Multimodal assessment of a BCI system for stroke rehabilitation integrating motor imagery and motor attempts: a randomized controlled trial

Abstract

Background

Brain–computer interface (BCI) technology based on motor imagery (MI) or motor attempt (MA) has shown promise in enhancing motor function recovery in stroke patients. This study aimed to evaluate the effectiveness of BCI-based rehabilitation in improving motor function through multimodal assessment, and to explore the potential neuroplastic changes resulting from this intervention.

Methods

We conducted a randomized double-blind controlled clinical trial with multimodal assessment to evaluate the efficacy of a BCI system for enhancing motor recovery. A total of 48 ischemic stroke patients completed the study (25 BCI, 23 control). The BCI group used an 8-electrode electroencephalogram (EEG) system, a virtual reality training module, and a rehabilitation training robot for real-time motor intention-based feedback. The control group used identical BCI devices but without displaying real-time data and feedback. Participants underwent 20-minute upper and lower limb training sessions for two weeks. Motor function (Fugl-Meyer Extremity scale), electromyography (EMG), and functional near-infrared spectroscopy (fNIRS) were assessed pre- and post-intervention.

Results

The BCI group demonstrated significantly greater improvement in upper extremity motor function compared to the control group (ΔFMA-UE: 4.0 vs. 2.0, p = 0.046). EEG results of the BCI group showed a significant decrease in both DAR (p = 0.031) and DABR (p < 0.001) compared to baseline. EMG analysis revealed that BCI treatment resulted in significant increases in deltoid and bicipital muscle activity during both shoulder and elbow flexion movements compared to baseline (p < 0.01). fNIRS results indicated enhanced functional connectivity and activation in key motor-related brain regions, including the prefrontal cortex, supplementary motor area, and primary motor cortex in the BCI group.

Conclusion

BCI-based rehabilitation using an attention-motor dual-task paradigm significantly improved upper limb motor function and enhanced motor and cognitive network activity in stroke patients. Multimodal assessment supports the potential of BCI rehabilitation as an effective tool for leveraging neuroplasticity and promoting motor recovery.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12984-025-01723-8.

Introduction

Stroke, a debilitating neurological disorder, results in motor function impairments in approximately 60% of patients [1]. Rehabilitation of motor dysfunction following a stroke remains a critical focus, yet even with multimodal treatment approaches, 40% of stroke survivors continue to experience varying levels of disability [2]. Among these, the inability to fully recover function in the upper limb significantly impacts not only patient independence but also brings a substantial economic burden to families and society [3]. Therefore, the need for effective stroke rehabilitation, especially the upper limbs, is critical in the long-term continuum of stroke care and will remain a priority for the foreseeable future.

The Brain–computer interface (BCI) technology based on motor imagery (MI) or motor attempt (MA) has emerged as a promising enhancement to conventional rehabilitation approaches in stroke [4–6]. Both MA and MI tasks can activate sensorimotor areas similar to real motion thereby aiding in motor function recovery after stroke [7–9]. Non-invasive BCI methods can achieve the effect of rehabilitation training by collecting the EEG signal of the patients’ MI or MA, using computer systems to decode the patients’ brain activity and utilizing external devices to assist limb movements [10, 11]. Previous randomized controlled trials (RCTs) have shown that MI-BCI systems significantly improved motor function assessments in more than 60% of stroke patients, demonstrating their potential to restore motor function by inducing activity-dependent brain plasticity [6, 12, 13].

Compared to MI, MA is reported as a more effective, natural and verifiable trigger task in BCI training, and is more easily detected in EEG for patients paralyzed in the upper limbs after a stroke [11, 14]. MA can address concerns of task specificity that often arise in MI paradigms, which makes maintaining consistent focus on the motor task can be challenging, particularly for those with cognitive impairments [11]. Studies on motor-impaired patients have revealed both differences and similarities in cortical activations and sensorimotor integration during MI and MA tasks, with activating distinct neural circuits [11, 15] while partially overlapping in premotor and parietal areas, basal ganglia and cerebellum. These findings underscore the importance of integrating both attempted and imagined movements in neurofeedback training for motor promoting or restoring. In this study, we integrate MI and MA tasks within the BCI system to enhance motor function in stroke patients. The BCI simultaneously generates descending control signals and ascending sensory feedback (bidirectional stimulation), effectively stimulating the relevant brain areas [16–18]. MI involves the mental simulation of movement, while MA requires the patient to actively attempt a movement [11, 19]. Both tasks engage brain regions involved in motor control, such as the motor cortex, and are interpreted by the BCI system through EEG signals. The system decodes these signals in real-time, providing feedback to the patient, which enhances their motor intentions and supports motor recovery.

