Abstract

Background:

Phantom limb pain (PLP) is a challenging neuropathic pain condition following limb amputation or brachial plexus injury. Non-pharmacological interventions such as motor imagery training, phantom motor execution and mirror therapy have shown potential to alleviate PLP by engaging sensorimotor circuits, but their effects are debated. We developed GHOST, a portable EEG-based brain–computer interface (BCI) coupled with immersive virtual reality (VR), allowing patients to control a virtual limb via motor imagery in real time, as a neurofeedback-based rehabilitation tool.

Methods:

We conducted a single-center exploratory pilot trial to assess the feasibility and preliminary efficacy of this device. Seven patients with chronic upper-limb PLP (amputees or brachial plexus avulsion, pain ≥3/10) underwent 10 training sessions over 2 weeks. Daily pain diaries (distinguishing continuous pain vs. paroxysmal pain episodes) were recorded for 1 month before and 1 month after the intervention, with follow-up to 6 months. Motor imagery ability, anxiety-depression (HADS), and quality of life (SF-36) were also evaluated.

Results:

Six patients completed ≥8 sessions. All participants achieved BCI control of the virtual hand, with high success rates (>70%) even as task difficulty increased, demonstrating system feasibility. No adverse events occurred. Compared to baseline, patients experienced a significant short-term reduction in paroxysmal pain (frequency and intensity of pain “flare-ups”), with up to >80% median decrease in weekly cumulated pain episode intensity (p < 0.001). Three of five patients also reported around 30% improvement in average daily pain during the first post-training month. HADS anxiety/depression scores showed a non-significant improving trend. By 3–6 months post-training, pain levels had largely returned to pre-intervention values.

Conclusion:

This pilot study supports the safety and feasibility of EEG-BCI with VR for PLP and suggests that it can yield short-term analgesic effects, particularly on paroxysmal pain. These findings support the hypothesis that sensorimotor re-engagement could effectively target maladaptive neural processes underlying PLP, while warranting confirmation in controlled trials. Future work will optimize the training protocol and investigate neuroplastic changes associated with this BCI-VR intervention.

1 Introduction

Amputation or partial/complete deafferentation of a limb often results in the experience of the “phantom” limb (Kooijman et al., 2000), whereby patients perceive the absent limb through sensory or kinesthetic hallucinations (so-called “phantom sensations”). Frequently, these patients also suffer from a specific type of chronic neuropathic pain—phantom limb pain—that is complex to treat and generally resistant to combined medical and surgical approaches (Ephraim et al., 2005). Two types of pain are described: continuous pain of variable intensity, commonly described as a burning background pain and paroxysmal pain episodes characterized by electric shock-like sensations lasting from a few seconds to several hours, referred to as painful exacerbations. These two pain types can occur independently. The pathophysiology of this pain remains poorly understood. One of the leading proposed mechanisms is “maladaptive plasticity” resulting from an abnormal reorganization of the primary sensorimotor cortex contralateral to the affected limb (Flor et al., 2006). This reorganization is thought to occur as a consequence of non-use or underuse of the sensorimotor brain region following severe deafferentation of the limb. In upper limb cases, the extent of this cortical plasticity appears to correlate with both the incidence and severity of phantom pain (Karl et al., 2001).

Despite a low level of evidence, it has been shown that motor imagery training—where patients imagine movement of the affected limb—can reduce pain, potentially by inducing cortical reorganization (Beaumont et al., 2011). However, as the ability to imagine movement is a subjective cognitive faculty that is difficult to evaluate, its clinical application remains challenging. Other noninvasive therapeutic strategies, such as mirror therapy (Giraux and Sirigu, 2003; Foell et al., 2014), are also frequently used. In mirror therapy, patients observe the movements of the intact limb reflected in a mirror placed symmetrically, thereby creating the visual illusion of the limb’s presence (Ramachandran and Rogers-Ramachandran, 1996) and, in turn, activating the corresponding motor circuit. These considerations render the phenomenon an attractive therapeutic target.

Based on these elements, we developed a noninvasive neurofeedback medical device named GHOST, which combines motor imagery and mirror therapy techniques. Its concept was first presented at the National Congress of the Société Française d’Évaluation et de Traitement de la Douleur (SFETD, Bordeaux, France, 2016). The device falls within the category of brain–computer interfaces (Daly and Wolpaw, 2008) and provides the user with real-time feedback on their imagery of movement in a virtual reality (VR) environment (Leeb and Pérez-Marcos, 2020). Specifically, when a patient—equipped with an electroencephalograph (EEG) connected to a brain decoder—successfully imagines the requested movement (opening/closing of the affected hand), they see, through a VR headset, a first-person view of an avatar performing the corresponding movement. This constitutes active feedback via a motor imagery training program with immersive visual return.

