Neuromodulatory effects of voluntary-motor-effort-driven wrist-hand robot in sensorimotor dynamics after stroke Authors: Yunxia Huo, Weidong Li, Fanjie Ding, Di Ao, Huijing Hu, Xiaoling Hu, Le LI Published: 2026 ABSTRACT BackgroundVoluntary motor effort (VME)-driven neuromuscular electrical stimulation (NMES)-robots have demonstrated superior rehabilitative efficacy in post-stroke wrist–hand motor restoration compared with purely mechanical systems. However, their neuromodulatory mechanisms remain unclear. This study aimed to investigate cortical responses elicited by an electromyography (EMG)-driven NMES-robot assisting wrist–hand movements in individuals with chronic stroke, and to evaluate robot-nervous system interactions, using EMG to capture VME.MethodsCortical responses during wrist–hand extension were recorded in participants with chronic stroke (n = 18) and unimpaired controls (n = 20) using functional near-infrared spectroscopy (fNIRS). The EMG-driven NMES-robot delivered somatosensory stimulation with different combinations of mechanical stimulation (MS) to the targeted joints and NMES to the forearm muscles, with or without VME.ResultsCombined MS and NMES co-stimulation activated wider cortical areas than MS or NMES alone in both groups. Compared with the unimpaired group, the stroke group exhibited greater ipsilesional activations in response to NMES (p < 0.05), while MS induced comparable activations in the contralateral hemisphere to the stimulated limb in both groups (p < 0.05). When VME was required, the stroke group exhibited wider contralesional involvement with MS only (p < 0.05) but reduced contralesional involvement during sole NMES (p < 0.05), compared to those in the co-stimulation. Additionally, the MS and NMES co-stimulation resulted in similar functional connectivity patterns between the stroke and unimpaired groups.ConclusionCombined mechano-electrical somatosensory co-stimulation with MS and NMES to the affected wrist-hand, when voluntary motor effort is generated from the residual muscles, could elicit extensive cortical activation approaching normal patterns.Trial registrationChinese Clinical Trial Registry with the identifier ChiCTR2200057839.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12984-026-02062-y. Background Motor restoration of distal joints in the upper limbs, (e.g., wrists and hands), is challenging after a stroke and is associated with considerable impairment in extensor function but relatively preserved voluntary flexion [1]. During rehabilitation, stroke survivors commonly achieve higher motor gains in proximal joints than in distal joint, primarily because of compensatory muscle activity from the proximal joints [2]. This compensatory activity leads to learned disuse of distal muscles in the long term, which is a key factor limiting further recovery of distal motor function [3]. Rehabilitation robots are commonly used in long-term motor function rehabilitation after a stroke. These devices mechanically assist human therapists in completing labor-intensive physical interventions with repetitive physical exercises. By supporting impaired joints, compensating for gravitational forces, and guiding functional movements, rehabilitation robots facilitate somatosensory and proprioceptive input through mechanical stimulation (MS) in stroke survivors [4]. Compared with robotic designs relying on continuous passive movements, those promoting central-to-peripheral voluntary motor effort (VME) in the paretic limb (i.e., VME-driven robot) are more effective in facilitating rehabilitative neuroplasticity [5]. During physical exercise involving the paretic limb, central-to-peripheral VME is a primary driver of functional neuroplasticity for motor recovery post-stroke [6]. VME promotes descending excitation from the motor cortex to the target muscle, not limited to the cortical excitability alone as in motor imagery with mental practice only [7]. However, clinical evidence indicates that robot-assisted training with pure MS often resulted in poorer motor restoration in wrist–hand function compared to more proximal upper-limb segments [8]. More recently, VME-driven mechanical robots integrated with neuromuscular electrical stimulation (NMES) (i.e., VME-driven NMES-robots) are more effective in achieving wrist–hand motor function restoration than robots that provide purely MS during training [9]. Such VME-driven robots have offered major benefits in terms of alleviating co-contraction and improving the contraction levels of upper-limb distal muscles in chronic stroke survivors [10]. However, studies on this topic have reported the effects of treatment in terms of behavioral changes alone, employing clinical assessments and measurement of peripheral muscle activation, such as through electromyography (EMG) patterns, and they have not comprehensively examined the neuromodulatory effects associated with rehabilitation in the central nervous system. This has limited the understanding of both the immediate interaction and the long-term effects between the multi-modal assistance provided by VME-driven NMES-robots and the residual nervous system responses post-stroke. Further research is required to determine the sole and/or combined impacts of NMES and robot-provided MS on dynamic cortical modulation in the absence of VME, and to elucidate the combined effects of VME and the multi-modal assistance from the robot and NMES. Effective motor restoration in a paretic limb after a stroke requires repeated enhancement of targeted neuromuscular pathways on the hemiplegic side through neuromodulation by VME and somatosensory feedback, a process that recruits both descending and ascending neural tracks [11]. VME-driven NMES-robots provide multimodal somatosensory feedback by providing MS to joints and muscle-specific electrical stimulation, facilitating motor function restoration [9]. Exploring sensorimotor cortical dynamics is essential for elucidating the neuromodulatory