Augmented Reality Neurofeedback

• Augmented Reality Neurofeedback is a spectrum of closed-loop systems that acquire neural signals and deliver spatially registered feedback to support self-regulation and command execution.
• It leverages methodologies such as SSVEP-based detection, motor imagery decoding, and AR-guided neuronavigation to translate brain activity into meaningful, context-embedded feedback.
• Applications include assistive control, pediatric intervention, neurorehabilitation, and non-invasive brain stimulation, with promising empirical results demonstrating improved precision and user engagement.

Augmented reality (AR) neurofeedback is a class of closed-loop systems in which a measured variable related to neural or neurotechnology state is mapped onto spatially registered feedback in an augmented environment. Across the current literature, the term spans at least three distinct but related categories: direct neurofeedback from ongoing brain activity; AR-mediated brain-computer interface (BCI) control paradigms that provide implicit performance-linked feedback rather than explicit self-regulation training; and adjacent AR-guided neurotechnology systems, such as neuronavigation for transcranial magnetic stimulation (TMS), that provide real-time operator-facing feedback without measuring brain state. The distinction is technically important. Some studies implement classical neurofeedback targets such as mu-suppression-derived social attention, some implement motor imagery (MI)-driven AR feedback that stabilizes BCI control, and some demonstrate enabling infrastructure for real-time AR-based feedback loops without constituting neurofeedback in the narrow clinical sense (Lyu et al., 6 Mar 2025).

1. Conceptual scope and definitional boundaries

AR neurofeedback is not a single paradigm but a spectrum of closed-loop designs. In the narrowest sense, it refers to systems that acquire neural signals, estimate a target state in real time, and present that state back to the user through AR so that the user can learn self-regulation. "Eggly" exemplifies this formulation: a consumer EEG headband provides a social attention index derived from mu suppression, and that index continuously and discretely drives events in a mobile tabletop AR game for autistic children (Lyu et al., 6 Mar 2025). The MI-based AR-robot system is also neurofeedback in a substantive sense because AR is used as a control-stabilization mechanism between EEG decoding and robot commitment, with direction-congruent feedback over the task space rather than an abstract monitor (Wang et al., 25 Sep 2025).

A broader but still neurofeedback-relevant category consists of AR-BCI systems in which the user does not explicitly learn to regulate an endogenous latent state for its own sake, but where brain activity is embedded into a meaningful AR interaction loop. The 2017 immersive SSVEP navigation study is explicit on this point: it is best understood as an early feasibility study of an AR-enabled SSVEP BCI rather than a neurofeedback study in the classical self-regulation sense. The user attends to flickering targets, the system detects stimulus-specific steady-state visual evoked potentials, and the resulting class is mapped to navigation commands. The feedback is therefore task-based and action-based rather than an explicit neural score, yet the architecture still operationalizes a closed-loop brain-to-environment system in immersive AR (Faller et al., 2017).

A third category is adjacent AR-guided neurotechnology. The two neuronavigation papers on non-invasive brain stimulation describe real-time tracking of patient and TMS instrumentation, a Unity digital twin, and AR projection of a brain model or estimated stimulation point onto the patient scene. These systems provide clinician-facing spatial guidance rather than participant-facing neural self-regulation. They are therefore not direct neurofeedback, but they show how AR can visualize hidden internal state in situ and support closed guidance loops around neurotechnology workflows (Hu et al., 23 Jan 2026).

This taxonomy matters because the same phrase, “AR neurofeedback,” can otherwise collapse distinct technical questions: whether the controlled variable is endogenous or exogenous, whether the feedback is explicit or implicit, whether the user is participant or operator, and whether the loop is designed for neural self-regulation, command selection, or stimulation guidance.