Since its inception, a class III evidence study using a similar principle—albeit substituting magnetoencephalography (MEG) for EEG and a robotic hand for feedback—demonstrated a significant reduction in phantom pain lasting one week, after three days of training, in 12 patients (Yanagisawa et al., 2020b). Although MEG is a powerful diagnostic and research tool, its bulk, complexity, limited availability, and cost make it unsuitable for routine therapeutic deployment. More recently, a multicenter, double-blind randomized controlled trial (Lendaro et al., 2025) evaluated two extended-reality (XR) interventions in 81 patients, comparing phantom motor execution (PME; real-time myoelectric decoding from the residual limb overt execution of phantom movements to control a virtual limb) with phantom motor imagery (PMI; imagined movements). Both groups exhibited substantial reductions in pain, and PME did not demonstrate superiority over PMI (non-significant between-group difference).

Our study provides a proof of concept for a portable device intended for clinical use. The aim was to evaluate both the feasibility and the potential effects of the device on upper-limb phantom pain in patients with either upper limb amputation or severe brachial plexus injury. The intervention is based on the principle of directly targeting maladaptive plasticity within sensorimotor circuits.

2 Patients and methods2.1 Patients

Patients were selected from those treated at the Pain Evaluation and Treatment Center of the University Hospital of Nantes (Nantes, France) or referred by specialized centers in the region. They presented with pain fulfilling the characteristics of neuropathic phantom limb pain affecting the upper limb. Inclusion criteria were as follows: an average pain score of at least 3/10 on the Numerical Rating Scale (NRS); a brachial plexus injury or an upper-limb amputation at least at the wrist; partial or complete motor or sensory deficit for more than 6 months; and absence of cognitive disorders or central nervous system lesions, particularly somatosensory ones (verified by brain and spinal MRI in patients with traumatic injuries).

Patients participated in a single-center, pilot, exploratory, uncontrolled, non-randomized, open-label, prospective clinical trial evaluating a noninvasive medical device as a technological innovation. As this was an exploratory study, the sample size (n = 7) was determined based only on the center’s recruitment capacity during the inclusion period, consistent with sample sizes reported in the literature for non-invasive interventions using neurofeedback techniques. The study spanned 7 months according to the schedule, and comprised: (1) a one-month pre-intervention period (m-1) during which subjects recorded daily evaluations of continuous pain (“minimal,” “maximal,” “average” terms refer to patients’ self-reported minimum and maximum pain levels, and their mean pain rating over the course of the day, on the Numerical Rating Scale) and paroxysmal pain (“duration,” “intensity,” “number” of pain peaks); (2) an experimental device usage period consisting of 10 sessions of 2 h each, conducted daily over 2 weeks (w1-w2, 5 sessions per week); and (3) a subsequent follow-up period (w3-w4 and m1-m6). Patients were required to complete at least 8 of the 10 scheduled training sessions over the 2-week period. Daily pain evaluations were recorded by the patients in a self-assessment diary from day 1 to day 30 following the first session, in the same manner as during the pre-intervention period, and then weekly until the end of the inclusion period (m6). Questionnaires assessing anxiety-depression (Hospital Anxiety and Depression Scale, HAD) and quality of life (SF-36) were completed during the information visit, at day 30, and at the end of the protocol. A subjective evaluation questionnaire of motor imagery ability (the French version of the MIQ-RS: Motor Imagery Questionnaire-Revised Second Version) was administered during the information visit.

A “BCI Competence Test” was performed at day −30. The purpose of this test was to determine whether the EEG-based brain decoder algorithm could detect a specific electrical signature associated with the change between resting state and when the patient executed a series of motor intention tasks with the affected upper limb (see Experimental Procedure). This criterion was mandatory for a patient’s inclusion in the study, as it constituted the indispensable basis for device training. In other words, in the absence of an identifiable signature, no feedback could be triggered. A maximum of three attempts were allowed (Figure 1).

Consolidated standards of reporting trials (CONSORT) flow diagram for the GHOST pilot trial.