mechanisms underlying such interventions because it reveals characteristic network reorganization patterns such as strengthened interhemispheric connectivity and restored ipsilesional activation [12, 13]. In this study, we used functional near-infrared spectroscopy (fNIRS) to investigate the neuromodulatory effects of an EMG-driven NMES-robots for wrist–hand rehabilitation, a representative VME-driven system developed by our team [14]. By detecting residual EMG signals in targeted upper-limb muscles, the EMG-driven NMES-robot capture VME and synergistically activate two independent assistive modalities: a closed-loop NMES modality and a compliant MS modality. In the closed-loop NMES modality, the detected EMG signals trigger the application of electrical stimulation through surface electrodes in the form of square-wave pulses, thereby inducing controlled muscle contractions. Compliant MS uses pneumatic actuation to support joints in extension and flexion, thereby promoting somatosensory and proprioceptive inputs. The EMG-driven NMES-robot also have a passive operation mode that implements predefined NMES protocols and offers gravity-compensated MS without VME involvement, which is beneficial for stroke survivors with severe motor impairments or during rest intervals. In the previous trial study, the EMG-driven NMES-robot offered considerable benefits with respect to improving wrist–hand movement in chronic stroke survivors [15]. These benefits may be attributable to the repeated reinforcement of specific neuromuscular pathways, with a process potentially modulated by VME in the descending pathway and multimodal somatosensory feedback. Multiple studies have indicated that combining VME with peripheral stimulation can substantially enhance cortical activation and neuroplasticity in individuals with upper-limb dysfunction. For example, Cui et al. [16] used fNIRS to evaluate cortical responses to NMES during different action states. They reported that NMES applied during action execution elicited greater cortical activation compared with NMES alone, which suggests a synergistic effect between peripheral stimulation and VME. Kato et al. [17] combined paired associative stimulation with voluntary limb movement and reported that this approach enhanced corticospinal excitability. Such reports are particularly relevant because neuroplasticity in the post-stroke brain is often characterized by reduced network connectivity and diminished cortical activation compared to unimpaired individuals, meanwhile these neurocircuit alterations are also projected to deviated motor performance assessed by clinical behavioral scores [18]. Studies also have indicated that multi-modality stimulation may offer benefits with respect to neural regulation. For example, a study involving local vibration to muscles and task-oriented therapy demonstrated improvements in motor learning and function, suggesting potential neuromodulatory benefits [19]. Liu et al. identified distinct cortical activation patterns during soft robotic glove training and virtual reality-based rehabilitation, highlighting the differential engagement of sensorimotor networks in varying therapeutic contexts [20]. In the present study, we examined the cortical dynamic neuromodulatory effects of EMG-driven NMES-robot assistance for wrist–hand extension in stroke survivors. We specifically focused on cortical responses to the sole and/or combined effects of NMES and MS in the absence of VME and on cortical responses to the presence of VME and NMES-robot multi-modal assistance. We used fNIRS to assess cortical responses, with cortical activation and network connectivity as the primary outcomes for evaluating neuromodulatory effects. We hypothesized that combining NMES with MS, whether delivered with or without VME, would elicit stronger cortical somatosensory responses than either modality alone, as reflected by increased activation and enhanced network connectivity approaching normal patterns. Furthermore, we hypothesized that VME would enhance the sensorimotor network connectivity across all stimulation modes compared with those without VME. Methods In this study, we compared the immediate cortical responses during the use of multimodal somatosensory stimulation provided by an EMG-driven NMES-robot in individuals with chronic stroke to those observed in unimpaired controls. The cortical responses to different assistive schemes were assessed by measuring activation level, network connectivity and topological variations. This study was conducted in accordance with the Declaration of Helsinki. Participants The current study protocol was approved by the institutional review board of Northwestern Polytechnical University and was registered on the Chinese Clinical Trial Registry with the identifier ChiCTR2200057839. Informed consent was obtained from all participants, who were chronic stroke survivors (stroke group: N = 18, mean age: 57.22 ± 13.70 years; female/male:6/12) and demographically matched unimpaired individuals (unimpaired group: N = 20, mean age: 61.56 ± 6.60 years; female/male:11/9). We included an unimpaired control group to provide normative benchmarks for quantifying poststroke neurophysiological performance. Inclusion criteria for chronic stroke were: (1) confirmation of stroke through either computed tomography or functional magnetic resonance imaging; (2) at least 6 months after onset, with ages ranging from 18 to 75 years; (3) the spasticity at the wrist and fingers were ≤ 3 as measured by the Modified Ashworth Scale (MAS); (4) Montreal Cognitive Assessment (MOCA) score ≥ 21; (5) passive range of motion of the wrist joint (extension: 45° to flexion: 60°); (6) the voluntary surface electromyographic (sEMG) signals can be detected on the affected wrist–hand extensors (i.e. three standard deviations (SDs) above the baseline). Inclusion criteria for the unimpaired controls were: (1) aged between 50 and 75 years; (2) without cardiovascular