction is feasible in a motor validation paradigm. It does not tell us that source-based neurofeedback is categorically superior to well-delivered sensor-level neurofeedback. It does not tell us that every clinic should chase MRI-based pipelines. And it certainly does not tell us that better localization automatically produces better outcomes. Those are separate questions.

Still, the paper gives clinicians something valuable: permission to raise the bar on conceptual specificity. Even in ordinary practice, we can be more careful about artifacts, more explicit about why we chose a target, more willing to revise a protocol that is not producing the expected response, and more humble about the gap between signal interpretation and clinical truth. That, to me, is where the paper becomes genuinely useful.

My practical takeaway is simple. Keep the ambition of source-based thinking, even when using ordinary equipment. Think in networks rather than only electrodes. Individualize frequencies, sites, and pacing rather than forcing clients into preset bins. Use complementary regulation tools when the autonomic system is clearly part of the bottleneck. And stay clinically honest. A sophisticated map is wonderful, but only if it helps the person in the chair become more stable, more flexible, and more able to live their life. That is still the standard that matters most.

Conclusion

Guarnieri and colleagues present a careful and technically ambitious proof-of-concept showing that real-time source reconstruction from high-density EEG is feasible, low-latency, and meaningfully comparable to a conventional offline workflow in a simple motor task. That alone makes the paper worth revisiting. It addresses one of the field’s more stubborn constraints: how to preserve the real-time advantages of EEG while moving closer to anatomically interpretable neural targets.

For neurofeedback, the implications are promising but still preliminary. The study supports the idea that source-based feedback is technically plausible and may eventually help refine target selection beyond scalp-level approximations. At the same time, it stops well short of demonstrating clinical superiority, broad generalizability, or ease of implementation in routine care. Those questions remain open.

Even with those caveats, this is the kind of archive paper that still feels alive. It captures a transition point in the field, where computational neuroscience, signal processing, and neurofeedback begin to speak the same language. That is good news for clinicians and researchers alike. Better physiological specificity does not solve every clinical problem, but it gives us a better starting point for asking smarter ones.

References

Guarnieri, R., Zhao, M., Taberna, G. A., Ganzetti, M., Swinnen, S. P., & Mantini, D. (2021). RT-NET: Real-time reconstruction of neural activity using high-density electroencephalography. Neuroinformatics, 19, 251-266. https://doi.org/10.1007/s12021-020-09479-3