Around 50% of patients with stroke suffer from limited dexterity in the affected upper limb (UL), ranging from slightly reduced dexterity to paralysis [1, 2]. Almost half of the patients with initial severe paresis, without a functional hand, and paralysis do not reach a satisfactory level of UL motor function, leaving many with an unusable UL and dependent on help [3]. Since most daily life activities are bimanual, impaired UL function compromises daily activities, quality of life, and vocational reintegration. Intensive training during the first weeks to months is crucial for regaining motor function [4]. During this period, the brain is most susceptible to reorganization [5]. However, it can be challenging to provide meaningful training for patients who can hardly move their UL voluntarily. Contrary to patients with some UL function who can participate in interventions such as constraint-induced movement therapy or task-specific training, there is a paucity of evidence-based and effective interventions for patients with severe paresis. Treatment options with some benefit include mirror therapy, mental practice, functional electrical stimulation, and advanced technologies, such as EMG biofeedback, robotic assistance, and exoskeletons, usually only available in specialized rehabilitation centers [6–8]. A relatively recent treatment alternative are restorative brain–computer interface (BCI) systems based on new treatment strategies that pair brain activity with contingent sensory feedback. Feedback modalities may include visual, motor and somatosensory stimulation. Visual feedback can involve 3D computer graphics (e.g., virtual arm model movement on the screen) or real object manipulations, while motor and somatosensory stimulation can be delivered by functional electrical stimulation (FES) or robotic (e.g. exoskeleton) devices executing/assisting the desired movements [9]. FES stimulates paretic sensorimotor nerves and induces artificially generated movements and afferent sensory feedback. Thereby, such restorative BCIs create a real-time feedback loop that can facilitate plasticity by strengthening neural connections [10–12]. Moreover, principles of experience-dependent neural plasticity are applied since patients are cognitively engaged in the training comprising many challenging repetitions, which is particularly relevant for patients with severe paresis who are unable to participate in more active interventions [13].

Several reviews conclude that the evidence for BCI training is still scarce, though promising, both as a restorative and an assistive approach [14–17]. Different systems and heterogeneous patient groups, mostly in the chronic phase after stroke, make comparisons challenging. Still, improvements in UL motor function in patients in the chronic phase were significant and seemed to last after completing BCI-based therapy [18–20], though it can be debated if these improvements were clinically meaningful. In a recent study, BCI training combined with FES was found to increase resting state connectivity between the hemispheres and within the motor network, which correlated with improvements in UL motor function in chronic patients [21]. Other studies have found a stronger desynchronization of alpha and beta bands in the ipsilesional hemisphere after BCI training [22, 23], and emphasized the importance of using BCI methods to improve the sensorimotor rhythms associated with movement preparation, as an alternative to modifying the brain activity during movement [24].

However, the evidence of therapeutic effects from randomized controlled clinical trials including BCI technology for different patient populations is still insufficient [16]. Few small-scale studies targeted patients with severe paresis in the first weeks after stroke. Pichiorri et al. [22] included 28 patients and found larger improvements for the BCI group in combination with visual feedback than for the control group. Morrone et al. [25] included eight severely impaired individuals in the subacute stage but focused on usability only.

Demonstrating possible treatment effects in the subacute phase can be challenging due to spontaneous biological recovery [26]. Furthermore, patients in the subacute phase frequently suffer from multiple impairments and psychological distress, making it more challenging to achieve levels of training intensity even remotely close to those suggested by pre-clinical research [27]. Most importantly, not all patients with severe UL paresis may have the underlying biological capacity in terms of cortico-spinal tract integrity to respond to the BCI intervention [28]. In this context, BCI training seems promising, but the evidence is insufficient and challenging to generate.

In this study, we examined the effectiveness of BCI training for patients with severe upper limb paresis in the subacute phase after stroke. Moreover, we investigated potential changes in cortical activation in the BCI group. We hypothesized that BCI training combined with visual feedback and FES would result in better UL recovery in the BCI group. We also expected changes in motor imagery-related electroencephalography (EEG) patterns towards enhanced laterality of the event-related sensorimotor oscillations when approaching the end of therapy in the BCI group, coinciding with the UL recovery process. Furthermore, we hypothesized better UL recovery for patients with preserved cortico-spinal tract (CST) integrity, regardless of the type of intervention.

Forty patients were included from August 2019 to February 2022. Due to the Covid-19 pandemic, the inclusion had to be interrupted for approximately 6 months in total. Follow-up assessments could be obtained for 35 patients, 5 patients had discontinued their participation, 4 in the BCI group, and 1 in the control group, Fig. 2. Reasons for discontinuing were mainly unrelated to the study, such as worsening medical instability (n = 1), and early discharge (n = 2). One participant in the BCI group experienced muscle soreness due to the FES, which could not be resolved by decreasing the stimulation intensity. In the control group, one patient retracted their consent.