diseases, neurological disorders, mental health conditions; (3) experiencing no limitations in activities of daily living; (4) dominate side is right. Table 1 provides the demographic characteristics of the participants, and the Upper Extremity Fugl-Mayer Assessment Scores (UE-FMA) of stroke group. Table 1Demographic characteristics of participantsCharacteristicsStroke group (n = 18)Unimpaired group (n = 20)Age in years (mean ± SD) (range)57.22 ± 13.70 (25–75)61.56 ± 6.60 (52–70)Gender (male/female)12/69/11Affected side (right/left)12/6NAStroke type (ischemic/hemorrhagic)9/9NATime after stroke (month) (mean ± SD) (range)12.14 ± 8.59 (6–36)NAUE-FMA scores (mean ± SD) (range)19.72 ± 16.75 (4–52)NA Table 1Demographic characteristics of participantsCharacteristicsStroke group (n = 18)Unimpaired group (n = 20)Age in years (mean ± SD) (range)57.22 ± 13.70 (25–75)61.56 ± 6.60 (52–70)Gender (male/female)12/69/11Affected side (right/left)12/6NAStroke type (ischemic/hemorrhagic)9/9NATime after stroke (month) (mean ± SD) (range)12.14 ± 8.59 (6–36)NAUE-FMA scores (mean ± SD) (range)19.72 ± 16.75 (4–52)NA Demographic characteristics of participants Experimental setup The experiment was conducted in a controlled and quiet environment. Figure 1A illustrates the experimental setup with an EMG-driven NMES-robotic system (Xiamen Zhanhong Chuangjian Technology Co., Ltd., China) [10, 21] which provided MS and NMES assistance to the targeted wrist–hand joint, synchronized with a multi-channel near-infrared system (NirSmart, Danyang Huichuang Medical Equipment Co., LTD., China), which was used to continuously measure and record the concentrations of oxy hemoglobin (oxy-Hb) and deoxyhemoglobin (deoxy-Hb) during the experiment. The targeted joint motion in this study was wrist–hand extension. Fig. 1Experimental setup integrated with a multimodal somatosensory stimulation platform. A Custom-designed setup integrating an EMG-driven NMES-robot and fNIRS. B The fNIRS opotode arrangement covering ROIs. C Experimental paradigm: Conditions: without VME and with VME; somatosensory inputs: MS, NMES, and MS and NMES. D Block design with a randomized trial sequence, consisting of six distinct tasks, each repeated three times. The detailed descriptions of experimental tasks are listed in Table S1 Experimental setup integrated with a multimodal somatosensory stimulation platform. A Custom-designed setup integrating an EMG-driven NMES-robot and fNIRS. B The fNIRS opotode arrangement covering ROIs. C Experimental paradigm: Conditions: without VME and with VME; somatosensory inputs: MS, NMES, and MS and NMES. D Block design with a randomized trial sequence, consisting of six distinct tasks, each repeated three times. The detailed descriptions of experimental tasks are listed in Table S1 The EMG-driven NMES-robot consisted of a wearable glove with a textile bracing, two-channel NMES and EMG detection electrode pairs, and a control box (650 g) powered by a 12 V direct-current battery. The wearable glove had an embedded musculoskeletal structure with a fabric-based pneumatic artificial muscle acting as a soft actuator, which was connected to a lightweight 3D-printed exoskeletal connector fixed on the palm side [10]. During the experiment, the pneumatic muscle generated compliant flexion and extension forces through pressurized air, and the 3D-printed connector served as a mechanical interface for transmitting actuator forces to the finger joints. This design provided MS to the targeted joint through synchronized pneumatic compression and guided joint excursion. During inflation, the pneumatic muscle (range:0–60 kPa) achieved a range of motion of 0° to 60° wrist–hand extension, with a constant angular velocity of 10°/s. Two-channel NMES was applied to the wrist–hand extensors (i.e., the extensor carpi ulnaris (ECU) and the extensor digitorum (ED)), to facilitate wrist–hand extension. Because of the identical anatomical positions of the ECU and ED two pairs of reusable surface electrodes (5 × 5 cm2, PALS Neurostimulation Electrodes, Axelgaard Manufacturing Co., Ltd., Fallbrook, CA, USA) were used for both NMES and EMG detection. The NMES outputs were square-wave pulses with a constant amplitude of 70 V, a stimulation frequency of 40 Hz, and a manually adjustable pulse width ranging from 0 to 300 µs, with a threshold pulse width set to evoke maximal muscle contraction [14]. A reference electrode (2 × 3 cm2, Blue Sensor N, Ambu Inc., Ballerup, Denmark) was attached to the skin surface over the olecranon to minimize common-mode noise. To facilitate phasic wrist–hand movements, ECU-ED muscle union was used as a voluntary neuromuscular driver, with it used to control the MS and NMES assistance provided by the proposed system. EMG-driven control was adopted to facilitate this proves. For each motion phase, a value of three times SDs above the EMG baseline in the resting state was set as a threshold. Once the EMG activation level of the ECU-ED muscle union reached the preset threshold, NMES was applied to the union, and MS was provided to the wrist–hand structure through an inflated pneumatic muscle to facilitate joint extension throughout the motion phase. Before the experiment, all participants were instructed to sit quietly for 5 min to ensure they acclimated to the environment. They were then asked to assume a comfortable seated position with their torso approximately 10 cm from the table’s edge. Their upper limbs were extended forward, and their elbows were maintained at approximately 170° of flexion, supported by a cushion (Fig. 1A). Next, the pneumatic muscles were dynamically calibrated at 10°/s to match individual maximal wrist–hand extension ranges. The intensity of NMES was then calibrated (2 mA steps) to evoke maximal wrist–hand extension within a tolerable range (16–36 mA), with discomfort considered to occur at a Visual Analog Scale score of < 3/10. A multi-channel near-infrared system (NirSmart, Danyang Huichuang Medical