In post-stroke rehabilitation, biofeedback added to physiotherapy improves upper- and lower-limb motor recovery, and EMG biofeedback of swallowing helps the return to oral feeding. It is a rehabilitation adjunct, not a treatment for the stroke itself. EEG neurofeedback proper remains little studied; brain-computer interfaces that drive a robot or electrical stimulation are a different technology, rated separately.
Updated :
What the research shows
EEG neurofeedback — standard amplitude training
(SMR/mu, feedback-only) — motor recovery of the paretic limb
Ages
Adult (18+)
Techniques
Sensorimotor rhythm (SMR)
AAPB4EfficaciousNeuroLogic3Probably efficaciouslower vs AAPB
Level 3 (AAPB level assigned jointly to biofeedback and neurofeedback; disaggregated here.) The AAPB's Level 4 rests on surface EMG; its only two EEG studies are Rayegani 2014 (n = 46, four arms: hand function on the JHFT improved in all three treatment arms, SMR power increased in the neurofeedback arm) and Vourvopoulos 2019 (n = 4, BCI-virtual reality feedback, no significant improvement on any clinical scale). The feedback-only literature since adds Ribeiro 2023 (SMR neurofeedback meta-analysis: stroke SMD 0.31, 95 % CI 0.03-0.60 vs other therapies, I² = 46 %, which the authors read as no demonstrable clinical benefit) and Cha 2025 (randomised cross-over, n = 15 subacute patients: greater ipsilesional mu suppression after real than sham feedback, no clinical endpoint). Multiple controlled studies with small, largely neurophysiological effects and no independent replication on a functional endpoint: Level 3. Closed-loop BCI with a robot, FES or exoskeleton is rated separately (see other methods).
EEG neurofeedback — amplitude training (alpha/theta) — post-stroke cognitive impairment
Ages
Adult (18+)
Techniques
Alpha training
AAPB0Not ratedNeuroLogic2Possibly efficacious
Level 2. Gupta 2026: pilot randomised trial in subacute stroke, n = 15 (five per arm), 20 sessions of EEG neurofeedback or brainwave entrainment added to treatment as usual over three weeks; both intervention groups improved processing speed (digit symbol substitution) and delayed recall relative to treatment as usual, entrainment additionally improving verbal working memory, mood and quality of life. Adequate dose and clinician-delivered, but five participants per arm and blinding unstated. He 2026 reviews 30 BCI-based cognitive rehabilitation studies in vascular cognitive impairment (696 stroke participants): attention is the most consistently improved domain, BCI is usually combined with other technologies, and there is no pooled estimate. One small randomised trial with an adequate dose: Level 2.
Other neurofeedback methods
Closed-loop BCI with actuator (robot, FES, exoskeleton) — motor rehabilitation. The reinforcer is an external actuator, not a feedback display: this is not clinical neurofeedback
Ages
Adult (18+)
Techniques
Closed-loop BCI with actuator (robot, FES, exoskeleton)
AAPB0Not ratedNeuroLogic3Probably efficacious
Level 3. Rated as a distinct rehabilitation technology: the EEG classifier triggers a robot, exoskeleton or functional electrical stimulation that moves the paretic limb, so the therapeutic reinforcer is contingent motor and proprioceptive stimulation rather than a feedback display, and the AAPB joint rating never tested it. Qin 2026 (Cochrane, 43 RCTs, n = 1628): vs conventional therapy, upper-extremity motor function MD 4.67 (95 % CI 2.25-7.09; 10 studies, 611 participants; low certainty); vs active controls SMD 0.58 (95 % CI 0.23-0.92; 14 studies; very low certainty) and balance MD 3.25 (95 % CI 1.07-5.43); vs sham-BCI SMD 0.22 (95 % CI −0.08 to 0.52; 9 studies, 279 participants; low certainty). Yang 2026 (24 studies, n = 846): FMA-UE SMD 0.31 (95 % CI 0.18-0.45, I² = 17 %), WMFT SMD 0.45, no significant effect on activities of daily living or spasticity. Nojima 2022 (16 studies, n = 382): SMD 0.48 (95 % CI 0.16-0.80, I² = 45 %), SMR-focused algorithms most effective. Superiority over conventional therapy replicated across many independent groups, but certainty is low to very low throughout and the sham comparison is null; under Parsons 2026 the sham-null is not itself a reason to lower, but it does not lift the level to 4.