2. Signal modalities, decoding strategies, and closed-loop architectures

The literature includes two primary neural modalities for AR-linked feedback: steady-state visually evoked potentials and motor imagery EEG. SSVEP-based systems dominate early AR-BCI work because they provide high bandwidth and require only minimal calibration and training. In the immersive AR/VR feasibility study, EEG was acquired with three sintered AgCl electrodes in a bipolar setup around $O1$, with one electrode 2.5 cm anterior to $O1$, one 2.5 cm posterior to $O1$, ground at $Fpz$, impedances below $5~k\Omega$, band-pass filtering from 0.5 to 100 Hz, notch filtering at 50 Hz, and sampling at 256 Hz. Classification used the harmonic sum decision method. One class was selected “as soon as the sum of the responses at the base frequency along with the second and third harmonic components exceeds the sums for all other responses throughout a dwell-time of 1 s for the navigation stimuli and 1.5 s for the on/off stimulus,” followed by a 3-second refractory period (Faller et al., 2017).

A faithful generic reconstruction of that detector is

$S_i = R(f_i) + R(2f_i) + R(3f_i),$

with class $i$ selected if

$S_i > S_j \quad \forall j \neq i$

throughout the dwell-time interval. The paper does not print this notation exactly, but it accurately reflects the described algorithm. The command mapping was fixed: $12$ Hz for turn $45^\circ$ left, $O1$0 Hz for turn $O1$1 right, $O1$2 Hz for step forward, and $O1$3 Hz for toggling navigation stimuli or detectors on or off (Faller et al., 2017).

The 2023 HoloLens framework retains the SSVEP basis but shifts to a more modern split architecture in which Unity on the headset displays flickering buttons and feedback, while EEG preprocessing and classification run externally in Python. The hardware comprises a first-generation Microsoft HoloLens and an Emotiv Epoc+ headset with 14 saline electrodes at $O1$4 and $O1$5, sampled at 257 Hz. Three commands were implemented in AR: Create Cube at 12 Hz, Create Sphere at 8.57 Hz, and Delete All at 10 Hz. Each trial contained 5 seconds of flicker followed by 5 seconds of rest, and recorded trials therefore contained 1285 samples per channel (Mustafa et al., 2023).

Its adaptive ensemble recognizer addresses inter-subject variability through eight models total: Support Vector Machine and Random Forest, each trained with four preprocessing variants—CAR + PCA, CAR only, PCA only, and neither CAR nor PCA. The paper gives the CAR equation exactly as

$O1$6

After CAR, FFT is applied, features are selected from a 1 Hz window centered on each target frequency and harmonic, PCA is used to reduce dimensionality and runtime, and Z-score normalization is applied. Online classification then uses weighted voting in which the vote for the predicted label from a given model is multiplied by that model’s training accuracy (Mustafa et al., 2023).

MI-based AR neurofeedback differs materially from SSVEP paradigms because the neural variable is not visually evoked by flicker. In the EEG-driven AR-robot system, an Emotiv EPOC X provides 14-channel EEG at 128 Hz. Signals are band-pass filtered between 8–16 Hz to isolate $O1$7 and low $O1$8 rhythms, features are extracted as bandpower estimates, and a linear classifier implemented in EmotivBCI discriminates Left vs Neutral, Right vs Neutral, and Lift vs Neutral. The output is not committed directly to the robot. Instead, it first drives AR target navigation and confirmation, which acts as an intermediate stabilizing layer (Wang et al., 25 Sep 2025).

The pediatric mobile AR neurofeedback game uses a different arrangement. The front-end is an iPadOS AR application built with Unity, ARKit XR Plugin, and AR Foundation; a BrainCo FocusCalm headband provides five channels of raw EEG signals and computes a vendor-derived social attention index interpreted from mu suppression. That index ranges from 0 to 100, where 0 indicates complete loss of focus on observation or no mu suppression, and 100 indicates full focus or significant mu suppression. Communication between the Python back-end server and front-end is bi-directional TCP. The neurofeedback variable therefore enters the AR loop already as a scalar metric rather than as a researcher-defined in-paper decoder (Lyu et al., 6 Mar 2025).

These architectures share a recurrent systems pattern: sensing, state estimation, transmission to an AR rendering layer, and immediate feedback embedded in the user’s environment. What changes across the literature is the source of the state variable, the timescale of estimation, and whether the loop is meant to train self-regulation, stabilize sparse control, or support task execution.