Equipment Co., LTD., China) was used to continuously measure and record cortical responses during the experiment. A standard cap compatible with the international 10–20 system was used to reference the layout of NIR spectroscopy probes. The distance between the sources and detectors was standardized to 3 cm. A total of 40 channels were used to cover the region of interests (ROIs) [22], i.e., the prefrontal cortex (PFC), supplementary motor cortex (SMC), primary somatosensory cortex (PSC), primary motor cortex (PMC), and somatosensory association cortex (SAC) (Fig. 1B). Experimental protocol Figure 1C and D detail the experimental protocol featuring somatosensory stimulations (MS, NMES, and their combined delivery) under both with and without VME conditions. The experimental tasks (Fig. 1D) were presented in a random order to limit the influence of participants’ anticipating subsequent tasks. Tasks 1–3 were performed without VME, with wrist–hand somatosensory stimulation as follows: (1) MS only, (2) NMES only, and (3) combined MS and NMES for co-stimulation. Tasks 4–6 were performed when VME was generated from the affected forearm, with wrist–hand somatosensory stimulation by the same MS and NMES arrangements as in Tasks 1–3, and the details are presented in Table S1. Under the condition without VME (Tasks 1–3), participants were asked to stay passive while receiving the somatosensory stimulation delivered through the system. In this condition, the system delivered the MS and NMES in a preprogrammed mode without detecting EMG. Under the condition with VME (Tasks 4–6), participants were asked to execute motor tasks under the experimenter’s guidance, including “wrist–hand extension”, “extension maintenance”, and “relaxation”. When the surface EMG signals from the ECU-ED unit during “wrist–hand extension” exceeded the preset threshold, somatosensory stimulation was automatically triggered. During the 30-second stimulation phase, participants were instructed to perform a single, sustained wrist–hand extension and hold it for the entire duration of the stimulation period. Each task included three trials, with each trial consisting of a 30-s stimulation phase followed by a 30-s rest period (Fig. 1D). This block design and duration were selected to ensure effective capture of hemodynamic response (which typically peaks approximately 5–7 s after stimulus onset) and minimize participant fatigue, in accordance with the established fNIRS methodology [23–25]. A 40-channel fNIRS system was used to record the concentrations of oxy-Hb and deoxy-Hb throughout the experiment, including during baseline recording, during task execution, and after the task resting state. The sampling frequency was 11 Hz, with wavelengths of 730 and 850 nm. The hemiplegic limbs of stroke survivors and the dominant limbs of unimpaired individuals were assessed. fNIRS analysis Data analysis was performed using MTLAB (MathWorksTM Inc., Massachusetts, USA) and NirSpark software (NirSmart, Danyang Huichuang Medical Equipment Co., LTD., China). Offline fNIRS analysis (Fig. 2) was conducted to assess cortical activation, network connectivity, and topological characteristics to determine the underlying neuromodulatory effects across the experimental conditions specified in Table S1. Specifically, the cortical activation of ROIs was quantified using beta coefficients derived from a general linear model (GLM), with these coefficients serving as indices of “cortical activation”. The network connectivity was assessed using two complementary approaches, i.e., functional connectivity (FC) and effective connectivity (EC). The FC between ROIs was analyzed using Pearson correlation, while EC between ROIs was evaluated with Granger Causality Analysis (GCA). The topological characteristics of the brain network were further characterized using graph-theoretical metrics, consisting of clustering coefficient (CC), global efficiency (Eg), degree centrality (Dc), and nodal local efficiency (NLe). Table S2 presented a summary on these measures. Fig. 2Flow chart of fNIRS pre-processing and analysis Flow chart of fNIRS pre-processing and analysis The fNIRS data acquired from stroke survivors with radiologically confirmed right-hemisphere lesions were computationally mirrored across the midsagittal plane. Spatial transformation was then applied to standardize hemispheric alignment as follows: all lesioned hemispheres were represented in the left hemisphere space, and right-sided sensorimotor measurements consistently corresponded to paretic limbs during subsequent analyses. The ROIs on the affected side (ipsilesional) were labeled as iROI, i.e., iPFC, iSMC, iPSC, iPMC, and iSAC, whereas the ROIs on the unaffected side (contralesional) were labeled as cROIs, i.e., cPFC, cSMC, cPSC, cPMC and cSAC. In the unimpaired group, the right upper limbs were the dominant side, and the ROIs of right hemisphere were labeled as RROI, while the ROIs of left hemisphere were LROI. The quality of each channel was evaluated through a visual inspection by two independent and experienced researchers and the application of a quantitative criterion based on signal-to-noise ratio in which channels with a raw intensity standard deviation exceeding 15% of the mean were excluded. Motion artifacts were identified using a moving standard deviation method with a 1-s window and 0.5-s step. Windows with a local standard deviation exceeding six times the global standard deviation of the signal were classified as artifacts and corrected using cubic spline interpolation [26]. A band-pass filter (0.01–0.08 Hz) [27, 28] was then applied to remove both physiological artifacts (e.g., respiration and heart activity) and low-frequency baseline drift. The modified Beer-Lambert law was used to convert optical intensity signals into oxyhemoglobin and deoxyhemoglobin concentration changes. To further suppress global physiological interference and superficial hemodynamic contributions, Principal