Despite promising results achieved in early-stage clinical trials, BCI-based stroke rehabilitation is still an emerging field, and its therapeutic mechanisms remain largely unclear [20–22]. Furthermore, these systems often demand sustained concentration during training, which can be particularly challenging for cognitively impaired individuals [23]. Cognitive impairment affects 25–80% of stroke patients and may significantly limit the effectiveness of BCI-based treatments, especially for those with attention deficits [24]. Given that the amount of mental workload is an important issue in motor learning [25], increasing the appeal of training and providing more effective feedback may be an effective enhancement for BCI rehabilitation systems. Previous studies have highlighted the crucial role of attention in BCI, showing that BCI training can improve motor function, primarily through the activation of alpha waves associated with attention [26, 27]. Chung et al. [28, 29] have demonstrated that BCI-guided treatment outperforms conventional approaches by increasing patients’ attention indices in the Fp1 and Fp2 regions, and these indicators can reflect the motor intention of the subjects to some extent. Thus, we applied an attention-motor dual-task paradigm to enhance patients’ attention and motor intention during BCI training, by monitoring the comprehensive level of attention in performing task.

In stroke rehabilitation, multimodal assessment methods are crucial for providing a comprehensive understanding of motor recovery. These methods integrate various measurement techniques, such as neuroimaging (e.g., fMRI, EEG), near-infrared spectroscopy (NIRS), electromyography (EMG), clinical evaluations (e.g., the Fugl-Meyer Assessment), and sensor-based motion analysis. NIRS allows for non-invasive monitoring of cerebral oxygenation and hemodynamics during motor tasks, providing a visual monitoring tool for post-stroke rehabilitation and a comprehensive understanding of the underlying mechanisms [30]. Studies have used NIRS to assess progressive brain plasticity after a stroke by evaluating the time course of hemodynamic patterns and changes in the interhemispheric balance between the cortical sensorimotor areas [31, 32]. By monitoring stroke-related changes in neuromuscular coordination and motor unit firing rates, EMG can be used for assessing muscle fatigue and detecting abnormal muscle activity [33, 34]. By combining data from these diverse sources, researchers and clinicians can gain deeper insights into the underlying neural mechanisms of recovery, assess the effectiveness of interventions, and tailor rehabilitation strategies to individual patient needs.

In this study, we aimed to evaluate the clinical practicability and effectiveness of our BCI-based rehabilitation system for stroke patients by employing a comprehensive multimodal approach, and investigate its potential mechanism. Our hypothesis is that this system utilizing prefrontal EEG signals for real-time feedback could enhance subjects’ motor intentions and cognitive engagement by integrating both MI and MA training, thereby improving motor function during rehabilitation. To this end, we conducted a randomized placebo-controlled blinded-endpoint clinical trial and applied a BCI-controlled pedaling training system.

Methods

This study employed a randomized double-blind controlled clinical trial design. Randomization was achieved using a coated random number table managed by an independent third party. Participants were randomly assigned to either the experimental group (BCI group) or the control group in a 1:1 ratio. The BCI group and the control group used identical BCI devices. The control group’s device recorded EEG signals without providing feedback, due to the training process showing its pre-trained exercise data rather than using real-time exercise data. This study was approved by the Ethics Committee of the First Affiliated Hospital of Xi’an Jiaotong University (No. XJTU1AF2022LSK-422), and all patients have signed informed consent.

Subject

The study involved patients with subcortical ischemic stroke confirmed by MRI, accompanied by hemiplegia in the rehabilitation center. The inclusion criteria were as follows: (1) Aged 35–79 years old; (2) Patients with first subcortical ischemic stroke onset from 2 weeks to 3 months; (3) Hemiplegia, and the muscle strength of the proximal upper limb was 1–3; (4) Right-handed; (5) Sitting balance level 1 or above (assessed by a 3-level balance scale, including static, self-dynamic, and external dynamic balance), was able to cooperate with assessment and treatment. Exclusion criteria: (1) Severely impaired cognition (Mini-Mental State Examination [MMSE] score < 20), unable to pay attention to and understand screen information; (2) Severe pain and limited mobility of the limbs. From November 2022 to January 2024, 90 stroke patients were initially enrolled and 38 were excluded for not meeting inclusion criteria. A total of 52 patients were randomized into the BCI group and the control group. Ultimately, 48 patients completed the 21-day BCI rehabilitation treatment (25 in the BCI group and 23 in the control group) (Fig. 1).

