Rehabilitation training robot using mirror therapy for the upper and lower limb after stroke: a prospective cohort study

Abstract

Background

This prospective cohort study was designed to investigate and compare the effectiveness of rehabilitation training robots versus conventional rehabilitation training on stroke survivors by monitoring alterations in brain network of stroke patients before and after robot intervention.

Methods

Between September 2020 and November 2021, stroke patients at four grade-A tertiary hospitals underwent limb rehabilitation training. Of the total of participants, 117 patients received conventional limb rehabilitation, 93 patients participated in upper-limb robot training, and 103 patients underwent lower-limb robot training. The measured outcomes included modified Barthel Index (MBI), Fugl-Meyer assessment subscale (FMA), and manual muscle testing (MMT). Functional magnetic resonance imaging (fMRI) was conducted on 30 patients to assess changes in the brain network. Data were mainly analyzed based on the Intention-to-Treat (ITT) principle.

Results

Post-interventional analysis utilizing linear mixed models in ITT analysis revealed that the robot training group had greater enhancements compared to the conventional limb rehabilitation training group. Notably, the shoulder flexor strength (P = 0.043) was significantly higher in the upper-limb group. On the other hand, hip flexor strength (P < 0.001), hip extensor strength (P < 0.001), knee extensor strength (P = 0.013), ankle dorsiflexion strength (P < 0.001) and ankle plantarflexor strength (P < 0.001) were significantly higher in the lower-limb group. In the upper-limb group, region-of-interest (ROI) -to-ROI analysis revealed enhanced functional connectivity between the left hemisphere’s motor control region and the auditory network. ROI-to-ROI analysis primarily showed enhanced interhemispheric functional connectivity in the lower-limb group, specifically between right the hemisphere’s motor control region (central opercular cortex) and left hemisphere’s primary motor area in the precentral gyrus.

Conclusions

According to our research findings, upper- and lower-limb rehabilitation robots demonstrated great potential in promoting motor function recovery in stroke patients. Robot-assisted training offers an alternative treatment method with comparable efficacy to traditional rehabilitation. Large-scale randomized controlled trials are needed to confirm these results.

Trial registration: The study was registered on the Chinese Clinical Trial Registry (ChiCTR1800019783).

Introduction

Stroke is the second most common cause of mortality worldwide. Recently, there has been a worrisome increase in the global incidence rate of stroke among young and middle-aged people (those under 55 years) [1]. Approximately 90% of stroke survivors have some kind of functional impairment, which imposes substantial social and economic burdens on the patients and their families [2]. Movement disorders are the most common challenge, characterized by the loss or limitation of muscle control, which often leads to limited functional capacity and reduced participation in daily activities [2,3,4,5].

Impairment of upper limb function, paralysis, poor spatiotemporal coordination and spasticity are highly prevalent among stroke survivors [6, 7] and can result in impaired reaching, grasping, and manipulation [8]. About one-third of ischemic stroke survivors experience reduced mobility and inability to walk, characterized by impaired gait and balance, which affects the patient’s daily self-care, quality of life, occupational and social integration, and increasing the risk of falls [9,10,11,12,13,14].

Mirror therapy (MT) is a type of rehabilitation therapy in which a mirror is placed between the subject’s arms or legs so that the patient can look at the mirror image of the healthy limb to create an illusion that the affected limb is also moving normally. This operation has the ability to stimulate different areas of the brain to produce movement, sensation and pain consciousness, resulting in promotion of motor function reconstruction and brain function remodeling of hemiplegic limbs [15, 16]. Rehabilitation training robots apply the same concept as mirror therapy.

Robotic mirror therapy (RMT) techniques are a new type of mirror therapy, that integrate robot technology [17]. RMT works by replicating the gait trajectory of the healthy side. It perceives the patient’s active motion intention in real time by using the angle, angular velocity, angular acceleration, and phase information of the sampling point, and adjusting the reference trajectory in real time according to the state of the affected side to drive its movement and achieve rehabilitation [18].

Furthermore, robot training may enhance brain function by altering cortical excitability, which supports neuroplasticity and creates an environment conducive to neuronal reorganization in response to treatment [19, 20]. The effectiveness of robotic neurorehabilitation is assessed using established, standardized clinical scales, including the Functional Ambulation Scale (FAC), the Fugl-Meyer Assessment (FMA), the 10 m and 6-min walking tests [21], etc. Clinical measures are less sensitive when assessing neurobiological effects of sophisticated forms of neurotherapy. The most comprehensive and objective method for evaluating the impact and effectiveness of robotic neurorehabilitation is to use functional magnetic resonance imaging (fMRI) [22]. Initially, fMRI was employed to evaluate the occurrence and progression of motor dysfunction [23]. The blood oxygenation level-dependent (BOLD) signal provides an indirect assessment of human brain activity with a very high spatial resolution, relying on variations in deoxyhemoglobin (deoxyHb) levels [24]. Lately, the emphasis has shifted to examining the properties of the functional organization of brain networks, offering a promising way to comprehend a wide-ranging alteration brought on by strokes [23, 25], particularly the time correlation measurement between various areas of the brain blood oxygen level-dependent (BOLD) signals during a resting state of functional connection [26, 27].