3. AR feedback design and interaction primitives

A central design axis in AR neurofeedback is whether feedback is detached and symbolic or embedded in meaningful scene elements. The literature repeatedly favors embedding. The immersive SSVEP study superimposed flickering targets on a real-world video scene through a head-mounted display and camera; the user’s attended target produced an SSVEP over occipital cortex, and the resulting class was mapped to navigation commands for an avatar in the augmented scene. In the AR scenario, the 3D graphics were tracked to a fiducial marker, so the user perceived the physical environment through camera video with superimposed computer-generated objects, including the avatar, waypoints, and flickering command stimuli. This is video-see-through AR rather than optical see-through, a choice that matters because visibility and contrast of flickering stimuli can be better controlled experimentally (Faller et al., 2017).

The interaction design of that system included an explicit no-control/control distinction. Subjects first remained in a 30-second no-control period, then activated navigation stimuli, navigated to a second gray waypoint, deactivated the stimuli, and remained in a final 30-second no-control period. The on/off toggle in the upper right allowed the system to distinguish no-control from control periods more robustly by enabling or disabling navigation commands. This explicit mode switching is an important practical feature because it reduces unintended activations in naturalistic scenes (Faller et al., 2017).

The HoloLens SSVEP framework uses virtual buttons rather than avatar-centered navigation. The architecture cleanly separates AR rendering from BCI processing: Unity on the HoloLens displays flickering buttons and later displays feedback; EEG recording and classification run externally in Python; classification results return via a Python-Unity socket interface; and Unity triggers the corresponding AR event. Selecting one of the flickering buttons causes Unity to instantiate or delete virtual objects at the HoloLens cursor location (Mustafa et al., 2023). This suggests a practical implementation path for early AR neurofeedback systems: host-side decoding with headset-side visualization.

In MI-based AR neurofeedback, the design emphasis shifts from frequency-tagged visual salience to direction-congruent embodied cues. The smartphone-based Unity video-see-through AR interface anchors a virtual interactive block and arrow indicator to each fiducial-tagged physical object. During Left or Right MI, the entire AR block together with its arrow produces a small but visible sway toward the neighboring target in the decoded direction; during Lift MI, the block locks onto the current target object and finalizes the decision. The study explicitly compares four conditions—No AR block, Static, Sham, and Neurofeedback—thereby isolating AR itself, motion per se, and correctly contingent visuomotor feedback (Wang et al., 25 Sep 2025).

The pediatric social-attention game uses scene-integrated feedback rather than explicit dashboards. The child places a tablet on a desk, taps to anchor the farm scene, and watches a social collaboration sequence involving two human-like characters, a bird, eggs, trays, and reward effects. The clearest continuous mapping is that bird height directly tracks the social attention index. The game also uses discrete stage mapping through adaptive thresholds

$O1$9

with $O1$0, $O1$1, $O1$2, and $O1$3. These thresholds divide performance into low, medium, and high stages, which control character movement speed, egg-laying speed, facial expressions, music tempo, and special rewards such as heart-shaped bubbles if high stage is maintained for more than 3 seconds and golden eggs when performance is non-decreasing for more than 3 seconds (Lyu et al., 6 Mar 2025).

A related but non-neural design appears in upper-limb rehabilitation AR feedback. There, HoloLens 2 overlays personalized 3D tunnels onto the real environment, and color-coded real-time feedback indicates how accurately a user follows therapist-defined trajectories. The feedback is kinematic rather than neural, but the design principle is closely related: the relevant error variable is embedded in spatially registered AR rather than detached monitor widgets (Luciani et al., 2024).

Taken together, these systems converge on several recurrent primitives: spatial registration to task objects or workspace, scene-embedded rather than peripheral feedback, explicit state transitions or confirmation stages, and temporally extended feedback that supports persistence rather than isolated event signaling.

4. Application domains

The present AR neurofeedback landscape is heterogeneous, with applications spanning assistive control, robotics, pediatric intervention, rehabilitation, and brain-stimulation workflows.

In assistive and hands-free control, the immersive SSVEP navigation study motivates AR as a friendlier and more intuitive real-world user interface that may compensate for low BCI bandwidth by making interaction more direct and contextually anchored in the physical world. The authors explicitly mention smart home control and user groups requiring hands-free operation due to temporary situational disability (Faller et al., 2017). The HoloLens SSVEP framework broadens this orientation toward mainstream AR-BCI applications for healthy users by replacing preexisting voice commands in an open-source HoloLens Unity project with EEG-based command selection (Mustafa et al., 2023).