Fig. 1.

CONSORT flow diagram

Control treatment

The control group used the identical equipment and training settings but received simulated feedback from their pre-recorded EEG data rather than their real-time EEG. Participants imagined cycling movements with both their upper and lower limbs while EEG data was recorded, but the EEG data was not used to control the BCI machine. In contrast to real-time neurofeedback paradigms, the motor intention and cycling speed displayed during training sessions were generated from pre-existing datasets, thereby eliminating real-time performance feedback throughout the experimental protocol. Both the BCI group and the control group received standard rehabilitation treatments, including acupuncture, physiotherapy therapy, occupational therapy and electronic biofeedback. The details regarding the type, intensity, and duration of the additional rehabilitation interventions are provided in the supplementary materials.

Brain–computer interface and intervention

The BCI rehabilitation system structure is depicted in Fig. 2A. The system consists of three components: (1) an 8-electrodes EEG collection system (Zhejiang Mailian Medical Technology Co., Ltd., Hangzhou, China) for scalp EEG data acquisition; (2) A virtual reality training system that presents a motivating game interface and converts EEG data into an motor intention score(Mscore), a quantitative metric reflecting motor intention level(as detailed in the supplementary materials); and (3) a pedaling training robot (Mailian, China) driving limb movements in a circular motion. The robot’s speed was controlled by the Mscore calculated from Fp1 EEG signals, serving as the BCI’s regulatory target [17].

Fig. 2.

BCI rehabilitation system introduction and experimental design. A The BCI system comprises an 8-electrode EEG collection setup, a virtual reality training system, and a pedaling training robot. Participants performed BCI-controlled motor imagery and cycling training for both upper and lower limbs. In the BCI group, motor intentions were visualized as a real-time Mscore on a monitor, with assistance provided when the score exceeded a set threshold, increasing with higher scores. In the control group, the Mscore was displayed as their pre-training score, with no feedback provided. B Left: Diagram of the subject undergoing BCI rehabilitation exercises. Upper right: The experimental protocol for limb exercise and rest. Lower right: The distribution of EEG electrodes (A1 and A2 as reference and ground) and fNIRS electrodes consists of 16 light sources (represented by purple circles) and 16 detectors (represented by blue circles), forming a total of 48 measurement channels (represented by gray shaded regions). C The experiment of rest and motor task. During motor task, EMG and fNIRs spectroscopy signals were simultaneously recorded. During the rest task, participants sat quietly with their eyes open and fNIRs signals were collected

Participants engaged in BCI-controlled motor imagery and cycling training for both upper and lower limbs. For participants with insufficient wrist strength, elastic bands were applied around their wrists during the training to ensure they could complete the cycling training effectively. They were instructed to vividly imagine cycling movements, with the intensity of their motor intentions reflected by the real-time Mscore displayed on a monitor. In BCI group, when the Mscore exceeded a predefined threshold, the cycling machine provided assistance. Higher Mscores increased the assistance and speed of cycling, while lower Mscores triggered auditory cues and reduced assistance. Each training session included visual feedback, synchronizing the on-screen pedaling motion with the patient’s actual pedaling, and displaying the patient’s Mscore in real-time via a progress bar and numbers.

Experimental design

The experimental design is depicted in Fig. 2B. Both the BCI group and the control group participated in daily 20-minute pedaling training sessions for upper and lower limbs, conducted five days a week over a two-week period, maintaining consistent training intensity. As shown in Fig. 2C, participants performed a motor task and a rest task at the beginning and end of the study. In the motor task, participants performed structured shoulder and elbow flexion movements with both the affected and healthy arms, respectively. This involved alternating 30 s of rest with 30 s of movement tasks across four blocks. During these movement tasks, electromyography (EMG) and near-infrared (NIR) spectroscopy signals were simultaneously recorded. During the rest task, participants sat quietly with their eyes open and NIR signals were collected.

Functional and behavioral assessment

Motor function assessment

The Fugl Meyer Assessment of Upper Extremity (FMA-UE) and lower Extremity (FMA-LE) was used to assess movement disorders [35]. FMA-UE ranges from 0 to 66 points, and FMA-LE from 0 to 34. Lower scores indicate greater impairment in limb functions, making it a standard measure for evaluating stroke-related damage and monitoring therapy progress. Motor function was assessed once before and after treatment, and these evaluations were conducted by a therapist blinded to treatment allocation.