Level 2. Sanders 2022: double-blind sham-controlled RCT, n = 24 chronic stroke, three training days; motor-cortex laterality increased within (P = 0.019) but not across days; primary behavioural outcome (Jebsen Taylor Test total) null (P = 0.116), secondary gain on gross-motor subtasks (P = 0.010), no change on ARAT or FMA-UE (both P > 0.5), reduced corticospinal white-matter asymmetry at 1 week (P = 0.008). Butet 2025: bimodal EEG-fMRI neurofeedback vs motor imagery without feedback, n = 30 chronic stroke, five weeks; FMA-UE increased significantly only in the neurofeedback group (P = 0.003 vs P = 0.633), between-group P = 0.048, 8/15 vs 3/15 responders, maintained at one month; no sham. Boukrina 2026: feasibility RCT in post-stroke reading impairment, 4 contingent vs 3 non-contingent stroke participants; reading comprehension improved more in the contingent group (difference 11.42 points, 95 % CI 1.12-21.71), no change in reading aloud. Independent groups, randomised comparators and protocol-specific neural change, but samples of 24, 30 and 7 with a null or marginal primary endpoint: Level 2.
Level 2. Tetsuka 2023: randomised, sham-controlled, double-blind, n = 30 acute stroke, a single day of prefrontal fNIRS neurofeedback; the post-intervention decline in spatial working memory was prevented in patients who raised right prefrontal activity — a conditional, immediate effect. Zhang 2026: fNIRS-BCI contingent feedback triggering observation-imitation videos vs non-contingent sham, n = 44; both groups improved upper-extremity function with no between-group difference, while real feedback enhanced beta desynchronisation bilaterally and ipsilesional HbO self-regulation. Lin 2026: multicentre assessor-blind RCT, n = 50 post-stroke depression, 12 weeks; fNIRS-guided neurofeedback with art therapy added to CBT reduced HAMD-17 more than CBT alone (−11.4 vs −8.6, P < 0.001, d = 1.1; remission 80 % vs 40 %, P = 0.009) — but the neurofeedback is bundled with art therapy, so its specific contribution cannot be isolated. Controlled studies with neural or confounded clinical effects: Level 2.
Biofeedback
Surface EMG biofeedback — upper- and lower-limb motor recovery
Ages
Adult (18+)
Techniques
EMG biofeedback
AAPB4EfficaciousNeuroLogic4Efficacioussame vs AAPB
Level 4 maintained (AAPB level assigned jointly to biofeedback and neurofeedback; surface EMG is the evidence the AAPB rating rests on.) Randomised superiority over therapeutic exercise or physiotherapy alone across four decades and independent settings: Basmajian 1975 (n = 20, strength and range-of-motion gains about twice those of exercise alone), Burnside 1982 (n = 22), Intiso 1994 (n = 16), Armagan 2003 (n = 27, vs placebo SEMG: greater active range of motion and SEMG amplitude), Kim 2017 (n = 30), Dost Sürücü 2021 (n = 40), Lirio-Romero 2021 (n = 38, single-blind, vs sham SEMG, greater F-MA-UE gain); Schleenbaker 1993 (8 studies, mean effect size 0.81), Moreland 1998 (8 studies, ankle dorsiflexor strength significant). The 2022-2026 window is consistent: Wang 2024 (10 RCTs, n = 303, SMD 0.44, 95 % CI 0.12-0.77; short-term SMD 0.33 significant, long-term not), Yao 2026 (8 studies, n = 549, total FMA MD 9.50, ADL MD 8.80, I² 69-98 %, prediction intervals crossing zero), Feng 2022 (Bayesian network meta-analysis, 45 RCTs, n = 3379: EMG-BF plus rehabilitation training ranked first for FMA-UE and pain in shoulder-hand syndrome, SUCRA 96.8 %). Munoz-Novoa 2022 found no difference between sEMG-driven and other interventions on FMA-UE (SMD 0.14, 14 studies) but pools biofeedback with EMG-triggered stimulation and robotics. Level 5 not assigned: no superiority over a bona fide reference treatment.