Following a stroke, robot-assisted training has demonstrated the potential for enhancing activities of daily living (ADL), including motor function and muscle strength [28,29,30,31,32,33,34,35]. Studies report variability in outcomes based on patient characteristics, device utilized, length and volume of training, control group, and the outcomes measured. However, it is uncertain whether robot-assisted limb training is more beneficial than traditional therapy for the same frequency and length of time [36,37,38,39].

Given the significance of limb function for everyday activities, a prospective cohort study was conducted on stroke patients who had undergone motor dysfunction. The study investigated the clinical effects of rehabilitation training robots and whether robot training was more effective than conventional rehabilitation training. Additionally, it also investigated the changes in brain networks of stroke patients before and after robot intervention.

Methods

Study design

This was a prospective cohort study designed to evaluate the therapeutic efficacy of conventional rehabilitation training versus robotic-assisted training in stroke patients. After the pretrial, the stroke patients were initially reluctant to enter the control group and preferred to try the robot rehabilitation training. The physicians recommended the patient enrollment based on their condition, which was in line with their clinical needs. Clinical registration was originally a randomized controlled study, changed to a cohort study. All patients included gave their informed consent. The study was registered on the Chinese Clinical Trial Registry with the unique identification number ChiCTR1800019783, and it was approved by the ethics council of Nanjing Medical University’s first affiliated hospital (No. 2019-SR-310).

Setting

According to their motor function and selection, patients were divided into three groups: the upper-limb robot group (Upper limb group), the lower-limb robot group (Lower limb group) and conventional training group (Control group). Enrolled participants were managed through an online system (http://47.102.217.116:8888/apoplexy/) with each center having its own account and password to log in. Patients in the upper-limb group received robot-assisted motor function training using a desktop-type upper limb rehabilitation robot developed by Southeast University (Fig. 1). Those in the lower limb robot group received robot-assisted motor function training using lower limb walking exoskeleton assistant training device Xwalk, (Fig. 2). Patients in the control group received conventional or traditional rehabilitation instruction from therapists. Both groups received two sessions every day, five days a week, for four consecutive weeks. Each session lasted for 20 min.

Fig. 1

Desktop type upper limb rehabilitation robot

Fig. 2

Lower limb walking exoskeleton assistant training device Xwalk

The desktop-type upper limb rehabilitation robot included a display screen, camera, control box, robot body, and a computer host (Fig. 1), and involved four training modes: passive training, mirror training, active training, and damping training. Each training mode was equipped with various rehabilitation games. The upper limb robot system is a mirror rehabilitation robotic arm system based on a desktop force feedback device. During the rehabilitation game training, the force feedback device collects the patient’s limb movement trajectory, force information, etc., which can flexibly adjust the rehabilitation training plans. The system has the characteristics of high positional accuracy, large feedback force, and large motion space. It guides rehabilitation patients to complete designated training tasks in stages according to their physical condition. By playing games, patients can improve their training enjoyment, attention, and initiative. Assist patients in promoting the formation of separation movements, stimulating residual muscle strength, enhancing muscle endurance, restoring joint coordination, and restoring joint flexibility. Specifically, when the patient is in a flaccid phase, the machine perceives the flaccid state of the affected limb and uses a passive or assisted mirror training mode where the healthy side drives the affected side. When the patient is in a spastic phase and the affected limb is in a spastic state, the machine perceives that the muscle tension on the affected side is significantly higher than normal. The machine will activate the spasticity protection mechanism, follow the limb movements, and no further training will be performed. Then, slowly stretch the spastic limb to reduce muscle tension, and when the muscle tension decreases, perform task-oriented active training on the affected limb. When the muscle tension of the affected limb decreases and enters the separation movement phase, the machine perceives that the muscle tension of the affected limb has decreased to a level close to normal, and performs task-oriented active resistance training on the affected limb. Highly repetitive and engaging upper limb exercises can enhance motor sensory input and stimulate the brain to generate motor plans, effectively improve the patient’s upper limb motor function and thus motivating the patient. The robot’s weight reduction system can be used to reduce the upper-limb physical and active muscle exercise demands to achieve optimal training [40].