EEG data acquisition and analysis during BCI training

EEG signals were acquired using an EEG cap equipped with two ear clip electrodes (one reference and one bias) and eight wet (saline) electrodes (Fig. 2B). During BCI real-time training feedback, adaptive filters were used to remove power-frequency interference at 50 Hz. A 5th-order IIR bandpass filter with a range of 0.5 Hz to 30 Hz was applied to eliminate low-frequency artifacts (such as eye movements or breathing) and high-frequency noise.

Continuous EEG data segments were collected during 10 BCI training sessions. These EEG data segments were then divided into two phases: the first 5 training sessions (stage 1) and the late 5 training sessions (stage 2). For the analysis phase, additional notch filtering and bandpass filtering were applied to the EEG data. A sliding window approach was used to calculate power. The analysis of EEG data focused on calculating the power spectrum energy ratios from the FP1 electrode [17]. Three specific ratios were computed: the delta/alpha power ratio (DAR), the delta/(alpha + beta) power ratio (DABR), and the (delta + theta)/(alpha + beta) power ratio (DTABR) [17]. Finally, the mean values of these metrics were calculated separately for the early (stage 1) and late (stage 2) training stages, allowing for comparative analysis between the two phases.

EMG data acquisition and analysis

EMG data were collected using electrodes placed on bilateral upper limbs (biceps brachii and anterior deltoid). Three EMG indices were calculated: (1) root mean square (RMS): reflects the average amplitude of the EMG signal; (2) integrated EMG (iEMG): represents the total electrical energy of the EMG signal; (3) averaged EMG (aEMG): indicates the average amplitude of the EMG signal over time [36].

fNIRS data acquisition and analysis

fNIRS measurements were obtained using the NirScan 6000 C (Danyang Huichuang Medical Equipment Co., Ltd., China), which consists of 16 light sources and 16 detectors, forming a total of 48 measurement channels. The 48 channels were positioned over the main brain regions, including primary somatosensory cortex (S1), primary motor cortex (M1), somatosensory association cortex (SAC), premotor cortex (PMC), supplementary motor area (SMA), frontal eye fields (FEF), and dorsolateral prefrontal cortex (DLPFC). The sampling frequency was 11 Hz and the wavelength was 730 nm and 850 nm. The coordinates of each optical pole on the Montreal Neurological Institute (MNI) brain template were determined using the 10/20 international standard lead system, combined with the coordinate positioning function of the SPM software, with reference to the coordinates and optical pole positions. Because we included patients with both right- and left-sided lesions, all imaging data from patients with a left-sided lesion were flipped horizontally before data analysis, so the affected hemisphere formed the right side of the image.

For fNIRS signals, a fourth-order Butterworth bandpass filter (0.01–0.1 Hz) is initially applied to eliminate artifacts, including cardiac interference. The motion artifacts were removed by spline interpolation, and the raw light intensity was converted to changes in the concentration of HbO₂ and HbR using the modified Beer–Lambert law. Given that HbO is more reproducible and stable over time, and shows the highest correlation with fMRI BOLD measures, the results will concentrate on HbO findings [37]. The preprocessed fNIRS data were imported to the NIRS-KIT toolbox based on the MATLAB environment for individual-level analysis [38]. In this study, Pearson correlation was used approaches for measuring functional connectivity (FC) in fNIRS channel in resting state data. T-tests or paired tests were applied to examine differences in FC. A general linear model (GLM) was used to evaluate channel-wise task-evoked neural activation for each participant. To assess differences activation of cortex after treatment between BCI group and control group during the motor task, t-tests were applied to identify channels with significant variations in brain activity. To control for multiple comparisons, the results were corrected using the false discovery rate (FDR) method (p < 0.05) [39].