EMG-triggered neuromuscular electrical stimulation — EMG as trigger for stimulation, not as feedback
Ages
Adult (18+)
Techniques
EMG biofeedback
AAPB4EfficaciousNeuroLogic4Efficacioussame vs AAPB
Level 4 (AAPB level assigned jointly; this row disaggregates the EMG-triggered stimulation studies AAPB counts toward its Level 4.) Separate row because the mechanism is stimulation contingent on a voluntary EMG threshold, not feedback of the signal itself. Bolton 2004 (5 studies, 86 participants, mean effect size 0.82 on arm and hand function) and Monte-Silva 2019 (26 studies, 782 participants: robust short-term effect on upper-limb body structure and function versus dose-matched therapies and no treatment, stronger in chronic than acute/subacute patients, no significant benefit at longer-term follow-up). Guerrero-Mendez 2026 (38 studies, n = 1132): EMG-triggered stimulation alone not superior to conventional therapy, but combined with other interventions it improved all ICF domains; 17 of 38 trials at high risk of bias. Randomised superiority over dose-matched therapy, replicated in independent settings, short-term: Level 4.
Positional, gait, balance and force feedback (force platforms, wearables, joint-angle and finger-force feedback)
Ages
Adult (18+)
Techniques
Other biofeedback
AAPB4EfficaciousNeuroLogic4Efficacioussame vs AAPB
Level 4 maintained (AAPB level assigned jointly; AAPB counts positional feedback as biofeedback.) AAPB base: Mandel 1990 (n = 37, no-treatment control vs SEMG vs positional feedback: walking speed increased in both feedback groups, significant in the positional-feedback group and maintained at 3 months), Colborne 1993 (crossover, n = 8, significant stride-length and velocity gains), Pachiappan 2020 (n = 30, balance feedback vs task-oriented activity, both improved). Seo 2025 (Stroke): double-blind RCT, n = 45, 18 sessions of three-dimensional vs one-dimensional finger-force feedback; ARAT change 3.5 (CI 2.2-4.8) vs 0.8 (CI −0.5 to 2.1), P = 0.005, maintained at 1 month — the methodologically strongest biofeedback trial in the window. Dai 2026 (10 RCTs, n = 304): gait velocity MD 9.78 cm/s (95 % CI 6.06-13.50; 95 % PI 2.14-17.41), step length MD 5.88 cm with a prediction interval crossing zero. Wang 2025 (13 RCTs, n = 304, wearable gait feedback): gait speed SMD 0.41 (P = 0.02), Berg Balance SMD 0.44, Timed Up and Go SMD −0.36, no ADL effect. Randomised superiority in independent settings, including one double-blind trial: Level 4.
Level 3. Not covered by the AAPB rating. Alyanak 2025: assessor-blind RCT, n = 33, 15 sessions of Mendelsohn manoeuvre and effortful swallow with game-based EMG biofeedback vs the same exercises with verbal feedback (dose-matched); FOIS (P = 0.038), PAS-liquid (P = 0.026) and DOSS (P = 0.003) improved in the biofeedback group only, post-intervention PAS-semisolid better than control (P = 0.031). Kang 2026: single-blind RCT, n = 60, sEMG-triggered immersive virtual-reality game training added to conventional swallowing therapy, greater gains in WST, SSA, FOIS and sEMG RMS (P < 0.05), contact time unequal. Nordio 2022 (pilot RCT, n = 17: FOIS improved in both arms, biofeedback additionally improved pharyngeal clearance and swallow safety on FEES, stable at 2 months), Hou 2024 (three-arm RCT, n = 90) and Wang 2026 (n = 30, respiratory-swallow coordination with bimodal biofeedback, greater FOIS gain at post-intervention and 1 month, P < 0.001) are positive; Benfield 2023 (feasibility RCT, n = 27): DSRS 3.2 vs 4.3 at two weeks, not significant, not powered for efficacy. Syntheses are borderline: Zhang 2026 (5 RCTs, n = 174) FOIS MD 1.26 (95 % CI 0.03-2.48, P = 0.047, I² = 62 %), no significant PAS change; Toledo-Rodríguez 2026 (5 RCTs): possible advantage on tube removal and quality of life, no difference in severity, very low certainty. Multiple small randomised trials, one assessor-blind with a dose-matched comparator, pooled effects borderline: Level 3.