The robot-assisted training used mirror therapy integrated training with the lower limb walking assistant training device Xwalk (Fig. 2) with rigid large and small leg bars, and sole and lumbar support to assist the patient in standing. The controller operated four DC servo motors at the hip and knee joints to assist patients with walking training. The device simulated a normal gait, providing symmetrical, high-intensity, repetitive rehabilitation to stimulate sensory input and promote the reconstruction of neural control mechanisms in patients with upper motor neuron injuries. During Xwalk training, the patients were instructed to make attempts to match the movement of the robot while receiving biofeedback on their hip and knee joint from a computer monitor in the back. The patients used the biofeedback to monitor their performance during robotic-assisted stepping, adjusting the timing and intensity of leg movements to reduce errors. The lower limb robot system can provide patients with auxiliary support and reduce the level of lower limb weight bearing. Real time motion data of both limbs of the patient can be provided, such as torque current values and real-time motion trajectories of each joint. Multiple parts can be adjusted, including waist width and leg length, to achieve precise adaptation. It has basic training modes, such as passive, assisted, and resistance. The passive training mode simulates natural gait and sets parameters such as walking cycle, step length, and step height to provide patients with correct sensory input. The assistance and resistance training modes can enhance patients’ active walking ability. Provide multiple training gaits, while real-time natural gait planning can be performed based on the patient’s height, weight, and training situation to correct abnormal gaits and help patients to regain walking function.

Upper and lower limb robot training focused on restoring motor function through precise, repetitive, task-specific exercises. It incorporated routine motor rehabilitation including balance and posture training, locomotion exercises, lower extremity function training, and ADL practice. The intervention was customized to each patient’s functional abilities. For patients with severe impairments, early sessions concentrated on static and dynamic postural tasks, trunk alignment, increasing lower extremity range of motion enhancement, and overground walking. As patients improved or began the program with higher function, they progressed to more advanced balance and gait exercises. Training data was recorded everyday on the Case Report Form.

Patients in the control group received conventional or traditional rehabilitation instruction from therapists. Each group received two sessions every day, five days a week, for four consecutive weeks. Each session lasted for 20 min. The specific content of upper limb training for patients includes passive movements, massage, and traction for the shoulder, elbow, and wrist joints of the upper limbs. Active movement of upper limb shoulder, elbow, and wrist joints is included to enhance arm strength. Task-oriented training, such as daily life simulations such as (e.g., lifting, grasping, and dressing), as well as functional activities for specific tasks (such as writing and using utensils). The patients’ lower limb training includes increasing lower limb muscle strength through resistance exercises, isokinetic exercises, and other methods. We will gradually enhance the strength of the core muscle group through movements such as turning over, moving the center of gravity forward, standing up, and maintaining sitting balance. Gradually improve balance ability through methods such as standing on one foot, standing on a balance board, and walking sideways. Correct walking posture and gait, while correcting walking posture and gait, and conducting gait training in daily life scenarios.

Participants

Four rehabilitation facilities recruited participants between September 2020 and November 2021, i.e., the First Affiliated Hospital of Nanjing Medical University, Guangzhou First People’s Hospital, the Affiliated Brain Hospital of Nanjing Medical University and BENQ Medical Center. In all, 330 stroke patients took part, including 120 undergoing traditional limb rehabilitation training, 100 receiving upper-limb robot training and 110 receiving lower-limb robot training.

The functional magnetic resonance imaging (fMRI) examination was conducted on 30 patients to assess brain network alterations. However, due to limited project funding, the high cost of fMRI tests, and the need for two scans per patient (before and after intervention), the budget could not cover all patients. To ensure the data were representative of the cohort, we randomly selected 10 patients from each group, minimizing bias in image data processing.

The participation criteria for the interventional study included: (1) Cerebral infarction or cerebral hemorrhage in accordance with the 2016 edition of the Chinese Guidelines for the Diagnosis and Treatment of Cerebrovascular Diseases and consensus; (2) stable vital signs; (3) less than 75 years old; (4) significantly limited range of motion in the joints; (5) heart and lung function that is capable of completing training; (6) good cognitive function, MMSE (>24), can understand and actively participate in training programs; (7) agree and sign this informed consent for clinical research. Exclusion criteria included: (1) patient with other neurological complications or musculoskeletal diseases (e.g. Parkinson’s disease, multiple sclerosis, severe joint contracture, osteoporosis, fracture, etc.); (2) cardiac and circulatory diseases such as heart failure, unstable angina pectoris, poorly controlled hypertension, etc.; (3) weight > 120 kg; (4) Previous psychiatric history, severe anxiety and depression, uncooperative or aggressive behavior; (5) patients were judged on whether they had poor compliance and could complete the study as required; (6) ongoing involvement in other clinical trials. Written informed permission was acquired from each participant.

The flow process and research design are displayed in Fig. 3.

Fig. 3

Patient enrollment. The flow chart illustrating the study design and patients screened and enrolled in the study

MRI acquisition

A 3.0 T Verio MRI scanner (Siemens, Erlangen, Germany) with an 8-channel parallel head coil was used to scan each patient, who had to remain awake and close their eyes while lying quietly. Both structural and functional images were acquired. Echo-planar imaging (EPI) was used to obtain resting-state functional images with the following settings: repetition time (TR) = 2000 ms; echo time (TE) = 21 ms; slice thickness/gap = 4 mm/0.6 mm; acquisition matrix = 64 × 64; flip angle = 78°; voxel size = 3.5 mm × 3.5 mm × 4.0 mm; and field of view (FOV) = 224 × 224mm². Sagittal T1-weighted images were obtained with the following parameters: TR/TE = 1900 ms/2.19 ms; acquisition matrix = 256 × 256; flip angle = 9°; voxel size = 1.0 mm × 1.0 mm × 1.0 mm; slice thickness/gap = 1 mm/0.5 mm.