In human-robot interaction, AR neurofeedback is used as an intermediate decision layer rather than a direct actuation channel. The zero-touch grasping system integrates MI-based EEG decoding, smartphone AR target selection, and autonomous robotic grasping with a MyCobot 280Pi and eye-in-hand RGB camera. The hand–eye relation is expressed as

$O1$4

where $O1$5, $O1$6, $O1$7, and $O1$8 denote base, end-effector, camera, and object frames. The AR layer absorbs variability in EEG control by requiring stabilized navigation and explicit confirmation before robot execution (Wang et al., 25 Sep 2025).

In pediatric neurofeedback, Eggly targets preschool children with autism spectrum disorder, approximately 30 months to 6 years old, and focuses on social attention and social-communicative functioning. The child is not asked to tap or manipulate controls during play; instead, the intended behavior is to maintain social attention on the observed collaboration so that the EEG-derived social attention index drives in-scene game events. The paper positions mobile AR as a way to lower access barriers, support personalization, and situate neurofeedback in familiar non-clinical settings (Lyu et al., 6 Mar 2025).

In neurorehabilitation, the HoloLens 2 upper-limb study is not direct neurofeedback, but it occupies an important neighboring region of the design space. It provides therapist-customizable, real-time 3D kinematic feedback for precision retraining in exoskeleton-based rehabilitation. This suggests that AR-mediated rehabilitation feedback spans a continuum from neural, to kinematic, to task-performance variables, all sharing a common concern with embodied, spatially precise feedback (Luciani et al., 2024).

In non-invasive brain stimulation, the neuronavigation papers demonstrate AR-guided spatial feedback for TMS. One implementation uses three CANYON CNE-CWC5 web cameras, AprilTag fiducials, a Unity digital twin, and an Android smartphone or AR-capable device to overlay a dynamic brain model and estimated stimulation target. These systems are operator-facing rather than participant-facing and do not estimate physiology in real time, but they contribute enabling infrastructure—registration, spatial alignment, digital twin updating, and in-situ visualization—that is directly relevant to future multimodal AR neurofeedback and adaptive neurostimulation platforms (Hu et al., 28 Jan 2026).

5. Empirical evidence and reported performance

The empirical literature remains small, methodologically diverse, and unevenly distributed between direct neurofeedback and adjacent AR-BCI work. Nonetheless, several quantitative results define the current technical baseline.

The immersive SSVEP feasibility study involved three healthy male participants aged 26–27, with two SSVEP-experienced and one BCI-naive. Two runs were performed in VR and two in AR for each participant. The outcome measures were intentional task-conform interactions in control state as $O1$9 per minute, unintentional interactions in control state as $Fpz$0 per minute, unintentional interactions in no-control state as $Fpz$1 per minute, and positive predictive value

$Fpz$2

For AR, the mean values were $Fpz$3, $Fpz$4, $Fpz$5, and $Fpz$6. For VR, the mean values were $Fpz$7, $Fpz$8, $Fpz$9, and $5~k\Omega$0. Two experienced subjects finished all runs successfully in both conditions, while the naive participant achieved moderate control only in the first run of the VR condition and evidently failed in AR (Faller et al., 2017).

The HoloLens SSVEP framework used six volunteers aged 18–26, with separate PC and HoloLens groups. The paper reports the information transfer rate formulas

$5~k\Omega$1

and

$5~k\Omega$2

On the PC, offline experiments using all channels yielded a mean accuracy of 80%. In online PC ensemble experiments using $5~k\Omega$3 and $5~k\Omega$4, accuracies were 77.29%, 72.22%, and 81.32%, with ITRs 91.28, 70.94, and 109.73 bits/min. On the HoloLens, offline experiments using $5~k\Omega$5 and $5~k\Omega$6 achieved a mean accuracy of 73.39%, while online HoloLens ensemble experiments produced accuracies of 75.17%, 77.76%, and 75.58%, with ITRs 82.38, 93.29, and 76.39 bits/min. A paired $5~k\Omega$7-test comparing ensemble versus best individual classifiers yielded $5~k\Omega$8, which the authors interpret as statistically significant superiority of the ensemble classifier (Mustafa et al., 2023).