The study was conducted in accordance with the Declaration of Helsinki, approved by the Comité d’Éthique Ouest VI – Brest (France) (N°1,111/DM2), and registered in the EudraCT (n° 2017-A03484-49) and Clinical Trial (n° NCT03889353) registries. Written informed consent was obtained from all patients after detailed explanation of the procedure, the training sessions, and any potential risks.

2.2 Presentation of the medical device

The medical device used was designed and assembled by the University Hospital of Nantes and the École Centrale de Nantes. It belongs to the category of neurofeedback brain–computer interfaces whose general principle is to measure the user’s brain activity in real time, extract qualitative or quantitative information, and provide feedback so that the user can voluntarily adapt their brain activity or mental task. Operating in a closed-loop system, our device is composed of three main hardware components and two software components: a virtual reality headset (Hardware: Head Mounted Display HTC Vive, hand tracking system Leap Motion Controller. Software: SteamVR v1497390325), a computer (Intel Core i7-7800X, base frequency 3.5 GHz, 16 GB RAM, MSI GEForce GTX 1080 Ti Armor OC with 11 GB video memory, running Windows 10), and a 64-channel EEG amplifier (g.tec Medical Engineering g.HIamp 80 CHANNEL AMPLIFIER, CE-marked European class II, certified by TÜV SÜD Product Service GmbH) along with an elastic EEG cap with standard 10–10 system placements plus 86 intermediate positions (64 active g.tec electrodes: g. SCARABEO, 1 g. SCARABEOgnd as ground electrode, 1 g. GAMMA earclip as reference), an adapter cable (g. HEADbox – active), and a preamplifier. In addition, a signal processing software (“brain decoder”) and a virtual reality software (developed with Unity 3D 5.6.0f3, 64-bit) provide the visual feedback via a virtual avatar (Figure 2).

Schematic illustration (artist’s rendering) of the GHOST device during use in a patient with a left upper-limb amputation. The system comprises the following hardware: an HTC vive virtual reality headset; a leap motion controller; a personal computer; a 64-channel EEG amplifier (g.tec Medical Engineering, g.HIamp 80); an elastic EEG cap equipped with 64 active electrodes (g.tec: 64 g.SCARABEO electrodes, one g.SCARABEOgnd ground electrode, and one g.GAMMA ear-clip reference electrode); an active adapter cable (g.HEADbox); and a preamplifier. In addition, two in-house software modules—a signal-processing pipeline (“brain decoder”) and a virtual reality application developed using the Unity 3D engine (v5.6.0f3, 64-bit)—provide visual feedback via a first-person virtual avatar. The patient is seated comfortably. Movements of the intact right upper limb are tracked by the Leap Motion sensor mounted on the VR system and are faithfully reproduced by the avatar displayed to the patient. In the virtual environment, the patient’s amputated limb is replaced by an intact virtual arm embodied by the avatar. This virtual limb is then animated (hand opening and closing movements) when kinesthetic motor imagery of the affected limb is detected from the EEG signal by the pre-calibrated brain decoder developed by our team. The patient performs different exercises along sessions, providing active feedback through a motor imagery training program with immersive visual feedback.

During use, the patient wears the EEG cap and VR headset through which they view a first-person representation of an avatar. Movements are tracked via an infrared sensor (Leap Motion Controller) and replicated by the avatar, thereby virtually restoring body integrity. The affected limb (amputated or paralyzed) is animated via the interface based on the kinesthetic motor imagery performed by the user. Thus, the device integrates neurofeedback based on motor imagery coupled with the principle of mirror therapy in an immersive virtual environment.

The signal processing software, also referred to as the “brain decoder,” was developed jointly by the University Hospital of Nantes and École Centrale de Nantes. Its primary functions are to: (1) acquire EEG signals directly from the EEG device; (2) display these signals in real time to monitor quality; (3) interpret the EEG signals in real time to determine the user’s mental state and transmit this information to the VR software; and (4) guide the user through various mental tasks while recording EEG signals to build a predictive model.

The virtual avatar feedback software was similarly developed by the same institutions. Its main functions are to: (1) immerse the user in a simple virtual world where the avatar does not exhibit the affected limb; (2) react to the user’s mental state—in that when the user thinks of opening and closing their affected hand, the avatar’s corresponding hand performs that movement; and (3) mirror the user’s movements in real time.

The information systems employed comply with medical cybersecurity standards.