Data measurement

The demographic and scale scores of three groups of patients were recorded for comparison. Assessments were performed before treatment, and after 4 weeks of treatment.

The primary outcome was ADL measured using the Modified Barthel Index (MBI) [41], a scale that evaluates 10 fundamental elements of everyday tasks linked to mobility and self-care. Each of the 10 elements has five ranks (1–5), with higher scores denoting greater independence. Complete independence is represented by a total score of 100, whereas total reliance is represented by a score of 0.

The secondary outcome was the Fugl-Meyer assessment (FMA) [42], including FMA-UE and FMA-LE [43], which evaluates the level of motor impairment in the paretic lower and upper extremities, with higher scores indicating better performance. The Action Research Arm Test (ARAT) was used to evaluate upper-limb function by examining grasp, grip, pinch, and gross-motion exercises [44]. The test uses various equipment, including various-sized wood blocks, a sharpening stone, a cricket ball, a glass and water jar, hollow tubes with varying heights and thicknesses, ball bearings, washers, and marbles of varying diameters. Each task receives a score ranging from 0 to 3, contributing to a total score of 57, which is divided among the many tasks using the various pieces of equipment mentioned above.

Manual muscle testing (MMT), a 6-point grading standard ranging from 0 to 5, was used to assess muscular strength, with higher levels indicating higher muscle strength and level 5 representing normal strength [45, 46]. The modified Ashworth scale, also a 6-point grading scale, was employed to evaluate muscle tone, ranging from 0, 1, 1+, 2, 3, and 4 levels, with higher levels indicating higher muscle tone, and level 0 representing normal muscle tone [47].

Functional magnetic resonance imaging (fMRI), which detects changes in cortical signals in patients, was employed to localize the cortical central functional areas and to provide indepth analysis of other brain functions. Functional connectivity (FC) study offered activation analysis and additional insights into the mechanism behind the brain network-level therapy effects of robot training [48].

The indicators measured in the upper-limb group included MBI, FMA-UE, ARAT, MMT and MAS, while the lower-limb group included MBI, FMA-LE, MMT and MAS. All the above indicators were measured in the control group. Additionally, before and after the intervention fMRI data were collected from 10 participants in each group.

Sample size

To detect a minimal clinical relevant difference (MCID) of 5.34-point in MBI [49] with an SD of 25 [50], a statistically significant signal with an F-test power of 80% and an alpha error of 5% required 72 patients per arm, per earlier results. Assuming a 15% dropout rate, the recruitment target was set at 83 participants per group.

Statistical analysis

Scale data

Baseline differences between the three groups were assessed using the analysis of variance (ANOVA) for continuous variables and the chi-square test for categorical data.

Data were mainly analyzed based on the Intention-to-Treat (ITT) principle, using chain equations for multiple imputations of missing data (20 groups). All patient baseline characteristics-center, age, gender, height, weight, stroke type, stroke side, stroke duration-were used as covariate variables for data imputation. The analysis adopted a linear mixed model, treating the results of each evaluation as dependent variables, e.g., the primary outcome MBI. Fixed effects included time, group, and the interaction between time and groups, individuals were included as random intercept terms in the model to address baseline imbalance, and baseline measured variables were controlled for treatment weighted inverse probability (IPTW) [51]. IPTW involves two main steps. First, the probability—or propensity—of being exposed, given an individual’s characteristics, is calculated. This is also called the propensity score. Second, weights for each individual are calculated as the inverse of the probability of receiving his/her actual exposure level. The application of these weights to the study population creates a pseudopopulation in which measured confounders are equally distributed across groups [52]. Each person is attributed to a propensity score, estimated using a logistic regression model, and assigned a conditional probability of receiving robot assisted therapy. The covariates include all baseline characteristics of the patient (center, age, gender, height, weight, stroke type, stroke side, stroke duration). Create a pseudo population by taking the reciprocal of the probability of each individual receiving actual treatment (i.e., 1/PS for individuals receiving robot assisted therapy and 1/(1-PS) for individuals in the control group). The stable weight calculated by multiplying the original weight by the unadjusted treatment probability is used to ensure accurate estimation of variance. This study reports the interaction term’s β value (with the 95% confidence interval) between group and time as the therapeutic effect. Maximum likelihood was used to fit all models, and all models were checked for normality and homogeneity of variance.