Direct AR neurofeedback evidence is strongest in the MI-driven AR-robot system, though the study remains a small pilot with three healthy adults aged 22–24. Baseline MI calibration achieved 93.1 percent accuracy, false activation rate 8.3 percent, decision time $5~k\Omega$9 s, and ITR 14.8 bits/min. The paper defines ITR as

$S_i = R(f_i) + R(2f_i) + R(3f_i),$0

and defines the sustained control index as

$S_i = R(f_i) + R(2f_i) + R(3f_i),$1

In the AR ablation, accuracy was 75.6% for No AR, 76.2% for Static, 81.0% for Sham, and 96.9% for Neurofeedback. Decision time was 3.88 s, 3.76 s, 4.16 s, and 4.00 s respectively. False positive rate was 0.0%, 0.0%, 11.1%, and 2.8%. ITR was 11.9, 9.3, 15.7, and 21.3 bits/min, and the paper states that Neurofeedback achieved the highest ITR with improvements significant relative to all three baselines at $S_i = R(f_i) + R(2f_i) + R(3f_i),$2, Holm-corrected. In the full closed-loop grasping experiment, selection-to-confirm time was $S_i = R(f_i) + R(2f_i) + R(3f_i),$3 s, planning time $S_i = R(f_i) + R(2f_i) + R(3f_i),$4 s, execution time $S_i = R(f_i) + R(2f_i) + R(3f_i),$5 s, total cycle time $S_i = R(f_i) + R(2f_i) + R(3f_i),$6 s, and grasp success rate $S_i = R(f_i) + R(2f_i) + R(3f_i),$7, with the only reported failure due to EEG misclassification (Wang et al., 25 Sep 2025).

The pediatric mobile AR neurofeedback game provides early feasibility evidence rather than controlled efficacy. In the single-session field study with five autistic children, 1932 seconds of performance logs and 8611 events were collected. Children completed the game in about 4–5 minutes. The percentage of time in medium-performance stage was 28% for P1, 68% for P2, 62% for P3, 14% for P4, and 85% for P5, with about 23 transitions per child, 11 up-switches, and 12 down-switches. In the three-week deployment with one child across eight sessions, mean social attention index values were 31.22, 21.43, 34.37, 44.32, 45.76, 46.60, 50.47, and 44.67, which the authors interpret as an overall upward trend (Lyu et al., 6 Mar 2025).

A broader AR-mediated feedback result comes from upper-limb rehabilitation. In 15 healthy subjects, end-effector error was $S_i = R(f_i) + R(2f_i) + R(3f_i),$8 cm for No Holo, $S_i = R(f_i) + R(2f_i) + R(3f_i),$9 for Holo C1, $i$0 for Holo C2, and $i$1 for Holo C3. Joint-space error was $i$2 deg for No Holo, $i$3 for Holo C1, $i$4 for Holo C2, and $i$5 for Holo C3, with significant effects of Exercise and Condition and a non-significant Exercise × Condition interaction. Although this is not brain-signal neurofeedback, it provides quantitative evidence that AR can improve precision in closed-loop rehabilitation feedback (Luciani et al., 2024).

For AR-guided neurostimulation, the neuronavigation framework reports depth-estimation standard deviations of 0.07–0.09 mm, angular standard deviations of 0.04–0.06°, average localization error of 4.94 mm, and a case study with ten participants with a medical background indicating high usability. This is not neurofeedback in the strict sense, but it is relevant as adjacent evidence for the precision and usability of AR-based in-situ neurotechnology feedback (Hu et al., 28 Jan 2026).

6. Limitations, controversies, and future directions

The principal limitation of the field is definitional overreach. Several widely cited AR-BCI and AR-guidance studies do not constitute neurofeedback in the narrow sense because they do not train voluntary self-regulation of an endogenous neural feature. The immersive SSVEP navigation study and the HoloLens SSVEP framework explicitly demonstrate command selection, not continuous feedback about arousal, workload, affect, or self-regulation (Faller et al., 2017). The TMS neuronavigation papers do not acquire neural signals at all and are more accurately described as AR-guided stimulation or clinician-facing real-time neuronavigation feedback (Hu et al., 23 Jan 2026).

