This two-arm, parallel RCT was conducted in compliance with the Declaration of Helsinki and registered with the Chinese Clinical Trial Registry (ChiCTR2400082786). The study received approval from the Ethics Committee of Wuxi Central Rehabilitation Hospital (WXMHCIRB2024LLky018) and adhered to the CONSORT reporting guidelines (Fig. 1, Supplementary Table S1). A total of 40 stroke patients were enrolled from the Rehabilitation Medicine Department of Wuxi Central Rehabilitation Hospital, China, between April and December 2024. Table 2 presents the demographic and clinical characteristics of the participants.

The inclusion criteria were as follows: (1) unilateral subcortical lesions in the right hemisphere confirmed by MRI or computed tomography; (2) aged between 30 and 80 years; (3) stroke onset within 2 weeks to 3 months; (4) persistent hemiplegia causing hand dysfunction, with Brunnstrom hand stages II-IV, and at least one of the following motor abilities: (a) partial upper limb function [Action Research Arm Test (ARAT) gross motor score ≥ 3]; (b) manual muscle test (MMT) score ≥ 2 on shoulder flexion and either elbow flexion or extension; or (c) active shoulder flexion or abduction through ≥ 50% of the passive range of motion in gravity-eliminated positions, combined with any of the following motions: elbow flexion, elbow extension, wrist flexion, wrist extension, finger flexion, or finger extension [17]. (5) Passive range of motion in the affected upper limb near normal; (6) Mini-Mental State Evaluation (MMSE) score > 20 [18], with the ability to understand and cooperate with assessments and training; (7) motor imagery ability assessed using the Kinesthetic and Visual Imagery Questionnaire (KVIQ-20), with a score of > 55 indicating sufficient proficiency [19]. (9) Written informed consent was obtained from the participants or their guardians.

The exclusion criteria included the following: (1) coexisting neurological disorders such as peripheral nerve injuries, Parkinson’s disease, or multiple sclerosis; (2) severe dysfunction or failure of major organs (heart, liver, kidney, etc.); (3) cognitive impairments, aphasia, or conditions preventing understanding or following instructions; (4) posterior circulation infarction; (5) severe upper limb spasticity (modified Ashworth ≥ 3) or joint contractures; (6) skin injuries, infections, or hypersensitivity to pain; and (7) pregnancy.

Due to the lack of prior reports on the efficacy of brain-computer interface-based soft robotic gloves for upper limb function in subacute stroke patients, it was impractical to determine an exact sample size a priori. Therefore, we set 18 participants per group (36 total), which aligns with the minimum recommended sample size for pilot trials [20]. Accounting for a potential 10% dropout rate, we plan to enroll 40 participants in this pilot study.

The participants were randomly assigned to either the BCI-SRG group or the soft robotic glove (SRG) group by one of the authors (L.X. or Y.X.), who was not involved in the intervention or evaluation and used a computer-generated, sealed-envelope method without adjustment factors. Figure 1 shows a schematic overview of the experimental procedure. Both groups received conventional upper-limb rehabilitation, including therapist-guided active and passive limb exercises, task-oriented exercises, ADL training, and neuromuscular electrical stimulation. Therapy was delivered for 60 min per day, 5 days per week. In addition to traditional rehabilitation, the BCI-SRG group received training with a soft robotic glove by a BCI, whereas the SRG group used the same glove operated by a soft robotic control system. Both groups underwent 20 intervention sessions over 4 weeks (5 sessions per week), with motor function assessments conducted at two time points: preintervention and 4 weeks postintervention. fNIRS was used to assess brain activity at baseline and after 4 weeks.