Sensitivity analyses were conducted using ITT analysis without data imputation and Per-protocol (PP) analysis, applying the same analysis method as the main analysis. All statistical analyses were conducted in R 4.2.3 (R Foundation for Statistical Computing, Vienna, Austria).

fMRI data

Preprocessing

Functional and anatomical data were preprocessed using CONN (RRID:SCR 009550, version 22.a) [53, 54] and SPM [55] (RRID:SCR 007037 version 12.7771) with a flexible preprocessing pipeline [56]. Anatomical data were normalized into standard MNI space, segmented into grey matter, white matter, CSF, and lesion tissue classes. The data was then resampled to 1 mm isotropic voxels using SPM unified segmentation and normalization algorithm [57, 58] with an alternative tissue probability map (TPM) extended to include custom lesion masks.

Denoising

Functional data were denoised using a standard denoising pipeline [56] that included: regression of possible confounding effects based on motion parameters and their first order derivatives (12 factors); white matter timeseries (5 CompCor noise components); and CSF timeseries (5 CompCor noise components) [59], outlier scans (below 54 factors) [60]. Bandpass frequency filtering of the BOLD timeseries [61] between 0.008 and 0.09 Hz.

First-level analysis

Global Correlation maps (GCOR), representing network centrality at each voxel, were computed as the average of all short- and long-range connections between a voxel and the rest of the brain [62]. Connections were calculated using bivariate correlation coefficients between BOLD timeseries from voxel pairs, based on singular value decomposition of z-score normalized BOLD signals (subject-level SVD) with 64 components for each subject and condition [50]. Seed-based connectivity maps (SBC) and ROI-to-ROI connectivity matrices (RRC) were derived using 964 ROIs to characterize functional connectivity patterns.

Group-level analyses used a General Linear Model (GLM [53]). The differences in GCOR between follow-up and baseline measurements across groups (Upper limb or Lower limb) were tested to assess brain connectivity changes over time, controlling for age, gender, duration, and motor scores (ARAT or FMA-LE). Moreover, correlations between changes in brain connectivity and motor function improvements before and after the intervention were analyzed.

Changes within the somatomotor network of the Yeo 17 Networks (Schaefer 400 parcels) [63] before and after the intervention were analyzed using a ROI-to-ROI approach. The analysis of Functional Network Connectivity (FNC) utilized parametric multivariate statistics for cluster-level inferences [64]. The primary cluster-forming threshold was set at P < 0.05, with cluster-level correction for multiple comparisons using the False Discovery Rate (p-FDR). An uncorrected threshold of P < 0.05 was applied for individual connections, ensuring robust detection of significant connectivity changes while balancing the risk of Type I errors.

Results

Baseline characteristics

Between September 2020 and November 2021, 344 stroke survivors were consecutively recruited, among whom 313 were assigned to three groups, with 93 cases in the upper-limb robot group, 103 cases in the lower-limb robot group and 117 cases in the control group. A total of 313 (94.84%) received a 4-week assessment. The overall patient enrollment process is presented in Fig. 3. The baseline demographics (gender, age, height, weight) and stroke characteristics (type of stroke, or hemiplegic side) were comparable among the three groups (P > 0.05) (Table 1). After adjusted by the IPTW, the imbalance among groups has been addressed, and the standard mean differences (SMD) are almost all <0.1 (Table 1).

Table 1 baseline characteristics of the enrolled subjects

Comparisons between upper limb group and control group

There was no significant difference in MBI between the control group and upper-limb group. Notably, the MMT results, revealed that shoulder flexor muscle strength (P = 0.043) was significantly higher in upper-limb group compared with the control group (Table 2).

Table 2 MBI, FMA, ARAT, MMT and MAS score at 4 weeks rehabilitation between upper limb group and control group (Mixed effects model—ITT with MI)

In the upper limb group, changes in GCOR maps between pre- and post-intervention were observed at MNI coordinates (14, −72, 46), with a cluster size of 41 and a corresponding cluster-level qFDR-corr value of 0.007. Seed-to-voxel analysis identified two clusters, with peak MNI coordinates at (58, −22, −2) and (64, −16, −24), and corresponding cluster sizes of 47 (qFDR-corr = 0.006) and 32 (qFDR-corr = 0.029), respectively. ROI-to-ROI analysis identified enhanced FC between the left hemisphere’s motor control region and the auditory network (Fig. 4.).

Fig. 4

Functional and anatomical data (upper limb). A The difference in Global Correlation maps (GCOR) between pre- and post-intervention. B Seed-to-voxel analysis in the upper limb group. C ROI-to-ROI analysis in the upper limb group

Comparisons between lower limb group and control group

No significant difference in MBI was found between the control and lower-limb groups. However, the lower-limb group showed significantly higher strength in hip flexors (P < 0.001), hip extensors (P < 0.001), knee extensors (P = 0.013), ankle dorsiflexors (P < 0.001), and ankle plantarflexors (P < 0.001) (Table 3).