2.3 Experimental procedure

Each session was conducted in a dedicated room meeting comfort and safety standards, under the joint supervision of the investigator physician and an associated engineer.

The “BCI Competence Test” realized at day −30 was conducted as follows. The patient, with the EEG cap in place, was comfortably seated in front of a computer screen. They first underwent two “baseline” 1-min rest phases. In the first, the patient was instructed to remain calm and relaxed with eyes open and gaze fixed; in the second, to remain calm with eyes closed and without movement. Then, the patient performed 9 series of tasks, each series consisting of 6 trials per task (i.e., 54 trials per task and 162 trials per session). In each series, the interface randomly cued one of the following tasks: (1) imagine opening and closing the right hand at 0.5 Hz (“right hand”); (2) imagine opening and closing the left hand at 0.5 Hz (“left hand”); (3) remain calm, relaxed, with eyes open and fixed (“relax”). All patients were instructed to perform kinesthetic motor imagery, i.e., mentally rehearsing the movement by focusing on the sensation of executing it, without overt movement. Once the recording was complete, the acquired data were immediately analyzed and the software calculated the success rate of the brain decoder’s classification algorithm, which had to reach a threshold of 70% for the patient to be definitively included. In case of failure, the test could be repeated up to three times. The success rate was defined as the ratio of the number of correct decisions to the total number of decisions made by the brain decoder.

Once enrolled, patients participated in one session per day (excluding weekends) (Figure 3), with a total of 10 device sessions scheduled over 2 consecutive weeks (w1-w2). Each session included the following main steps: (1) Device setup, installation (EEG cap and VR headset) – 30 min; (2) “Open-loop Calibration” phase similar to that in the “BCI Competence Test” – 15 min; (3) “Closed-loop Training” phase followed by “Free Use” – 30 min; and (4) a debriefing period – 30 min.

Flowchart of the experimental procedure.

The “Closed-loop Training” phase, was designed to train the patient to generate increasingly stable and detectable EEG signatures through motor kinesthetic imagery, guided by neurofeedback delivered via the animation of the corresponding virtual limbs of the avatar. To this end, the engineer determined along the session both the quantity and difficulty of the exercises based on the patient’s past performance, fatigue, and emotional state. Adjusting the difficulty aimed to promote progress while preserving motivation. During the initial sessions, patients practiced exercises designed to generate stable EEG signatures (i.e., sustaining each signature for a few seconds). As proficiency increased, the instruction was to extend the duration of signature maintenance. Once mastered, patients were trained to alternate between different mental states according to temporal constraints (For example, one task could be: “right hand” motor imagery for 6 s, “relax” for 3 s, then “left hand” motor imagery for 3 s). Gradually, both the number of alternations and the duration spent in each state were increased whenever feasible. During each training phase, the following parameters were recorded: (1) number of attempts for each task, (2) duration of each task, (3) number of mental states per task, (4) number of successes and failures, as well as (5) the time required to validate an attempt. We did not aim to standardize the training beyond its duration in order to preserve the flexibility required for any individualized rehabilitation process.

The “Free Use” phase, conducted at the end of each session, allowed patients to control the avatar continuously in an unsupervised mode. Although this usage scenario was more natural, it remained cognitively demanding, requiring sustained attention since the brain decoder was active at all times.

2.4 Signal processing and “brain decoder”

The brain decoder’s main function is to determine which mental task the patient is currently performing (“right hand” motor imagery, “left hand” motor imagery, or “relax”).

EEG data were acquired either during the BCI “Competence Test” or, at each session, during the “Open-loop Calibration” phase (Figure 3). Signals were first band-pass filtered using a finite impulse response (FIR) filter to retain the frequency band of interest (8–30 Hz), encompassing the alpha (8–12 Hz) and beta (13–30 Hz) rhythms that typically modulate during motor imagery. For each trial, EEG data from 1 to 6 s were extracted and segmented using a 2-s sliding window with 50% overlap, yielding a 3D data structure of size (n_windows × n_channels × n_samples). After filtering and windowing, class-wise statistical characteristics were computed and windows identified as outliers were excluded from the dataset (the artifact-rejection procedure will be described in a future publication).