Component Analysis (PCA) followed by Independent Component Analysis (ICA) was conducted. The number of principal components removed during PCA was determined by a cumulative variance-explained threshold of 0.90, aiming to remove global physiological interference while minimizing the risk of discarding genuine hemodynamic signals. This threshold is consistent with the range reported in stroke-related fNIRS studies [29, 30]. Subsequently, ICA was performed on the PCA-corrected signals, and independent components were rejected if their power spectra exhibited dominant energy outside the hemodynamic response frequency band of interest (0.01–0.08 Hz) [27, 28]. For each experimental task, oxy-Hb time courses were first temporally aligned across all trials, and then overlapped and averaged to generate condition-specific hemodynamic response profiles corresponding to their reliable reflection of neuronal activity-induced vascular responses [23]. For evaluation of task-related cortical activation, a GLM was used to calculate the beta coefficients for each subject and channel across different tasks. The GLM was defined as follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y={x}_{1}{\beta\:}_{1}+{x}_{2}{\beta}_{2}+{x}_{3}{\beta}_{3}+e$$\end{document} where Y is the response variable, which is typically a measure reflecting brain activity, such as ∆oxy-Hb; x is the design matrix, which captures the permutations of experimental conditions; β is the estimated parameter value, which reflects the effect of each experimental condition on cortical activity (crucial for subsequent analyses), and e is the error term. Connectivity (i.e., FC and EC) analysis is a method for characterizing the interactions between different neuronal activity patterns at the cortical level. FC refers to the statistical dependence between two spatially separated regions, reflecting the synchrony of fluctuations in neural activities over time. In this study, the FC between two ROIs was analyzed using Pearson correlation of their time series, providing an index of their temporal co-activation. EC facilitates the quantification of how neural activity in a brain region influences activity in another region, offering insights into directed interactions and causal relationships in information flow in neural network [31]. In the present study, we conducted GCA for EC analysis. To evaluate the main information flow between different ROIs, we calculated the Dominant Direction of Information Flow (DIF). We denote the DIF from one ROI to another using dEC, defined as follows:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${dEC}_{x\to\:y}=\frac{{EC}_{x\to\:y}-{EC}_{y\to\:x}}{{EC}_{x\to\:y}+{EC}_{y\to\:x}}$$\end{document} where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{x\to\:y}$$\end{document} is equal to [-1,1], 0<\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{x\to\:y}$$\end{document}≤1 refers to information flow from signal x to signal y; -1≤\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{x\to\:y}$$\end{document}<0 refers to information flow from signal y to x, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{x\to\:y}=0$$\end{document} refers to the absence of information flow between signals x and y. On the basis of the connectivity calculation results, we conducted graph theoretical analysis (GTA) by using the Gretna toolbox [32] to delineate the functional segregation and integration features. Global variables such as CC and Eg, and local variables such as Dc and NLe were also calculated. Eg characterizes the efficiency of a system transporting information in parallel, CC measures the degree of clustering in a network, NLe represents the fault tolerance level of a network in response to the removal of a node, Dc reflects the importance of an individual node in a network [33, 34]. Statistical analysis All statistical analyses were conducted using IBM SPSS Statistics software (v. 26.0, IBM, Armonk, NY, USA), as summarized in Fi g 3. All precalculated neurophysiological parameters, including cortical activation, network connectivity, and topological characteristics, were statistically analyzed. Data normality and homogeneity of variance were assessed at the group level with the Shapiro-Wilk test and the Levene test, respectively, to determine whether a parametric or a nonparametric approach should be selected. After the Shapiro-Wilk test of normality and Levene test of homogeneity, cortical activation, EC, CC, Eg, and Dc, were confirmed to be normally distributed and homogeneity (p > 0.05), while FC and NLe were non-normally distributed (p ≤ 0.05) and was thus analyzed using nonparametric methods. For each experimental task, one-sample t-test was employed to evaluate task-related cortical activation by testing the beta coefficients against zero. To evaluate task-dependent neuromodulatory effects, within-group differences under each experimental condition (without VME: Tasks 1–3; with VME: Tasks 4–6) were assessed using one-way repeated-measures analysis of variance (ANOVA) for normally distributed variables (cortical activation, EC, CC, Eg, and Dc) and the Friedman test for non-normally distributed variables (FC and NLe). Between-group differences in cortical responses (stroke group compared with unimpaired group) during the same tasks were assessed through the independent-samples t-test for cortical activation, EC, CC, Eg, and Dc, and the Mann-Whitney U test for FC and NLe. The effect of VME (i.e., Task 4 compared with Task 1, Task 5 compared with Task 2, and Task 6 compared with Task 3) was examined through paired t test for cortical activation, EC, CC, Eg, and Dc, and Wilcoxon’s tests for FC and NLe. The false discovery rate (FDR) correction was applied independently to each set of results derived from a specific statistical test and metric. For instance, in the assessment of the task-related cortical activation, FDR correction was performed on the family of 10 p-values generated from the 10 ROIs. For