In short
Clinical reading
AAPB Level 4 assigned jointly to biofeedback and neurofeedback, carried by surface EMG (some twenty trials since 1975, six meta-analyses). NeuroLogic separates the modalities: limb SEMG biofeedback 4, EMG-triggered stimulation 4, positional, gait and force feedback 4 (Seo 2025, double-blind), dysphagia 3 (new); amplitude EEG neurofeedback 3 (only two EEG studies in the AAPB base: Rayegani 2014, Vourvopoulos 2019), post-stroke cognition 2; closed-loop BCI with actuator 3 (Qin 2026, Cochrane, 43 RCTs), fMRI 2, fNIRS 2.
Protocols
SEMG on paretic groups (15-18 sessions integrated into physiotherapy), EMG-triggered neuromuscular stimulation, positional, gait, balance and finger-force feedback; submental SEMG with Mendelsohn manoeuvre and effortful swallow for dysphagia. Amplitude EEG neurofeedback (SMR/mu) remains experimental in this indication; closed-loop BCI protocols with a robot, FES or exoskeleton, sometimes in virtual reality, are a distinct motor-rehabilitation technology, rated separately and not as clinical neurofeedback.
Limits
Uneven methodological quality, varying definitions of biofeedback across studies, frequent confounding with rehabilitation intensity; effects mainly short-term (Wang 2024, Monte-Silva 2019); no clinical-fidelity trial of amplitude EEG neurofeedback (clinician-adjusted thresholds, 20-40 sessions); an adult-only literature — no paediatric data.
Study base
Mature biofeedback literature — dozens of RCTs and six meta-analyses of surface EMG since the 1970s, plus an active 2022-2026 dysphagia and gait-feedback literature; the neurofeedback base is small, recent and dominated by BCI-with-actuator trials. 2022-2026 base: 31 publications indexed in the archive (18 biofeedback, 13 neurofeedback/BCI).
Brendan's perspective
This is one of the rare places where I go below the AAPB, and only because their level 4 was assigned jointly to biofeedback and neurofeedback. Disaggregate the modalities and the surface EMG literature keeps its 4 easily: four decades of randomised superiority over physiotherapy alone. The EEG side does not. Rayegani 2014 and Vourvopoulos 2019 are what AAPB actually has, and nothing since resembles clinical-fidelity amplitude training — no qEEG-derived protocol, no thresholds adjusted by hand, no 20 to 40 sessions, no transfer work. Level 3 is a statement about what has been studied, not about what the method could do. I also separate closed-loop BCI with a robot or FES: there the reinforcer is contingent proprioceptive stimulation, not a signal handed back to a person. Its sham comparison is null, which on my reading of sham is not a null. What I would tell a client: after a stroke, biofeedback is a rehabilitation adjunct with real evidence, and EEG neurofeedback is not yet part of the answer on its own. One qualification to that, and it is the one that matters in practice. Nothing about staying at 3 for neurofeedback implies I would deliver it alone. The approach I would actually put in front of a stroke patient is the combination — surface EMG work inside physiotherapy, with EEG training alongside it — and on the strength of the biofeedback evidence that combination sits back at a 4. The level on this row is a statement about EEG neurofeedback studied in isolation, which is how the literature has studied it. It is not a statement about the best available approach, and those two things come apart here more sharply than almost anywhere else in this book.
Monte-Silva et al. (2019) Electromyogram-related neuromuscular electrical stimulation for restoring wrist and hand movement in poststroke hemiplegia: A systematic review and meta-analysis doi:10.1177/1545968319826053
Rayegani et al. (2014) Effect of neurofeedback and electromyographic-biofeedback therapy on improving hand function in stroke patients doi:10.1310/tsr2102-137
Lirio-Romero et al. (2021) Electromyographic biofeedback improves upper extremity function: A randomized, single-blinded, controlled trial doi:10.1016/j.physio.2020.02.002
Wang et al. (2024) Electromyographic biofeedback therapy for improving limb function after stroke: A systematic review and meta-analysis doi:10.1371/journal.pone.0289572
Seo et al. (2025) Biofeedback Training for 3-Dimensional Finger Force Control to Improve Upper Limb Function Poststroke: An RCT doi:10.1161/STROKEAHA.125.050965
Butet et al. (2025) EEG-fMRI neurofeedback versus motor imagery after stroke, a randomized controlled trial doi:10.1186/s12984-025-01598-9
Alyanak et al. (2025) Efficacy of Game-Based EMG-Biofeedback Therapy in Post-Stroke Dysphagia: A Randomized Controlled Trial doi:10.1007/s00455-025-10819-1