Following artifact rejection, features were extracted using Common Spatial Patterns (CSP) (Blankertz et al., 2008; Grosse-Wentrup and Buss, 2008)—which computes spatial filters maximizing variance differences between the conditions to be discriminated (MNE-Python implementation) (Gramfort et al., 2013)—and classified using Linear Discriminant Analysis (LDA) [scikit-learn implementation (Pedregosa et al., 2011)]. Decoder performance was assessed using k-fold cross-validation, with accuracy as the performance metric. The number of folds (k) was set to the number of available runs (typically 9 for the “BCI Competence Test” and 6 for the “Open-loop Calibration” phase).

After cross-validation, if the mean CSP–LDA accuracy exceeded 70%, the CSP–LDA model was fitted on the full open-loop dataset and subsequently used during the “Closed-loop Training” phase. An example of CSP-learned spatial patterns is shown in Figure 4. To confirm that the identified signatures reflect genuine sensorimotor EEG activity, ERD/ERS time–frequency spectrograms can be computed. For each open-loop trial, we extracted an EEG segment from −1 s to 6 s. Power spectral density over time was estimated using the Discrete Prolate Spheroidal Sequences (DPSS) multitaper method (implementation available in MNE-Python). ERD/ERS was then expressed as a percent change relative to baseline by subtracting the mean power during the reference interval (−1 to 0 s) and dividing by the mean baseline power. An example computed from an open-loop session of subject 01–05 is shown in Figure 5.

Scalp topographies of EEG spatial patterns (patterns 0–5) derived from data recorded during the BCI competency test (patients 01–05). Each panel shows the scalp distribution of one spatial pattern extracted from the EEG recorded during the BCI competency test, enabling discrimination between the different mental states performed by the patient (“right hand” motor imagery, “left hand” motor imagery, or “relax”). Dots indicate electrode locations. Colors represent the relative contribution (weight) of each sensor to the pattern (arbitrary units; see color bar), with warm and cool colors indicating opposite polarities (the sign is arbitrary). These spatial patterns highlight the cortical regions that contribute most to the decoding of the mental tasks (e.g., sensorimotor areas for motor imagery).

Event-related desynchronization/event-related synchronization (ERD/ERS): time–frequency maps computed at electrodes C3, Cz, and C4 for the motor imagery tasks of subject 01–05, based on data from the BCI Competence Test session. A decrease in signal power in the alpha band is observed during left-hand motor imagery, most prominently at C4, located over the contralateral sensorimotor cortex, as described in literature.

During the “Closed-loop Training” phase, EEG was continuously acquired, FIR band-pass filtered, and segmented using a 2-s sliding window with an overlap of 15/16 (corresponding to an update every 62.5 ms, i.e., 16 decisions/s). For each window, the fitted CSP–LDA model was applied to label the signal into one of the three learned classes. The resulting label (Figure 6) was then sent to the VR application to animate the avatar.

Operator interface during the “Closed-loop training” phase. Left: Raw EEG signal; Right: Top panel showing the instruction provided to the patient during the brain decoder calibration phase (e.g., “left hand” motor imagery) and the bottom panel displaying the real-time classification result (predicted probability of belonging to a given mental state).

2.5 Statistical analyses

Given the small sample sizes collected in this study, our analyses primarily relied on descriptive statistics: median, standard deviation, and quartiles.

Given the repeated-measures structure and substantial inter-individual variability in pain trajectories, we employed a linear mixed-effects model (LMM) instead of traditional non-parametric approaches such as the Wilcoxon signed-rank test. LMMs explicitly account for within-subject correlation by incorporating random intercepts for each participant, thereby modeling baseline heterogeneity in pain perception. This approach is particularly suited for longitudinal neurorehabilitation studies with small samples, where inter-subject variability and unbalanced data are expected.

The primary fixed effect of interest was time period, defined in three levels: pre-intervention (m–1), intervention (w1–w2), and post-intervention (w3–w4). Random intercepts were included at the subject level. Model parameters were estimated using restricted maximum likelihood (REML). To assess the global effect of time, we performed a likelihood ratio test (LRT) comparing the full model to a reduced model excluding the period factor. Wald z-tests were used to evaluate the significance of fixed-effect coefficients, with results reported as estimated marginal means (differences between periods), standard errors, and 95% confidence intervals.

All analyses were conducted in Python using the statsmodels library (Seabold and Perktold, 2010), following standard recommendations for mixed-effects modeling of repeated-measures and neurophysiological data (Luke, 2017).