Table 3 MBI, FMA, MMT and MAS score at 4 weeks rehabilitation between lower limb group and control group (mixed effects model—ITT with MI)

In the lower limb group, no significant differences were found in GCOR between pre- and post-intervention. However, changes in GCOR were strongly associated with changes in the FMA-LE motor score symptom score, with a peak activation at MNI coordinates (2, −50, 34) and a cluster size of 244 (qFDR-corr < 0.001). Seed-to-voxel analysis identified two significantly activated clusters, with peak MNI coordinates at (12, −40, −2) and (54, −52, 14), and corresponding cluster sizes of 44 (qFDR-corr < 0.001) and 41 (qFDR-corr < 0.001), respectively. ROI-to-ROI analysis confirmed enhanced interhemispheric FC, specifically between the right hemisphere’s motor control region (central opercular cortex) and the left hemisphere’s primary motor area in the precentral gyrus (Fig. 5.).

Fig. 5

Functional and anatomical data (lower limb). A The difference in Global Correlation maps (GCOR) between pre- and post-intervention. B Seed-to-voxel analysis in the lower limb group. C ROI-to-ROI analysis in the lower limb group

Sensitivity analyses

The results of the sensitivity analyses are shown in Tables 4 and 5. Treatment effects remained consistent between ITT analysis with raw data and the PP analysis. Additionally, the ITT analysis with MI demonstrated robustness.

Table 4 MBI, FMA, ARAT, MMT and MAS score at 4 weeks rehabilitation between upper limb group and control group (sensitivity analysis)

Table 5 MBI, FMA, MMT and MAS score at 4 weeks rehabilitation between lower limb group and control group (sensitivity analysis)

Discussion

Upper-limb

This study was conducted during the 2019 coronavirus disease (COVID-19) pandemic. It describes the clinical outcomes of robot-assisted training on stroke patients, including activities of daily living, motor functions, muscle strength, functional and anatomical data.

The MCID of FMA-UE was found to be 4.25–7.25 [65]. Before and after intervention, the mean difference for the control group was 6.02, and the mean difference for the upper limb robot group was 6.87, exceeding the MCID in the literature, indicating that both conventional treatment and upper limb robot treatment have clinical significance for the rehabilitation of stroke patients. But there was no statistical difference between the two groups, and the improvement value of the upper limb group was slightly higher than that of the control group, indicating that within the 4-week intervention, the two groups were equivalent, and the upper limb robot group was slightly better than the control group.

The MCID of the ARAT scale was 6.0 [66]. Before and after intervention, the mean difference between the control group was 3.32, and the mean difference between the upper limb robot group was 3.27, both of which did not reach the MCID, indicating that both conventional treatment and upper limb robot treatment did not produce clinical significance in fine motor function of the upper limb within 4 weeks. Because the functions of the upper limbs, especially the hands, are extremely delicate and diverse, including grasping, pinching, and other movements, which require high coordination and precise control, and require larger neural control areas in the brain. A 4-week intervention is not sufficient to bring about changes in fine motor skills in the upper limbs.

The findings showed that shoulder flexor muscle strength was significantly higher in the upper-limb group. This outcome is consistent with an earlier study on robot-assisted arm training that employed a device with an end effector for 40–105 min daily over four weeks in addition to traditional therapy [66,67,68,69,70].

Muscle weakness, an obvious symptom after stroke [71], is one of the main factors that slow down the recovery to normal physical fitness in patients [72]. Paralysis on the side of the body on the opposite the brain lesion is the most notable consequence [73], although a correlation between insufficient strength on the same side and walking speed has also been observed [74]. Paresis, defined as a change in the ability to generate normal levels of muscle strength [75], can lead to abnormal posture and stretching reflexes, as well as loss of autonomous movement [76, 77]. Restoring hand function is challenging due to its complex, precise movements. Moreover, hand function is primarily controlled by the anterior central gyrus, a larger area of the left cerebral cortex plays a key role in right-handed individuals. Stroke-induced upper limb motor dysfunction mechanisms is attributed to pathologically reduced cortical excitability and impaired limb innervation. Repetitive motor training of the affected limb stimulates the corresponding cortical representative area of the cortex to be enlarged and the efficiency of nerve signalling to be increased. Robotic training provides a specific context for motor relearning, emphasing on visual and auditory feedback to correct abnormal movement patterns [78, 79]. Highly repetitive and engaging upper limb exercises can enhance motor sensory input and stimulate the brain to generate motor plans, effectively improve the patient’s upper limb motor function and thus motivating the patient. The rehabilitation robot’s exoskeleton device and support system promote autonomous patient movement by adjusting active or passive modes. The robot’s weight reduction system can be used to reduce the upper-limb physical and active muscle exercise demands to achieve optimal training [40].