A 20-channel active electrode system (Neuracle Technology Co., Ltd., China) was used to acquire continuous EEG signals. During the acquisition, twenty channels (Fp1, Fp2, F7, F3, Fz, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, Oz, and O2 (Fig. 2b)) were selected, the sampling rate was 1000 Hz, and the electrode impedance was maintained below 10 kΩ to ensure signal quality. The signals were bandpassed with a 2–60 Hz filter and a notch filter (48–52 Hz) to remove artefacts and power line interference, respectively. The soft robotic glove module (Nanjing Reseader Medical Technology Co., Ltd., China) (Fig. 2c), comprising the robotic glove control box and the soft robotic glove, was responsible for providing mechanical feedback to the participants.

System design and protocol for BCI-SRG training. a A participant engaged in the BCI-SRG training session and wore the soft robotic glove while EEG signals were recorded. b EEG electrode placement according to the 10–20 system. c Soft robotic glove worn by the participant and equipped with sensors for feedback. d Configuration of fNIRS channels. e Task fNIRS measurement protocol, detailing a 5-cycle procedure of preparation, grasping task, rest, and end. L: left hemisphere; R: right hemisphere

To enable training-free online motor imagery classification, we initially employed mu suppression scores derived from the C3 and C4 channels to quantify the event-related desynchronization magnitude and established a threshold for online classification. The target channel data were transformed into the frequency domain using the Fast Fourier Transform, which enabled the calculation of energy within the mu band. Mu suppression was computed using the formula: Musupp = − (Muptask − Muprest)/Muprest*100%, where \(}_}}}\) represents the EEG energy during the MI task and where \(}_}}}\) denotes the EEG energy during the resting state.

Before training, the therapist provides detailed instructions on the treatment procedure. Following this, the patient is asked to focus on a black cross displayed on the screen for 60 s to collect resting-state EEG data. After 2 s of preparation, each session begins with an image of a cup displayed on the screen, prompting the patient to imagine grasping the cup with their affected hand. This task lasts for 5 s. If successful, a 15-s congratulatory animation is shown, and soft robotic hand assistance is triggered to aid the patient in performing the task. If unsuccessful, a 2-s encouraging animation is provided. After the animation, the patient rested for 5 s before the next trial. Each training session consisted of 75 trials, which were grouped into 15 rounds of 5 trials each. The average session duration in the BCI-SRG group was approximately 20–25 min, depending on the proportion of successful detections. The detailed procedure is illustrated in Fig. 3.

Flowchart of the BCI-SRG training protocol, outlining the sequence of tasks: preparation, motor imagery, task execution, and feedback

In the SRG training group, patients were seated comfortably in front of a screen wearing a soft robotic glove. Training consisted of 10-s gripping tasks followed by 5-s rest intervals, with a total of 75 repetitions, lasting 20 min in total. A left-hand grasping video was displayed on the screen, and patients were instructed to imagine performing the grasping motion with their left hand in synchronization with the video.

The primary outcome of this study was the ARAT score, a validated measurement of upper limb motor function with proven high reliability and validity in stroke patients [21]. The ARAT consists of 19 items across four subscales: grasp, grip, pinch, and gross movements. Each item is scored on a 4-point scale (0–3), with a maximum score of 57 for one upper limb. Higher scores indicate better function. Even if some upper limb motor function is present, a score of zero is assigned in this assessment if the hand is unable to perform grasping, gripping, or pinching tasks. The secondary outcomes included the FMA-UL and MBI scores. Additionally, cortical activation changes were evaluated using fNIRS to assess motor cortical activity at the first and final neurofeedback sessions.

