Who, how, and how much? A scoping review of brain-computer interface studies for upper-limb stroke rehabilitation

@misc{bousfiha2026much,
  title = {{Who, how, and how much? A scoping review of brain-computer interface studies for upper-limb stroke rehabilitation}},
  author = {Camile Bousfiha and Erin Soller and Marion Couton and Juliane Le Treis and Célestine Allombert-Blaise and Laurent Bougrain and Paolo Bartolomeo and Fabrizio de Vico Fallani},
  howpublished = {HAL},
  year = {2026},
  month = oct,
  url = {https://inria.hal.science/hal-05837835v1},
  note = {CC BY 4.0},
}

TY  - UNPB
TI  - Who, how, and how much? A scoping review of brain-computer interface studies for upper-limb stroke rehabilitation
AU  - Camile Bousfiha
AU  - Erin Soller
AU  - Marion Couton
AU  - Juliane Le Treis
AU  - Célestine Allombert-Blaise
AU  - Laurent Bougrain
AU  - Paolo Bartolomeo
AU  - Fabrizio de Vico Fallani
PY  - 2026
DA  - 2026/10/24
PB  - HAL
UR  - https://inria.hal.science/hal-05837835v1
AB  - Brain-computer interfaces (BCIs) are among the neurotechnologies proposed for upper-limb rehabilitation after stroke, particularly for patients with severe motor impairments. Although meta-analyses report moderate shortterm improvements in upper-limb motor function, heterogeneity in paradigms, instructions, feedback, dosage and patient characteristics limits comparability and clinical translation. We aimed to map BCI-based upper-limb rehabilitation after stroke to identify factors limiting comparability and clinical transferability. This scoping review followed JBI methodology and PRISMA-ScR guidelines. PubMed, Scopus, the Cochrane Library and IEEE Xplore were searched from the earliest available date to 15 April 2026 for longitudinal studies of closed-loop BCI in adults with stroke and upper-limb motor impairment. Data on population characteristics, task instructions, brain-signal decoding, feedback, protocol structure and dosage were charted and summarized without meta-analysis. Ninety-three studies met the search criteria, including 2397 stroke patients. Of these, 1619 patients received a BCI intervention, with the remainder assigned to control groups across 48 comparator studies. Studies predominantly included chronic patients with severe upper-limb motor impairment, with a median BCI group size of 12. Group-size-weighted control accuracy was 69.7% (20 studies), while 12.9% of patients did not reach the 60% threshold (23 studies). Motor imagery was the most common task (69 studies), followed by motor attempt ( 19) and action observation (5). However, 37 studies did not specify the task used and fewer reported the instructions provided. Sensorimotor rhythm was the most commonly decoded signal (69 studies). Feedback was mainly delivered through exoskeletons (44), virtual environments (32), functional electrical stimulation ( 27) and/or end-effector robots (13). A typical 1-hour session comprised a median of 80 trained-task trials performed over 17.5 minutes. Protocols consisted of a median of 14 sessions over 4 weeks, corresponding to 6.7 hours of BCI training. Rehabilitative BCI protocols are highly diverse, limiting interpretation of their therapeutic effect. Details on patient instructions, signal decoding and feedback modality were often missing, while BCI training dose remains limited relative to standard rehabilitation guidelines. More consistent and comparable BCI protocols are needed to support reliable evaluation of efficacy and translation into clinical practice.
ER  -