Statistical analysis

In this study, we used the G-power (v3.1.9.2; Franz Faul, University of Kiel, Kiel, Germany) for calculation of the sample size. Based on the pre-experimental results and previous research results. The expected change in Fugl-Meyer Assessment of Upper Extremity Motor Function (ΔFMA-UE) was set at 6 points for the BCI group and 3 points for the control group. The standard deviation was 3.5 points, with α = 0.05 (two-sided) and power (1-β) = 0.8. At least 44 patients were needed in order to make an adequate group size, thus a sample size of 22 subjects per group. The data were analyzed using SPSS version 20 (SPSS, Inc., Chicago, IL, USA). Statistical comparisons between groups utilized t-tests, chi-square tests, or Kruskal-Wallis rank-sum tests as appropriate. Within-group comparisons before and after treatment employed paired t-tests or Wilcoxon signed-rank tests. For the comparison of EEG and EMG measurement, baseline-correct was performed by calculating the absolute difference between values. The significance level was set at p < 0.05. Multivariate regression analysis was used to compare the primary outcome between the two groups after adjusting covariates. Covariates were selected based on clinical experience, previous literature (e.g. age), and their associations with the outcomes of interest or eliciting a > 10% change in the effect estimate (α = 0.05, two sides) [40, 41].

Results

Demographic information

The basic demographic and clinical information of the recruited subjects is listed in Table 1. There were no significant between-group differences in the characteristics of patients, including age, sex, time poststroke, stroke type and FMA at baseline (p > 0.05).

Table 1.

Baseline characteristics of patients

	BCI	Control	P-value
N	25	23
Age, years	58.6 ± 12.8	58.3 ± 12.8	0.928a
Sex			0.145b
Female	5 (20.0%)	9 (39.1%)
Hemiplegia			0.571b
Left	11 (44.0%)	12 (52.2%)
Right	14 (56.0%)	11 (47.8%)
Time from onset, days	33.3 ± 18.5	41.8 ± 19.1	0.124a
mAS-elbow flexion			0.546b
0	6 (24.0%)	4 (17.4%)
1	8 (32.0%)	11 (47.8%)
1+	10 (40.0%)	6 (26.1%)
2	1 (4.0%)	2 (8.7%)
mAS-elbow extension			0.133b
0	18 (72.0%)	15 (65.2%)
1	6 (24.0%)	4 (17.4%)
1+	0 (0.0%)	4 (17.4%)
2	1 (4.0%)	0 (0.0%)
FMA-UE	23.0 (11.0–46.0)	18.0 (11.0–29.0)	0.634c
Upper limb	18.0 (11.0–26.0)	14.0 (11.0-20.5)	0.944c
Wrist	0.0 (0.0–2.0)	0.0 (0.0–1.0)	0.733c
Hand	2.0 (0.0–9.0)	1.0 (0.0-3.5)	0.901c
FMA-LE	18.0 (11.0–26.0)	14.0 (11.0-20.5)	0.664c
Upper-speed-Pre (rpm)	42.52(7.63–62.29)	45.10(7.73–57.74)	0.655 a
Upper-speed-Post (rpm)	49.48 (4.98–61.44)	48.68(7.96–65.17)	0.882 a
Lower-speed-Pre (rpm)	33.77(3.91–45.08)	30.53(5.66–40.88)	0.356 a
Lower-speed-Post (rpm)	35.21(5.67–45.17)	35.66(29.69–43.13)	0.850 a

Values are presented as mean ± SD, median (Q1-–3) or N (%)

FMA-UE Fugl-Meyer assessment of upper extremity motor function, FMA-LE Fugl-Meyer assessment of lower extremity motor function, mAS modified Ashworth Scale, rpm revolutions per minute.

^aEvaluated using t test. ^bEvaluated using chi-square test. ^cEvaluated using Kruskal-Wallis rank-sum test

Comparison of the motor function before and after treatment

After treatment, the median FMA-UE score in the BCI group improved to 27.0 points (Q1–Q3, 14.0–50.0), compared to 22.0 points (Q1–Q3, 13.0–30.5) in the control group. Specifically, proximal upper extremity scores were 23.0 (Q1–Q3, 13.0–30.0) and 17.0 (Q1–Q3, 12.5–22.5), respectively, while wrist and hand scores did not show significant changes. The median dFMA-UE was 4.0 (Q1–Q3, 3.0–6.0) for the BCI group and 2.0 (Q1–Q3, 1.0–4.0) for the control group. Adjusting for age, time of onset, and mAS of elbow extension, the BCI group demonstrated an additional improvement of 2.8 points (95% CI: 0.1, 5.4; p = 0.046) in FMA-UE compared to the control group (Table 2). For FMA-LE, the median scores after treatment were 24.0 points (Q1–Q3, 18.0–29.0) in the BCI group and 23.0 points (Q1–Q3, 18.5–28.0) in the control group. The BCI group showed a minimal improvement of 0.5 points (95% CI: − 0.6, 1.5; p = 0.392) compared to the control group (Table 2).