3 Results

Seven patients (5 men and 2 women, aged 28–65 years) were consecutively recruited at the University Hospital of Nantes between May 9, 2019, and November 9, 2021, meeting the inclusion criteria (Figure 1). One patient withdrew consent for professional scheduling conflicts. Data from one patient (01–03) could not be analyzed due to the presence of outlier values (see Discussion). Two patients experienced pain related to amputation (one traumatic, the other oncological), while the remaining four had brachial plexus injuries (one radiation-induced complication and three due to traffic accidents causing brachial plexus avulsions). The average duration since the injury was 93 months (range: 7–360 months). All included patients had refractory chronic pain evolving over similar time periods. The main clinical characteristics of the patients participating in the study are detailed in Table 1.

IDAge/SexDiagnosisDuration (months)Medical treatmentsSurgical treatmentsAlternative therapies01–0158/FRadiation-induced right brachial plexus injury (incomplete, C5)21Pregabalin 600 mg/day; Paracetamol 3 g/day; Piroxicam 20 mg/dayBrachial plexus neurolysis; C7–T2 DREZotomyrTMS, self-hypnosis, physiotherapy01–0228/MTraumatic right brachial plexus injury (complete, C5)44Duloxetine 90 mg/day; Marinol 7.5 mg/dayMusculocutaneous nerve neurotization via spinal accessory nerve (with intercalated sural nerve graft)Physiotherapy only01–0345/MTraumatic left arm amputation at shoulder level (work accident)7Morphine sulfate 30 mg/day (extended-release); morphine sulfate 40 mg/day (immediate-release); Amitriptyline 20 mg/day; Pregabalin 400 mg/day; Duloxetine 60 mg/dayLeft shoulder amputation (work accident)TENS, mirror therapy, motor imagery training01–0465/FOncologic left arm amputation at shoulder level (sarcoma)20Pregabalin 200 mg/day; Duloxetine 90 mg/dayLeft shoulder amputation (sarcoma)Mirror therapy, motor imagery training01–0538/MTraumatic left brachial plexus injury (C5–C6)105Paracetamol 1 g as needed; intolerance to analgesic medicationsNeurolysis of C8–T1 roots; nerve graft from C5 to musculocutaneous nerve; 2nd metacarpal elevation and derotation osteotomy; metacarpophalangeal arthrolysisQutenza (capsaicin 8% patch), physiotherapy01–0656/MTraumatic right brachial plexus injury (C5–C6)360Tramadol/Paracetamol combination (2/day); Pregabalin 250 mg/day; Duloxetine 60 mg/dayC5 nerve graft; spinal accessory nerve neurotizationMirror therapy, motor imagery training, TENS

Demographic and clinical characteristics of the study participants (n = 6).

For each patient, age, sex, diagnosis, duration of pain since injury (in months), ongoing medical treatments, previous surgical interventions, and alternative or non-pharmacological therapies are reported at inclusion. Diagnoses include traumatic or radiation-induced brachial plexus injuries and upper-limb amputations. Medical treatments reflect stable analgesic regimens at the time of enrollment. Surgical and alternative therapies were performed prior to inclusion and were not modified during the study period.

The reported pain exhibited the typical features of neuropathic phantom pain, with DN4 scores ranging between 5 and 9 [mean 6.83 ± 1.47, median 6.50 (6.00–8.00)]. According to the inclusion criteria, patients had an average pain score of at least 3 on the numerical scale. Three patients experienced between 11 and 20 pain crises per day, with one patient reporting more than 20 crises. Two-thirds of the patients also reported spontaneous pain lasting more than 11 h per day. All patients were receiving multidisciplinary care at pain treatment centers, and the various pharmacological treatments prescribed for neuropathic pain were in line with current recommendations. One patient had undergone a Drezotomy, and four patients with brachial plexus injuries had undergone nerve transfers/grafts/neurolysis. Additionally, five patients had previously tried nonpharmacological treatments (transcutaneous electrical nerve stimulation, self-hypnosis, mirror therapy, motor imagery training, physiotherapy) without long-term efficacy.

3.1 Concept of motor imagery training under EEG-BCI control with VR feedback: feasibility3.1.1 Self-assessed motor imagery ability: MIQ-RS

The MIQ-RS (Motor Imagery Questionnaire-Revised Second Version) evaluates an individual’s ability to imagine motor movements both visually (IMV) and kinesthetically (IMK). The questionnaire comprises seven items where participants are asked to visualize or feel specific movements, rating the ease of performing the imagery on a Likert scale ranging from 1 (very difficult to imagine) to 7 (very easy to imagine).