FC and EC, the correction families consisted of 45 and 90 p-values, respectively, corresponding to the total number of spatial connections analyzed. For NLe and Dc, the correction family for each metric consisted of 10 p-values corresponding to the 10 ROIs. The Eg and CC had a family-wise number of 1. All p-values reported were corrected for FDR. Additionally, as previous studies [35, 36] have demonstrated that EC was associated with motor performance post-stroke, we conducted a partial correlation analysis between task-related EC and UE-FMA scores with the onset time of stroke considered as a covariate, to explore the association between the severity of motor injuries and neuroplastic impairments at the cortical level. The level of statistical significance was set at p< 0.05 in this work (Fig. 3). Fig. 3The logic flow of statistical analysis The logic flow of statistical analysis Results VME-induced cortical network connectivity Compared with FCs under the somatosensory stimulations without VME involvement, those with VME demonstrated significantly higher connections in the sensorimotor network in both stroke and unimpaired groups when MS and NMES were applied together (Fig. 4B). Specifically, in the unimpaired group, VME significantly increased the FC in the following region pairs: LPFC-LSMC (Z = 3.13, p = 0.02, r = 0.71), RPFC-LPSC (Z = 3.09, p = 0.02, r = 0.69), RPFC-LPMC (Z = 2.91, p = 0.03, r = 0.65), RPFC-RPMC (Z = 3.13, p = 0.02, r = 0.70), RPFC-RPSC (Z = 2.99, p = 0.03, r = 0.67), and RPFC-RSAC (Z = 3.47, p = 0.02, r = 0.77). In the stroke group, VME significantly increased the FC in the region pairs of iPFC-iSAC (Z = 3.42, p = 0.02, r = 0.83), cPFC-iSAC (Z = 3.24, p = 0.03, r = 0.75), and cPSC-cSAC (Z = 2.85, p = 0.047, r = 0.67) but significantly decreased the FC in the region pair of cSMC-iPMC (Z = 3.07, p = 0.03, r = 0.72) (Fig. 4B). In addition, the involvement of VME significantly increased the hemodynamic DIF into the LSMC in the unimpaired group, as evidenced by the significant increases in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{RPMC\to\:LSMC}$$\end{document} during Task 4 (MS with VME) (t (19) = 4.43, p = 0.03, Cohen’s d = 0.99), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{LSAC\to\:LSMC}\:$$\end{document}during Task 5 (NMES with VME) (t (19) = 4.20, p = 0.04, Cohen’s d = 0.94), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{LPSC\to\:LSMC}$$\end{document} during Task 6 (MS and NMES with VME) (t (19) = 4.20, p = 0.04, Cohen’s d = 0.94) (Fig. 4A). In the stroke group, VME induced distinct DIF patterns. Specifically, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{iPSC\to\:cPFC}$$\end{document} (t (17) = 4.50, p = 0.03, Cohen’s d = 1.06) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{cPSC\to\:cPMC}$$\end{document} (t (17) = 4.14, p = 0.03, Cohen’s d = 0.97) were significantly higher in Task 4 than in Task 1 (MS alone). In contrast, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{iPSC\to\:cPSC}$$\end{document} (t (17) = 4.31, p = 0.02, Cohen’s d = 1.02) was significantly higher, whereas \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{cPFC\to\:iPSC}$$\end{document} (t (17) = -3.82, p = 0.04, Cohen’s d = 0.90) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{cSAC\to\:iPSC}$$\end{document} (t (17) = -4.49, p = 0.02, Cohen’s d = 1.06) were significantly lower in Task 5 compared with Task 2 (NMES alone) (Fig. 4A). These alterations may reflect an adaptive reorganization of cortical connectivity post-stroke. The detailed numerical results are summarized in Table S4. Fig. 4Condition-dependent modulation of cortical connectivity patterns. A EC reorganization with and without VME. B FC changes between conditions with and without VME Condition-dependent modulation of cortical connectivity patterns. A EC reorganization with and without VME. B FC changes between conditions with and without VME Cortical responses to somatosensory stimulation with and without VME Responses in the prefrontal and somatosensory cortex during VME Responses in the prefrontal and somatosensory cortex during VME Responses in the prefrontal and somatosensory cortex during VME The stroke group exhibited distinct neuromodulatory patterns during VME. Cortical activation analysis suggested that somatosensory stimulation with VME engaged different cortical regions. Specifically, Task 4 predominantly elicited contralesional activation, i.e., in cPSC (t = 3.32, p = 0.04, Cohen’s d = 0.78) and cSAC (t = 3.03, p = 0.04, Cohen’s d = 0.71), whereas Task 5 primarily induced ipsilesional activations, i.e., iSAC (t = 3.42, p = 0.03, Cohen’s d = 0.81). Task 6 synergistically integrated the cortical activation patterns of the two modalities, leading to enhanced ipsilesional hemispheric iPMC activation (t = 2.95, p = 0.04, Cohen’s d = 0.70) during movement (Fig. 5A). However, one-way repeated-measures ANOVA showed no significant task-related differences among VME tasks (Tasks 4–6) in the stroke group. The detailed numerical results are summarized in Table S5.1. Fig. 5Cortical processing during somatosensory stimulations with VME. A Group-level cortical activation patterns. B Intergroup differences in FC and EC. C Intergroup differences in NLe and Dc Cortical processing during somatosensory stimulations with VME. A Group-level cortical activation patterns. B Intergroup differences in FC and EC. C Intergroup differences in NLe and Dc By contrast, in the unimpaired group, Task 4 selectively activated the RPFC (t = 3.29, p = 0.04, Cohen’s d = 0.73), whereas Task 5 demonstrated preferential engagement with the LSAC (t = 3.38, p = 0.03, Cohen’s d = 0.76) (Fig. 5A). Task 6 evoked significantly stronger bilateral cortical activations compared with either sole stimulation, particularly in the LPFC (F (2, 18) = 8.45, p = 0.00, η²ₚ = 0.51), LSMC (F (2, 18) = 5.31, p = 0.04, η²ₚ = 0.34), LPMC (F (2, 18) = 4.65, p = 0.04, η²ₚ = 0.54) and LPSC (F (2, 18) = 4.12, p = 0.04, η²ₚ = 0.13) (Fig. 