Table 2.

Multivariate regression analysis of motor function scales between treatment groups

	β (95%CI) P-value
	Adjust Ia	Adjust II
dFMA-UE	3.2 (0.7, 5.7) 0.015	2.8 (0.1, 5.4) 0.046b
Upper limb	1.7 (0.2, 3.1, 0.2) 0.029	1.6 (0.0, 3.1) 0.050b
Wrist	0.4 (-0.3, 1.1) 0.242	0.4 (− 0.4, 1.1) 0.314b
Hand	0.8 (-0.1, 1.7) 0.072	0.6 (− 0.3, 1.6) 0.181b
dFMA-LE	0.8 (-0.3, 1.9) 0.146	0.5 (− 0.6, 1.5) 0.392c

FMA-UE Fugl-Meyer assessment of upper extremity motor function, FMA-LE Fugl-Meyer assessment of lower extremity motor function

^aAdjusted for: age; ^bAdjusted for: age, time of onset and mAS of elbow extension; ^cAdjusted for: age, sex, and time of onset

Comparison of the EEG assessments

During BCI treatment Stage 1, there were no significant differences between the BCI and control groups in any of the EEG features (Fig. 3). During BCI treatment Stage 2, the BCI group had significantly lower mean DAR (p = 0.031) and DABR (p < 0.001) compared to the control group. The BCI group showed a significant decrease in both DAR (p = 0.009) and DABR (p < 0.001) compared to Stage 1, while the control group showed no significant change in any of the three indicators.

Fig. 3.

Comparison of EEG assessments (DAR, DTABR and DABR values; baseline correction) between the control and BCI groups for two treatment states during Upper-limb exercise *p < 0.05 indicates significant differences by t-tests or Wilcoxon signed-rank tests

Comparison of the EMG assessments

In the EMG assessments of the paretic side deltoid muscle activity, BCI treatment exhibited significant increases in RMS, iEMG, and aEMG compared to baseline during both shoulder and elbow flexion movements (Fig. 4A, B). During shoulder flexion, the control group show no significant changes in any EMG parameters. While during elbow flexion, the control group exhibited significant increases in EMG parameters compare to pre-treatment (Fig. 4A, B). In the EMG assessments of the paretic side bicipital muscle activity, the muscle activity significantly increased in BCI group (Fig. 4C, D). In contrast, the control group show no significant changes in any EMG parameters (Fig. 4C, D).

Fig. 4.

Comparison of pre- and post-treatment EMG assessments (RMS, iEMG and aEMG values; baseline correction) between the control and BCI groups for different muscles during different movements during Upper-limb exercise. *p < 0.05 and **p < 0.005 indicate significant differences by paired t-tests

fNIRS results for rest and motor task

The results demonstrated significant alterations in functional connectivity between key brain regions following BCI-based rehabilitation in stroke patients. The BCI group results showed notable changes in multiple brain regions after treatment, including bilateral prefrontal cortex, supplementary motor area, primary somatosensory cortex, and the bilateral dorsal parietal cortex (Fig. 5A). The changes in the control group were primarily localized in fewer regions compared to the BCI group, notably involving the left supplementary motor area (SMA.L), left primary motor cortex (M1.L), and right dorsal parietal cortex (DPC.R) (Fig. 5B). The results of the BCI group and control group after treatment showed that the BCI intervention elicited a more pronounced reorganization of functional networks, particularly involving the SMA.L, M1.L and the right prefrontal cortex (PFC.R) (Fig. 5C).

Fig. 5.

Different pattern of FCs (t-test or pair t-test, pFDR < 0.05) for between and within groups. The size of the node indicates the degree of the between and within groups differences. A Functional connectivity difference before and after treatment in BCI group. B Functional connectivity difference before and after treatment in control group. C Differences in functional connectivity after treatment between BCI group and control group

After flipping the NIR data of patients with left hemiplegia, the results showed that the BCI group had increased activation in key motor-related areas of the left hemisphere when performing right side movements. During left shoulder movement, the BCI group showed significant activation in dorsal parietal cortex (DPC.R), prefrontal cortex (PFC.L, PFC.R) and primary motor areas (M1.L, M1.R) (Fig. 6A). During right shoulder movement, the BCI group exhibited higher activation in prefrontal cortex (PFC.L, PFC.R), supplementary motor area (SMA.L, SMA.R) and primary motor areas (M1.R), while activation in activation in the M1.L was lower compared to the control group (Fig. 6B). For left elbow movement, the BCI group demonstrated significantly increased activation in dorsal parietal cortex (DPC.R), prefrontal cortex (PFC.R) and primary somatosensory cortex (S1.R) (Fig. 6C). Figure 6D illustrates the differences in fNIRS activation during right elbow movement, greater activation was observed in the BCI group, especially in the prefrontal cortex (PFC.L, PFC.R), dorsal parietal cortex (DPC.R) and primary motor cortex (M1.R).