A second limitation is sample size. The immersive SSVEP AR study used three healthy young men; the MI-driven AR-robot neurofeedback system used three healthy adults; the HoloLens SSVEP study used six volunteers split across PC and HoloLens conditions; the pediatric AR neurofeedback game involved five autistic children in a field study and one child in deployment; and the TMS AR-guidance case study involved ten participants with a medical background. These numbers support feasibility but not strong generalization.

A third limitation is methodological under-specification. The immersive SSVEP study does not describe a dedicated hardware synchronization circuit between EEG and stimulus onset presentation and does not report photodiode timing, explicit latency compensation, or detailed artifact handling (Faller et al., 2017). The HoloLens SSVEP framework does not report end-to-end latency despite relying on a socket bridge between Python and Unity, and it addresses motion interference with relatively lightweight processing rather than explicit sensor fusion or artifact decomposition (Mustafa et al., 2023). The MI-based AR-robot system does not specify the exact linear classifier variant, exact decision fusion among classes, or the numerical dwell time or threshold governing stabilization before target switching (Wang et al., 25 Sep 2025). The pediatric AR neurofeedback game relies on a vendor-derived social attention index and does not provide the underlying mu-suppression computation, feature extraction equations, or latency measurements (Lyu et al., 6 Mar 2025).

A fourth issue concerns comfort and ecological realism. SSVEP paradigms rely on conspicuous flicker targets, which may constrain long-duration use even though the 2017 study reports that all subjects were neutral regarding annoyance caused by oscillating stimuli (Faller et al., 2017). The HoloLens framework explicitly suspects unstable frame rates as a cause of shifted spectral peaks and therefore uses $i$6 Hz frequency windows rather than exact bin matching (Mustafa et al., 2023). The AR-robot neurofeedback system reports high mental demand and effort, each $i$7, and post-task fatigue $i$8 on a visual analog scale, indicating that improved performance does not imply low workload (Wang et al., 25 Sep 2025).

A fifth controversy concerns what variable should be fed back. SSVEP provides reliable online classification and can tolerate modest calibration, but it is exogenously driven by flicker and sustained gaze. This makes it well suited to command selection and attentional engagement with tagged objects, but less suitable for self-regulation of endogenous states such as relaxation or emotional control (Mustafa et al., 2023). MI-based systems avoid the dependence on flicker but remain vulnerable to subject-specific variability and prolonged training demands (Wang et al., 25 Sep 2025). Vendor-derived scalar metrics, such as the social attention index in Eggly, simplify deployment but reduce transparency and reproducibility (Lyu et al., 6 Mar 2025).

The likely future direction is multimodal convergence. The neuronavigation papers already provide real-time registration, digital twin updating, and anatomically anchored AR visualization (Hu et al., 28 Jan 2026). A plausible extension is to combine these spatial layers with online EEG, fNIRS, TMS-EEG, EMG, or motor-evoked responses so that AR no longer visualizes only geometry but also physiology and target engagement. A second direction is more robust mobile decoding under movement, because practical AR use implies head motion, changing viewpoints, and environmental distraction. A third is the design of feedback that is contingent and spatially grounded without becoming visually intrusive. The AR-robot study’s sham condition is especially instructive here: motion alone is not sufficient, and incongruent motion can actively destabilize control (Wang et al., 25 Sep 2025).

In its present state, the literature supports a cautious synthesis. AR can host real-time EEG-driven interaction loops, can embed feedback into meaningful contexts rather than abstract bars, can improve kinematic precision in rehabilitation-style feedback, and can deliver anatomically anchored spatial guidance in neurostimulation workflows. Direct evidence for therapeutic AR neurofeedback remains limited, but the technical ingredients—portable sensing, AR display, registration, adaptive thresholding, and scene-integrated feedback—are now sufficiently concrete to define an emerging research area rather than a purely speculative one.