5A). The detailed numerical results are summarized in Table S5.1. Comparative analysis revealed distinct cortical response patterns in the prefrontal and somatosensory motor network between the stroke group and the unimpaired group. During Task 4, the FC of iPFC-iPSC (Z = -3.41, p = 0.01, r = 0.55), iPFC-cPSC (Z = -3.17, p = 0.03, r = 0.51), and iPFC-cSAC (Z = -3.55, p = 0.01, r = 0.58) was significantly lower in the stroke group compared with the unimpaired group (Fig. 5B). The stroke group also exhibited a significantly lower DIF in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{cPMC\to\:iPFC}$$\end{document} (t =-3.75, p = 0.03, Cohen’s d = 1.25) compared with the unimpaired group. By contrast, a significant increase was observed in\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\:dEC}_{iPSC\to\:iPMC}$$\end{document} (t = 3.47, p = 0.04, Cohen’s d = 1.15) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{cPSC\to\:cPMC}$$\end{document} (t = 3.58, p = 0.03, Cohen’s d = 1.19) (Fig. 5B). Moreover, NLe was significantly elevated in both the ipsilesional and contralesional PMC (iPMC: Z = 3.14, p = 0.02, r = 0.51; cPMC: Z = 2.74, p = 0.03, r = 0.45) in the stroke group (Fig. 5C). During Task 5, a significant decrease was observed in the FC of iSMC-iPSC (Z = 3.14, p = 0.04, r = -0.51) and iSMC-iPMC (Z = 3.49, p = 0.02, r = -0.57) in the stroke group (Fig. 5B). A significant decrease was also observed in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{cPMC\to\:iSMC}$$\end{document}(t = -3.89, p = 0.04, Cohen’s d = 1.30), although a significant increase was observed in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{iPSC\to\:cPSC}$$\end{document} (t = 3.72, p = 0.04, Cohen’s d = 1.24) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{iPSC\to\:cSAC}$$\end{document} (t = 3.40, p = 0.048, Cohen’s d = 1.26) compared with the unimpaired group (Fig. 5B). Additionally, a significant decrease in Dc was observed in the iSMC (t = -3.00, p = 0.04, Cohen’s d = 1.00) that was accompanied by a significant increase in the Dc of the cSAC (t = 2.75, p = 0.04, Cohen’s d = 0.89) in the stroke group (Fig. 5C). Importantly, inter-group comparisons demonstrated that during Task 6, the stroke group exhibited a significant decrease in the FC of iSMC-iPSC (Z = 3.32, p = 0.04, r = -0.54) (Fig. 5B) and a significant increase in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{iPSC\to\:iSAC}$$\end{document} (t = 3.79, p = 0.049, Cohen’s d = 1.26) (Fig. 5B). The detailed numerical results are summarized in Table S5.2. These findings suggest that Task 6 reduced group-level differences in cortical network organization, resulting in more comparable connectivity patterns between stroke and unimpaired participants. 2.Cortical sensitivity to somatosensory stimulation 2.Cortical sensitivity to somatosensory stimulation Cortical sensitivity to somatosensory stimulation In both groups, Task 3 (combining MS and NMES) induced greater bilateral cortical activation (Fig. 6A) compared with Task 1 (MS alone) or Task 2 (NMES alone). This cortical response was particularly pronounced in the unimpaired group, where one-way repeated-measures ANOVA revealed significant differences at the RSMC (F (2, 18) = 4.71, p = 0.048, η²ₚ = 0.20), LPSC (F (2, 18) = 5.75, p = 0.04, η²ₚ = 0.23), RPMC (F (2, 18) = 5.45, p = 0.04, η²ₚ = 0.22), and LPMC (F (2, 18) = 5.39, p = 0.04, η²ₚ = 0.22)) (Fig. 6A (2)). In the stroke group, we also observed the same cortical response pattern that Task 3 induced greater bilateral cortical activation than Task 1 or Task 2. However, one-way ANOVA revealed no significant differences among different tasks, indicating that the stroke group may exhibit reduced cortical sensitivity to multimodal somatosensory stimulation. This attenuated response was further reflected in network connectivity and topological characteristics: aside from a significant task-dependent difference in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{cPFC\to\:cPSC}$$\end{document} (F (2,16) = 9.76, p = 0.04, η²ₚ = 0.37) (Fig. 6B (2)), no other global or local variables showed significant task-related changes in the stroke group (Fig. 6C (2)). Notably, Task 2 (NMES alone) elicited ipsilesional hemispheric activation in the stroke group, as evidenced by significant activation in the i-SMC (t = 3.50, p = 0.03, Cohen’s d = 0.83) — a response that was absent in the unimpaired group (Fig. 6A (1)), pointing to a distinct neurophysiological pattern post-stroke. The detailed numerical results are summarized in Table S6. Fig. 6Task-dependent cortical responses to somatosensory stimulations without VME. A Cortical activation patterns in the two groups. B EC variations across different experimental tasks. C NLe across different experimental tasks Task-dependent cortical responses to somatosensory stimulations without VME. A Cortical activation patterns in the two groups. B EC variations across different experimental tasks. C NLe across different experimental tasks Analysis of the unimpaired group revealed that the RPFC and LPSC were critically involved in the differentiation between distinct forms of somatosensory stimulation. Specifically, the RPFC received enhanced directed information flow from other cortical regions during the combined MS and NMES condition (Task 3) (Fig. 6B (1)), suggesting that the application of MS and NMES was associated with increased inter-regional cortical communication involving the RPFC. In addition, the LPSC exhibited a robust response to NMES, with a marked increase in NLe during both Task 2 and Task 3 compared to Task 1 ((χ² (2) = 11.42, p = 0.03, Kendall’s W = 0.29)) (Fig. 6C (1)). The detailed numerical results are summarized in Table S6.2. Correlation between clinical scores and fNIRS parameters Correlation analysis revealed clinically promising correlations between the DIF and UE-FMA score during somatosensory stimulation with VME. Specifically, a significant positive correlation was observed