Fig. 6.

Different activation of cortex during motor tasks (t-map). A Differences in activation of cortex after treatment between BCI group and control group during left shoulder motion. B Differences in activation of cortex after treatment between BCI group and control group during right shoulder motion. C Differences in activation of cortex after treatment between BCI group and control group during left elbow motion. D Differences in activation of cortex after treatment between BCI group and control group during right elbow motion

Discussion

Using a multimodal assessment approach, this study aimed to explore the effects and underlying mechanism of BCI-based neuromodulation rehabilitation on improving upper limb motor function in patients with stroke. We found that the BCI group, which underwent the attention-motor dual-task paradigm training, showed a significant improvement in upper limb motor function compared to the control group. The results of EMG analysis showed a stronger muscle contraction recruitment ability and compensatory activation of adjacent muscles in the BCI group. EEG signals revealed the decreased DAR, DABR, and DTABR in prefrontal lobe activity in BCI group. Furthermore, fNIRS imaging revealed a more significant activation of the lesion-side motor area and enhanced functional connectivity with the prefrontal lobe, supplementary motor area, sensory area, and contralateral hemisphere in the BCI group after treatment. These findings suggest that our BCI rehabilitation system combining MI and MA can enhance brain cognitive and motor area activity, increase inter- and intra-hemispheric compensation, and promote motor function recovery after stroke.

Consistent with our previous study, this study found that, compared to pedaling training alone, our BCI system enhanced patient focus by incorporating multiple feedback modalities, leading to further improvements in upper limb motor function in stroke patients [42]. Neuroplasticity underpins all post-stroke rehabilitation interventions [43], and three potential neuroplasticity mechanisms were engaged in our rehabilitation system. The first mechanism was neurofeedback training [44], where patients received real-time visualization of their pedaling action and attention index, allowing them to consciously regulate their motor actions and cognitive activity. The second mechanism involved operant conditioning [45], with rewards given for high participation and additional feedback provided during lower engagement. These two mechanisms likely explain why patient’s attention and motor intention enhanced in BCI-based rehabilitation systems, and the higher concentration in turn induces more effective feedback [28, 46]. Finally, the most important mechanism was that our BCIs establish a neural circuit aligned with Hebbian plasticity theory [47, 48]which involves activating central intentions, executing peripheral movements, and generating proprioceptive feedback. Therefore, by incorporating multiple feedback modalities such as visual, auditory, and somatosensory cues during motor imagery and attempting, our system provides valuable insights for developing BCI-based smart home or community rehabilitation tool [22].

Several studies highlight the importance of EEG feature selection for optimizing BCI rehabilitation in stroke patients [49–52]. Al-Qazzaz et al. [50] propose a new feature fusion framework for motor MI-based BCI, suggesting that integrating information from various EEG features can enhance BCI performance in stroke rehabilitation. Based on the causal interactions between the channels of EEG signal, Varsehi and Firoozabadi [51] introduce Granger causality as a method for selecting informative EEG channels, potentially leading to more efficient BCIs for MI. In our previous research, we identified a negative correlation between power ratio measures (DAR, DABR, DTABR) in the Fp1 and post-stroke motor function [17]. Accordingly, the current study utilized prefrontal EEG signals as a key modulation in the BCI training system, and we also found BCI group exhibited significant enhanced EEG power ratios in Fp1 area compared to controls after treatment. Our findings suggested that the Mscore calculated from these power ratios can serve as an effective target for BCI modulation.