The mean MIQ-RS scores among patients were 21.67 ± 7.69 [median 25.00 (13.00–27.00)] for visual imagery and 24.00 ± 5.90 [median 23.50 (20.00–26.00)] for kinesthetic imagery (Figure 7). These scores were lower than those observed in healthy subjects in the validation study of the French version of the MIQ-RS (Loison et al., 2013).

Motor imagery ability scores (MIQ-RS) in GHOST patients. Visual (IMV) and kinesthetic (IMK) motor imagery scores in the GHOST cohort, assessed using the MIQ-RS (7-point scale; higher scores indicate easier imagery; total scores range from 7 to 49 on the ordinate), were lower than normative values reported in healthy subjects in the validation study of the French version of the MIQ-RS, indicating greater difficulty in imagining movements in this patient population.

Despite these lower scores, all subjects succeeded in generating EEG signatures via motor imagery that were detectable by the “brain decoder” during the “BCI Competence Test.”

3.1.2 Concept of training: feasibility—progressive increase in difficulty

Four patients completed all 10 planned sessions, one completed 9 sessions, and one completed 8 sessions. Each session lasted on average 2 h, including 30 min for setup (cap, headset), 15 min for calibration, and 30 min of training (a fixed duration). The remaining 30–45 min were allocated for reception, installation (dressing/undressing, comfortable positioning), administrative verification, collection of clinical feedback, review of pain diaries, and discussion. The time commitment and cognitive engagement required by the sessions did not constitute a barrier for the patients.

To evaluate the impact of training on avatar control during EEG-BCI feedback sessions, data on the number of sessions, session duration, and exercise difficulty were collected. The exercises involved activating a series of mental tasks corresponding to different states (“relax,” “right hand,” “left hand”). Increasing difficulty was introduced via several parameters (e.g., the maximum time allowed to validate a task, the duration the task is held, and the length of imposed task sequences). A composite score was generated from these elements to provide a clear evaluation of the difficulty presented to the subjects.

Performance levels remained very high (overall median success rates >70% in exercises) despite an increase in the imposed difficulty level. The results regarding performance and progression in exercise difficulty are shown in Figures 8, 9.

Difficulty of the training exercises for each patient across sessions [t1–t5 performed on days 1–5 during week 1 (w1), and t6–t10 performed on days 8–12 during week 2 (w2)]. The solid black line denotes the median across all participants. During each session, patients performed a sequence of exercises requiring alternation between different mental states under temporal constraints (e.g., right-hand motor imagery for 6 s, relax for 3 s, then left-hand motor imagery for 3 s). When feasible, task difficulty was progressively increased by adding more alternations and/or modifying the duration of each state. The difficulty level was derived from a composite score combining (i) the number of mental states required per task and (ii) completion time expressed as a proportion of the allotted time.

Raw evolution of patients performance during training sessions [t1 to t5 performed on days 1–5 during week 1 (w1), and t6 to t10 performed on days 8–12 during week 2 (w2)]. The success rate was defined as the number of successfully completed exercises divided by the number of attempts within each session, for each patient. The solid black line denotes the median across all participants.

3.2 Effect of training on pain and anxiety/depression3.2.1 Effect on paroxysmal pain

Paroxysmal pain showed a pronounced reduction during the intervention period, as reflected by descriptive percentage changes relative to baseline. At the group level, median paroxysmal pain volume decreased by approximately 35% at w1, 55% at w2, 64% at w3, and 86% at w4 compared with the pre-intervention condition (see Table 3). Longitudinal mixed-effects modeling confirmed a significant reduction in paroxysmal pain volume during the post-training phase (w3–w4; coef = −92.6, p = 0.001), whereas no significant group-level effect was observed during the early training phase (w1–w2; coef = 6.8, p = 0.814) (see Tables 4, 5). Weekly analyses further showed a significant decrease at w4 compared with baseline (coef = −111.4, p = 0.001), with earlier weeks showing non-significant or heterogeneous effects at the group level. At the individual level, significant reductions were observed in most patients during both early and late phases, while one patient exhibited a transient early worsening followed by delayed improvement. Long-term follow-up (w12–w23) was not included in statistical analyses, as the change from daily to weekly self-assessments introduced a measurement bias that precluded reliable longitudinal comparison. Individual- and group-level results are illustrated in Figure 10 using violin plots.