Comparative efficacy of brain-computer interface training with different feedback modalities for post-stroke upper-limb motor recovery: a systematic review and network meta-analysis of randomized controlled trials

Abstract

Brain-computer interface (BCI) training may improve post-stroke upper-limb recovery through closed-loop feedback, but the comparative efficacy of different feedback modalities remains uncertain. This study compared BCI training with different feedback modalities for post-stroke upper-limb motor impairment. PubMed, Embase, Web of Science, Cochrane Library, and Scopus were searched from inception through April 2026 for randomized controlled trials of BCI-based upper-limb rehabilitation after stroke. BCI interventions were classified by feedback modality and feedback contingency into multimodal feedback (BCI-MM), robotic or exoskeleton-assisted feedback (BCI-Robot), functional electrical stimulation feedback (BCI-FES), and visual feedback (BCI-visual), while comparator interventions were classified separately, with usual rehabilitation (UR) as the reference. Random-effects network meta-analysis estimated WMDs with 95% CIs. The primary outcome was Fugl-Meyer Assessment for Upper Extremity (FMA-UE); secondary outcomes were the Action Research Arm Test (ARAT) and Wolf Motor Function Test (WMFT). Risk of bias was assessed using the Cochrane Risk of Bias 2 tool, and certainty was assessed using CINeMA. Twenty-two trials involving 956 patients across subacute, chronic, and mixed post-stroke stages were included. The FMA-UE network contained 11 connected nodes, with no global or local inconsistency. Compared with UR, BCI-Robot showed greater FMA-UE improvement (WMD = 6.28, 95% CI 2.70–9.86), as did non-BCI robotic training (WMD = 4.26, 95% CI 0.42–8.10) and BCI-MM (WMD = 4.03, 95% CI 1.00–7.06). BCI-FES did not differ significantly from UR. BCI-visual had the highest SUCRA value (77.6%), but its comparison with UR was non-significant, with very low certainty of evidence; no BCI feedback modality was superior to another. Secondary outcomes showed greater ARAT improvement with BCI-Robot versus UR, whereas WMFT findings were non-significant. Certainty for the main FMA-UE comparisons was low or very low. BCI-Robot and BCI-MM showed statistically significant FMA-UE benefits versus UR, although certainty was low. As non-BCI robotic training also showed benefit, the BCI-Robot effect cannot be attributed to BCI alone. No BCI feedback modality was statistically superior to another. Further high-quality direct-comparison trials are needed. https://www.crd.york.ac.uk/PROSPERO/view/CRD420261370474 , Unique Identifier: CRD420261370474.