A continuous-wave fNIRS system (NirSmart, Danyang Huichuang Medical Equipment Co., Ltd., Zhenjiang, China) with wavelengths of 730 and 850 nm was used to measure changes in the oxygenated haemoglobin (Δoxy-Hb) concentration, with a sampling rate of 11 Hz. fNIRS measurements were conducted by trained researchers (W.Z.). A total of 35 channels were set up, including 14 light sources and 14 detectors. The distance between each detector and the light source was 30 mm. On the basis of the standard Brodmann brain mapping method, the channels were divided into seven regions of interest (ROIs): the LDLPFC (9, 11, 12, 25, 26, 27), MPFC (6, 7, 8, 10, 22, 23, 24), RDLPFC (3, 4, 5, 19, 20, 21), LSMC (13, 14, 28, 29, 32, 33, 34, 35), and RSMC (1, 2, 15, 16, 17, 18, 30, 31). The montage configuration for the NIRS probe channels is shown in Fig. 2d and Table 1. Prior to the experiment, all the participants were instructed to sit still in a noise-free environment for 10–20 min to minimize haemodynamic responses caused by physical activity. A NIR gain quality check was conducted before each recording session to ensure optimal data acquisition, preventing both insufficient and excessive gain. To position the probes accurately on the scalp, the participants’ heads were covered with a cap and secured using straps to adjust and stabilize the emitters and receivers. The hair was cleared to maximize light‒tissue coupling and ensure close contact between the probes and the skin. fNIRS testing consisted of a baseline phase (30 s), a task phase (300 s), and an end phase (30 s). The task phase included five blocks, each comprising a 30-s grasp task followed by a 30-s rest period, for a total of 300 s (Fig. 2e).

Data inspection and preprocessing were performed using the Homer2 toolbox in MATLAB 2014a (The MathWorks Inc.), with the following steps: (1) Channel quality check: All channels from each participant were inspected for quality, and the results were recorded before preprocessing. (2) Signal conversion: Raw NIRS intensity values were converted to optical density signals. (3) Motion artefact removal: Motion artefacts induced by head movement during data acquisition were detected and corrected using the motion artefact removal algorithm with the following parameters: tMotion = 0.5, tMask = 2.0, STDEVthresh = 20.0, AMPthresh = 0.5, and pSpline = 0.99. (4) Spline interpolation: Following motion artefact detection, spline interpolation was applied to correct for any remaining artefacts. The detection window was then shifted to the next segment until the entire time series was processed. (5) Filtering: A bandpass filter (0.01–0.1 Hz) was applied to remove physiological noise and drift, including heart rate (~ 1 Hz), respiration (0.2–0.3 Hz), and Mayer waves (~ 0.1 Hz) [22]. (6) Conversion to HbO concentration: Applying the modified Beer‒Lambert law, the filtered optical density data were converted to HbO concentrations [23].

Data analysis was performed using SPSS 22.0 (IBM SPSS Statistics, Chicago, IL, USA) for statistical analysis and GraphPad Prism 8 (GraphPad Software Inc., San Diego, CA, USA) for visualization. The normality of the data distribution was assessed using the Shapiro‒Wilk test. Demographic data are presented as the means ± standard deviations (continuous variables were analysed via independent-sample t tests) and counts (categorical variables were analysed by the chi-square test). Nonnormally distributed data (ARAT, FMA-UL, and MBI) are expressed as the median results and quartiles. The Wilcoxon rank sum test was used for intragroup comparisons, and the Mann‒Whitney U test was used for intergroup comparisons.

fNIRS analysis was performed with MATLAB. HbO changes during the grasping task were analysed by extracting signals from each channel on the basis of predefined markers. The mean HbO values across five grasping trials were calculated. Intragroup contrasts (post–pre) were assessed using paired t tests, whereas between-group contrasts of change scores (post minus pre) were analysed using independent t tests. Multiple comparisons were corrected by the Benjamini‒Hochberg procedure to control the false discovery rate (FDR), with significance set at P < 0.05. For visualization of brain activation, t statistic maps were generated using BrainNet Viewer. Significant t values were retained, whereas nonsignificant values were set to zero. The adjusted t values were mapped onto brain models, with colour gradients reflecting the magnitude of activation changes. Spearman correlation analyses were performed to examine the relationships between changes in HbO levels (post–pre) during grasping tasks and improvements in the ARAT score within each group. Statistical significance was set at P < 0.05.