Patients in the two groups did not differ with regard to age, days post-stroke, stroke severity, and other characteristics at baseline. The ability to move the hand and to walk at stroke onset was based on a score of 0 for the hand item, and a score of 0 or 3 for the gait item on the Scandinavian Stroke Scale conducted at the acute hospital [36].

TMS examinations could be performed with 32 patients, seven had contraindications, and one declined. Demographic and medical information is presented in Table 1.

Patients in the BCI group received a mean of 10.6 (± 2.5) treatment sessions with the BCI system. As the BCI training was not an add-on but a part of UL training, the amount of training received was balanced. Patients received on average 11.15 h of physiotherapy and occupational therapy per week in both groups. The therapy did also address other impairments. The 3–4 h BCI training per week was part of the regular physiotherapy and occupational therapy sessions. A BCI session, including donning on doffing, lasted for 60 min and most patients did 160 repetitions during a session.

There was a statistically significant improvement within the groups from baseline to 3-month assessment, p = 0.012 for the control group, p = 0.007 for the BCI group for the main outcome measure ARAT. The median score for the BCI group changed from a baseline median (IQR) of 0 (0) for ARAT and 4 (3) for FMA to a 3-month follow-up score of 3 (11) for ARAT and 6.5 (21) for FMA. The control group had a baseline median (IQR) score of 0 (0) on ARAT and 4 (2) on FMA, and a 3-month follow-up score 0 (9) on ARAT and 4 (21) on FMA. The between-group differences were not statistically significant, neither for ARAT (p = 0.328) nor for FMA (p = 0.406).

In general, only a few patients (10/35) in our sample improved and reached the MCID of 6 points change for ARAT at three months post-stroke. Of those, more patients were in the BCI group (5/15 = 33%, CI95% 12–62%) than in the control group (5/20 = 25%, CI95%: 9–49%). Figure 3 displays the numbers for the BCI and the control group.

The logistic regression with “allocation” (BCI/control) and “MEP status” (MEP+/MEP−) as independent variables, and a MCID of ≥ 6 on ARAT as dependent variable, was performed on 29 patients with complete data for both ARAT at three months and MEP status. An interaction term of allocation and MEP could not be estimated due to the limited sample size. The model explained 46.2% (Nagelkerke R²) of the variance in achieving an MCID on ARAT. Of the two independent variables, only MEP status was statistically significant. Patients with MEP+ had 29.4 higher odds to achieve some improvement in UL function, see Table 2.

In this study, we compared BCI training to standard training in patients with severe UL paresis after stroke. Slightly more patients in the BCI group (5/15) compared to the standard training group (5/20) achieved a clinically meaningful improvement of at least 6 points on ARAT. However, the difference was not statistically significant. MEP status was much more vital for future UL function than the type of treatment patients received.

This is the first study examining BCI training in the subacute phase where mainly patients with UL paralysis participated. This is a group for whom we lack effective treatments and where in general many will suffer from persistent impairments [37, 38]. Though the patients were relatively homogenous with regard to their severe UL impairment, we did not exclude patients with MEP− or unknown MEP status. The lack of UL recovery potential in a substantial part of our patients may have distorted the results. The use of biomarkers, especially TMS, to identify and stratify patients, is endorsed by several studies [29, 39] and recommended by the Stroke Recovery and Rehabilitation Roundtable [40]. Yet its use is not unequivocal since some patients with MEP− still improve their UL function. Hoonhorst et al. found no added value of TMS within the first 48 h but at 11 days after stroke [41]. In the study by Lundquist et al. 2 of 12 patients with MEP− achieved some dexterity [42]. Powell et al. applied neuromodulation in both patients with MEP+ and MEP− [43]. They found an improvement of 4.2 points on FM and 1.8 points on ARAT in MEP− patients. While these improvements are statistically significant, they are below the MCID for the assessment tools and do not indicate distal dexterity. An exception is the analysis by Senesh and Reinkensmeyer where some patients presumably without CST integrity responded to intensive training in the chronic phase [44]. Still, they did not reach advanced levels of fine motor skills. Classic experiments with non-human primates with lesions of the hand representation in the primary motor area showed that the animals regained control over arm movements, though slower, but not over selective finger movements [45]. Though alternative pathways, such as the rubrospinal tract, reticulospinal, and ipsilesional connections have been regarded as a potential means of compensating for CST damage [46, 47], CST integrity still seems to be the most decisive factor in regaining fine motor skills [40, 48]. Nevertheless, we cannot rule out that some people may be better capable of compensating for lost CST integrity, at least to some degree [44].