In the upper-limb group, ROI-to-ROI analysis revealed enhanced functional connectivity between the left hemisphere’s motor control region and the auditory network. Studies have confirmed that there are unidirectional or bidirectional direct nerve fiber projections between the auditory and somatosensory cortices, as well as between the visual and auditory cortices [78]. The interaction between the motion and auditory systems takes the form of feedforward and feedback relationships. These interactions may be related to the “auditory action” system, similar to the mirror neuron system. The auditory system primarily affects motor output in a predictive manner. For instance, musicians receive input and feedback from multisensory information by repeatedly practicing hand and upper limb movements alongside auditory and visual information, which not only activates the parietal lobe but also enhances the functional connection between the auditory and motor cortices [79]. Studies analyzing the changes in FC of the brain network in upper limb amputees (ULAs) report that both the left precentral gyrus in the auditory network and the left precuneus gyrus in the dorsal attention network showed a decrease in intra network FC [40]. These examples demonstrate that when upper limb motor function is enhanced, the connection between the motor cortex and auditory network is strengthened. Conversely, when motor function is weakened, the connection between the two functions is weakened.

Upper limb functional indicators showing significant differences in shoulder flexor strength after four weeks of rehabilitation, robotic training appears to accelerate the process of upper limb rehabilitation. This and other data trend are an indication of better improvement compared with the control group.

Lower-limb

Gait training is not only about restoring the patient’s ability to walk, but also about the patient’s ability to walk and participate in the community [80]. Routine rehabilitation of stroke patients with gait abnormalities is carried out in steps. Gait training can only begin when there is adequate supportive strength in the trunk and limbs [81]. Previous studies show that chronic stroke survivors (>6 months after stroke) have decreased muscle mass and strength in both hemiplegic and non-hemiplegic limbs [82, 83].

Robot-assisted gait training has demonstrated equivalent or even better training effects and safety standards compared to conventional rehabilitation training. It has several notable advantages, including precise and controllable training patterns, repeatability, low energy consumption, timely feedback, and objectivity [84,85,86,87].

Conventional rehabilitation therapy for post-stroke motor function recovery often requires multiple therapists to manually guide the training. The limitations of conventional rehabilitation training method are becoming more apparent due to the rising incidence of stroke and the shortage of therapists. Efficiency in personnel utilization remains low, with significant variability in the skills and experience among therapist partly impacting treatment outcomes. Furthermore, the traditional rehabilitation training method fail to accurately record the training parameters such as joint movement speed, displacement and torque, but they cannot be accurately controlled. As a result, individual patient training sessions are therefore restricted. With the increased demand in personnel in the field of rehabilitation medicine, rehabilitation robots offer a capability of assisting or replacing therapists by streamlining traditional rehabilitation training methods. The combination of robotics and medical technology has helped patients rebuild the central nervous system, restoring their activities of daily living, global motor function, limbs and trunk strength, and gait improvement with enhanced safety and efficacy compared with conventional therapy [86,87,88,89,90].

The MCID of FMA-LE was found to be 6.0 [91]. Before and after intervention, the mean difference for the control group was 4.65 and for the lower limb robot group was 3.96, both of which did not reach MCID. This indicates that both conventional treatment and lower limb robot treatment have clinical significance for lower limb motor function within 4 weeks. The results showed that hip flexor muscle strength, hip extensor muscle strength, knee extensor strength, ankle dorsiflexion strength and ankle plantarflexor strength were significantly higher in lower-limb group compared with conventional therapy. These results were consistent with previous research [92, 93]. The effectiveness of the robot may have been driven by the following mechanisms: the robot’s bodyweight support enhanced the stroke patients’ walking stability and training effectiveness, provided continuous, consistent, repetitive, and intense training, which increased its efficacy, and ameliorated the cardiopulmonary function and blood circulation within the lower limbs. Moreover, the mirror rehabilitation robot allowed monitoring of the patient’s recovery status in real time and adjustment of the training mode. During the soft paralysis stage, the robot detected the affected limb’s condition and activated the passive or assisted mirror training mode, enabling the healthy side to guide the movement of the affected side. In the case of muscle spasm, the robot sensed the increased muscle tension in the affected side, thereby opening the spasm protection mechanism. The robot followed the limb activities without the need for prior training and reduced the muscle tension by slowly stretching the spasm limb. If the muscle tension decreased, the task-oriented active training of the affected limb was performed. As the muscle tension of the affected limb decreased and entered the separation movement phase, the robot detected the near-normalization of muscle tension on the affected side and initiated task-oriented active resistance rehabilitation training for the affected limb.