By collecting the EMG during participants performing shoulder and elbow flexion movements before and after BCI training, we investigated the change of muscle activation in stroke-affected limbs. EMG signals are increasingly utilized as a critical physiological marker to assess motion paralysis features in post-stroke patients [53]. Stroke survivors often exhibit co-activation of the anterior deltoid with biceps (flexor synergy), and posterior deltoid with triceps (extensor synergy), leading to restrictive and stereotypical movement patterns [54]. Previous studies have shown that training with a myoelectric-computer interface can reduce abnormal co-activation, restore more normal activation patterns, and improve upper limb function [54]. Similarly, our results showed a significant increase in EMG measurements of the biceps brachii and upper trapezius muscles on the paretic side in the BCI group after two weeks of training. We can infer that BCI training may promote more focused and efficient muscle use in targeted movements. Interestingly, we also found a notable increase in compensatory activity of the anterior deltoid during elbow flexion, which was absent in the control group. This suggests that BCI-based rehabilitation may encourage the engagement of compensatory strategies not typically seen with conventional therapy, potentially enhancing the overall motor output of the impaired limb. Our findings suggest that BCI can significantly improve motor function in stroke patients by enhancing targeted muscle activity and compensatory motor strategies, a result that supports the continued development and refinement of BCI technologies in rehabilitation.

Our fNIRS results revealed enhanced functional connectivity between motor-related brain regions after BCI-based treatment, particularly between the S1 and other key regions involved in motor control, such as the SMA and dorsal parietal cortex (DPC). Previous studies have shown that S1 connectivity might be an important marker in explaining stroke outcome, and S1 and SMA play critical roles in generating motor intentions and processing sensory feedback [55–57]. In our results, the enhanced connections between these regions post-treatment likely reflect improved coordination and integration of motor and sensory processes, critical for motor function recovery in stroke patients. Moreover, enhanced activity in the PFC has been widely reported during dual-task BCI training, indicating that the brain’s need for additional executive functions to manage inter-task interference training [58–60]. Consistent with these studies, our results suggest that by using attention-motor dual-task BCI training, our system could be implemented to enhance plasticity in the motor cortex, leading to improved recovery of motor function. Comparison with control group, we found the enhanced connectivity of PFC, SMA and M1 in BCI group after treatment, which is in line with the increased connection within BCI group. These results further demonstrate that our BCI-based rehabilitation enhances functional connectivity in key brain regions associated with both motor control and cognitive function, outperforming traditional therapy. In Fig. 6, similar patterns inter-group differences were observed in the activation maps of key motor and cognitive-related brain areas (PFC, SMA, S1 and M1) during movement tasks. Hence, our findings underscore the potential of BCI systems to leverage neuroplasticity and improve motor recovery in stroke patients by strengthening motor and cognitive networks.

There were some limitations of this study. Firstly, to ensure the reliability of the study’s findings, only patients with subcortical infarcts were included. The effectiveness of the BCI rehabilitation system in other stroke types remains to be validated. Secondly, one session of a real or sham BCI rehabilitation training for the lower limb was conducted daily to promote upper-lower limb coordination rehabilitation. While there was a trend of improvement in lower limb motor function, it was not statistically significant. This could be attributed to the study’s focus on upper limb motor function as an inclusion criterion and may require further validation with a longer intervention period. Thirdly, covariate screening for multivariate regression analysis did not find FMA-UE and FMA-LE eliciting a greater than 10% change in the effect estimate.

Conclusions

By combining MI and MA tasks in the attention-motor dual-task paradigm, our BCI rehabilitation system can effectively promote motor function recovery for stroke patients. A multimodal assessment approach involving motor function evaluation, EMG, and fNIRS demonstrated significant improvements in upper limb motor function in the BCI group compared to the control group. Enhanced muscle activation during shoulder and elbow flexion was observed after BCI-based treatment. EEG and fNIRS analyses revealed improved brain activity regulation and increased functional connectivity in motor and cognitive-related brain regions. Our BCI system, based on dual-task with attention feedback, could leverage neuroplasticity and improve motor recovery in stroke patients by strengthening motor and cognitive networks.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization: QJ, CZZ. Investigation: QJ, HJ. Methodology: HJ, YZW. Writing – Original Draft: HJ, YZW. Writing – Review & Editing: HJ, QJ. Visualization: QL, XH. Supervision: CZZ. Funding Acquisition: QJ, HJ. Data Curation, QL, XH, GJ, ZD, CMF, CZZ, ZSQ. Validation: YZW, YB. All authors read and approved the final manuscript.

Funding

This work was supported by Key R & D Program of Shanxi Province No. 2022SF-379 and Xi ‘an Jiaotong University Medical Engineering Interdisciplinary Program No. QYJC05.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of the First Affiliated Hospital of Xi’an Jiaotong University (No. XJTU1AF2022LSK-422), and all patients have signed informed consent.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.