In our study, most patients (15/16) with MEP− did not reach any dexterity in the affected UL. This seems to be in contrast to the results by Pichiorri et al. [22], the only other study with subacute patients. In this study, 14 patients without detectable MEPs participated and many of those improved. However, it has to be taken into account that this improvement was likely limited to proximal, i.e., shoulder and upper arm function, and did not include finger movements [49]. In general, their sample included patients with better UL motor function, with a mean FMA score of 23.4 ± 17.3 in the BCI group and 24.2 ± 18.2 in control, and less overall stroke severity, a mean NIHSS score of 9 ± 2.6 in the BCI group and 8 ± 2.3 in control, than our sample with a mean FMA score of 3.4 ± 2.0 in the BCI group and 4.7 ± 5.7 in control, and a mean NIHSS score of 15.1 ± 2.7 in the BCI group and 15.9 ± 4.9 in control. Consequently, our findings corroborate the need for the stratification of patients in research and clinical applications based on the recovery potential derived from biomarkers and clinical assessments [50].

In the present study, patients received a mean of 10.6 BCI sessions. We cannot tell if more BCI sessions would have led to better results. When applied in the chronic phase, frequently a much longer treatment phase of 20–25 training sessions [20, 51, 52], or 18–30 h was offered [23]. This was not achievable in our subacute setting with a limited length of stay and patients suffering from multiple impairments which also needed to be addressed. In many cases, it was quite challenging to conduct 3–4 sessions a week. Yet, some other studies also applied a similar number of sessions. Biasucci et al. [19] conducted 10 sessions of BCI-actuated FES compared to sham FES over 5 weeks in 27 patients in the chronic phase and found significant improvement in the BCI-FES group that was retained 6–12 months after the intervention. In the above-mentioned study by Pichiorri et al. [22], 3 weekly sessions for 4 weeks were provided. Both groups improved, but there was a significantly larger improvement in the BCI group. Not only the number of BCI sessions but also the number of trials within a session could be relevant. Compared to a study by Sebastian-Ramogosa et al. [20] applying the same BCI system (RecoveriX, g.tec) in patients in the chronic phase after stroke, our patients received not only fewer sessions but also fewer runs with fewer trials within the session, only 160 as compared to 240. This was mainly caused by the lack of endurance in our subacute sample where patients suffered multiple motor and cognitive impairments, e.g., most were wheelchair users at the time of treatment. BCI training was well tolerated by these severely impaired patients when confined to 160 trials, which was still more or equal to the number of trials targeted in other studies [19, 22].

EEG analysis revealed a significant increase of LI for less affected hand movement imagery indicating an increase in contralesional activation towards the end of therapy in the broad range of movement-related frequencies. Unilateral movement imagery induces predominantly contralateral brain activation, reflected in contralateral ERD [53]. Since our results indicate a weakening of ipsilateral cortical activation for less affected hand imagery this may reflect a more natural activation pattern for less affected hand imagery towards the end of therapy, in line with both ERD and LI results. This normalization was expressed in an engagement of the lesioned hemisphere, as well as the laterality of the event-related sensorimotor oscillations during the less affected UL motor imagery over the course of BCI training. The observed changes in EEG pattern over time correspond to the statistically significant change in clinical scores on ARAT and FMA within the BCI group, implying that improvement in UL function might occur due to plastic changes within the brain. We emphasize that all patients successfully controlled the BCI system using sensorimotor alpha and beta rhythms during therapy. ERD/ERS of the sensorimotor rhythms was present in all analyzed data, confirming that the severity of impairment did not affect the presence of cortical activation in both hemispheres during UL motor imagery.

The lack of changes in the affected hemisphere for the affected hand motor imagery in our study is in contrast to findings by Pichiorri et al. who found significantly stronger contralateral ERD of the alpha and beta bands for both affected and unaffected UL imagery. However, their BCI patients had substantially higher UL function at baseline (mean FMA 23.4 ± 17.3) and experienced some recovery. We can only assume that the absence of any significant difference in event-related desynchronization (ERD) during affected hand motor imagery throughout the course of therapy may be attributed to the severe impairment and very limited recovery in our patients.

We did not find a difference in UL function improvement between patients receiving BCI training and those receiving standard UL rehabilitation only. However, a substantial part of the patients included suffered from a loss of CST-integrity, which likely limited their UL recovery potential independent of the type of training. Consequently, we can neither refute nor confirm that BCI training could be effective for patients with preserved CST-integrity. Larger studies recruiting only patients with some preserved CST-integrity should be attempted. Moreover, the intensity of the BCI training provided in the present study was low. Ideally, BCI should be continued ambulatory or at home after discharge to achieve sufficient intensity. Evidence from this study may be important for further determination of the type, onset, intensity, and dose of BCI therapy in severely impaired stroke patients in the subacute phase.

The authors would like to thank the participating patients and project therapists.

IB and AS conceived and designed the study. ARP guided the analysis of the clinical data. CBL, SD, and ES provided advice for the study and revised the manuscript. AS conducted all analyses on EEG data. All authors read and approved the final manuscript.

This study was funded by Helsefonden, Denmark. Grant number 19-B-0147.

Data are available from the corresponding author upon request.