In the lower limb group, ROI-to-ROI analysis identified enhanced interhemispheric functional connectivity, specifically between the right hemisphere’s motor control region (central opercular cortex) and the left hemisphere’s primary motor area in the precentral gyrus. The stroke was accompanied by extensive changes in the structure and functional connections between hemispheres [94, 95]. These changes may be associated with neurological deficits and the dynamics of functional recovery after stroke. Several studies on stroke patients measured the interhemispheric FC, which is located within the sensorimotor network. Similarly, this study found that almost all patients experienced a decrease in interhemispheric FC intensity in the early stages after stroke [96,97,98,99]. Vahdat et al. [100] and Guerra Carrillo et al. [101] demonstrated that interhemispheric FC was enhanced after motor and perceptual training. Fan et al. [102] reported that robot-assisted bilateral arm movement therapy enhanced the interhemispheric FC and improved the motor function. Another study exploring the brain-computer interface revealed that the subjects achieved significant functional recovery after a 5-week brain computer interface intervention treatment. Results of the EEG analysis indicated that the functional connectivity between the damaged hemisphere motor areas of the subjects was enhanced before and after treatment [103]. A previous study investigated how the resting-state functional connectivity (rsFC) and motor outcomes in stroke recovery were affected by the combination of EEG-based BCI intervention and functional electrical stimulation (FES). Following BCI intervention, the motor network’s interhemispheric and network rsFC significantly increased at the group level [104]. Our imaging results are consistent with previous studies, indicating that the enhanced functional connectivity between hemispheres suggests the role of this lower limb robot in promoting neurological recovery in stroke patients.

The MCID of the MBI scale was 5.34 [49]. From the statistical results, the mean changes of the three groups before and after the intervention were greater than the MCID, indicating significant clinical improvement in the treatment of stroke patients. However, there was no statistical difference between the three groups before and after the intervention, and the lower limb robot group was only numerically superior to the control group, indicating that within the 4-week intervention, robot intervention and the conventional rehabilitation intervention were equivalent, but lower limb robot intervention showed certain advantages. In our study, MBI scores did not differ between the control and robot groups, likely due to several factors. While robotic training enhances muscle strength, through repetitive, precise limb movements, without a focus on functional tasks or ADL, individuals may not experience improvements in their ability to perform these tasks. After a stroke, the brain can self-reorganize, and with the formation of new neural pathways, muscle strength is initially improved [105]. However, this does not always translate into an improvement in functional abilities in daily activities. Psychological factors, such as motivation and self-efficacy, can affect the performance of daily life activities [106]. Even with increased strength, a person may lack the confidence to perform certain tasks. After a stroke, some patients may experience severe functional impairment and may not be able to recover to normal in the short term. In addition to being unable to take care of themselves in daily life, their work, economy, and other aspects will also be greatly affected, and even their status in the family and society will change. Secondly, patients often feel fear and worry about the disease and prognosis, which can lead to psychological disorders, in turn affecting the recovery of motor function. Stroke behavioral deficits refer to the behavioral changes that occur in individuals after a stroke. These changes may include sensory, motor, cognitive, and emotional disorders. Including difficulties in movement, coordination, and balance. as well as difficulty with memory, attention, and problem-solving [107]. Emotional changes such as anxiety, depression, and irritability are common. Utilizing these robots can improve various behavioral and psychological outcomes. Patients may experience reduced anxiety and depression levels, increased self-efficacy, and improved mood due to positive experiences during therapy.

Compared with traditional training, robot intervention has the following advantages. Mirror rehabilitation robots can promote motor recovery in patients with nerve injuries through precise repetitive training, thereby enhancing brain plasticity. The interactive nature of robotic systems may increase patient motivation and compliance during rehabilitation exercises [108]. These robots can be programmed to tailor exercises to individual patient needs, potentially leading to more effective rehabilitation outcomes. Robots can track progress and performance metrics, allowing therapists to adjust treatment plans based on real-time data [109].

However, there are also limitations, the high cost of mirror rehabilitation robots may limit accessibility for some patients and healthcare facilities [110]. Both patients and therapists may require training to effectively use these systems, which could be a barrier in some settings. The effectiveness of the intervention may vary among individuals due to differences in motivation, severity of disability, and other factors. Overreliance on robotic rehabilitation systems may weaken the effectiveness of traditional rehabilitation techniques, which are beneficial for stroke patients [111].

In terms of cost-effectiveness, although the initial cost of mirror rehabilitation robots may be high, including the purchase and maintenance costs of the robots, personnel training costs, and operating costs. However, in the long run, the use of rehabilitation robots can accelerate the readaptation process of stroke patients, reduce hospitalization time or rehabilitation courses, thereby offsetting initial costs [112]. Overall, although the initial cost may be high, the long-term benefits and savings often make robot intervention a cost-effective solution.

The results of this study must be interpreted in the context of certain limitations. A more trobust experimental design would include artificial mirror therapy as a control group, enabling direct comparison of changes in range of motion and strength between this group and the mirror rehabilitation robot group. Moreover, considering individual preferences, rehabilitation psychology, and economic factors would provide a more comprehensive understanding of the strengths and weaknesses of both mirror robot rehabilitation and traditional mirror therapy. Considering the evaluation of the relative benefits, limitations, cost-effectiveness, scalability, and accessibility of robot interventions make research design more comprehensive. This approach would be instrumental in tailoring personalized rehabilitation plans.

Conclusion

This study demonstrates that upper and lower limb rehabilitation robots can accelerate the recovery of motor function in stroke patients. Therefore, robot-assisted training may be an alternative treatment that provides persistent efficacy. In the future, large-scale randomized controlled studies are needed to validate our results.