between ipsilesional DIF output and UE-FMA score during Task 4 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d{EC}_{iPFC\to\:cPFC}$$\end{document}-UE_FMA, r = 0.51, p = 0.04;\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\:\mathrm{d}EC}_{iSMC\to\:cPFC}$$\end{document}-UE_FMA, r = 0.52, p = 0.03) (Fig. 7A). In contrast, significant negative correlations were observed between contralesional DIF output and UE-FMA scores, including\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{dEC}_{cPFC\to\:cSAC}$$\end{document} during Task 5 (r = -0.60, p = 0.01) and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\:\mathrm{d}EC}_{cPSC\to\:iPMC}$$\end{document} during Task 6 (r = -0.63, p = 0.007) (Fig. 7B). These observed correlations provide preliminary evidence of a potential relationship between interhemispheric information flow and motor outcome. However, due to the exploratory nature of these analyses and the limited sample size, these correlations should be interpreted with caution. Fig. 7Clinical-neural correlations: UE-FMA and EC during somatosensory stimulations with VME. A Ipsilesional outflow positively correlates with UE-FMA score. B Contralesional outflow negatively correlates with UE-FMA score Clinical-neural correlations: UE-FMA and EC during somatosensory stimulations with VME. A Ipsilesional outflow positively correlates with UE-FMA score. B Contralesional outflow negatively correlates with UE-FMA score Discussion In this study, functional near infrared spectroscopy was utilized to investigate the cortical responses by an electromyography-driven neuromuscular electrical stimulation-robot to facilitate wrist–hand extension in chronic stroke. Our goal was to explore the interaction between the robot and the nervous system, with EMG used to capture voluntary motor effort. Our results indicate that mechanoelectrical somatosensory stimulation involving both MS and NMES activated broader cortical areas compared with that involving MS or NMES alone. Task 6 (when both MS and NMES applied and VME involved) reduced the differences in the cortical networks between the stroke and unimpaired groups, leading to similar cortical network connectivity patterns in the two groups. VME reshaping sensorimotor dynamic connectivity VME significantly enhanced the hemodynamic information flow into the supplementary motor cortex in the unimpaired group, whereas no comparable modulation was detected in the stroke group (Fig. 4). Such absence of SMC modulation may reflect stroke induced network reorganization and lesion characteristics. Previous studies have shown that better motor recovery after stroke is associated with the restoration of brain functional organization toward patterns observed in the healthy brain [37], suggesting that promoting such restoration may represent a promising rehabilitation strategy. In this context, the lack of SMC directed modulation in the stroke group raises the possibility that the SMC may serve as a relevant cortical target for rehabilitation. Targeted re-engagement of this region during VME might facilitate the restoration of functional patterns that are typically observed in unimpaired individuals but disrupted after stroke. This interpretation is supported by converging evidence showing that the SMC plays a central role in internally generated movement initiation and coordination [38], that its activation patterns correlate with volitional control measures during VME in the affected hand [39], and that SMC-derived neurophysiological signatures can be decoded to drive robotic hand movements, thereby establishing closed-loop sensorimotor feedback systems [40, 41]. Future studies with larger and clinically stratified stroke cohorts are needed to determine how these factors influence SMC engagement during VME. Functional connectivity analysis revealed no significant differences between Task 4 and Task 1 or Task 5 and Task 2, whereas a significant increase was observed in Task 6 compared with Task 3 (Fig. 4B). The results indicated that VME added to a single-modality stimulation (MS or NMES alone) was insufficient to enhance FCs, whereas combining VME with multimodal somatosensory inputs (MS together with NMES) from the peripheral end enhanced the VME influence and elicited greater network connectivity at the cortical level. Such sensorimotor demand-gated amplification is consistent with previous study showing that FCs become stronger under more complex task conditions compared with those simpler ones [42]. Kang et al. reported that the dorsolateral prefrontal cortex increases its communication with the primary motor cortex specifically during movement planning and execution, indicating that prefrontal-motor coupling is further strengthened as task demands rise [43]. Complex dual stimulation during Task 6 triggered greater demands on planning and motor intention, thereby amplifying prefrontal-motor network engagement and revealing VME related FC differences. Although no significant differences were observed in network connectivity between Task 1 and Task 4, or between Task 2 and Task 5, these findings suggest that VME alone may not provide sufficient influence to enhance post stroke neural connectivity under the sole stimulation conditions. Conversely, the results also imply that using NMES or MS alone during passive movements, i.e., without VME, can evoke comparable neural network patterns in rehabilitative contexts. This may help explain the beneficial rehabilitative effects reported in studies employing continuous passive motions [44, 45]. We observed that Task 6 (when both MS and NMES applied and VME involved) emerged as the optimal method for inducing dynamic neuromodulatory effects in chronic stroke survivors. This approach demonstrated unique advantages in mitigating the differences observed in the cortical networks between the stroke and unimpaired groups, resulting in a cortical connectivity pattern in the stroke group that closely resembled that of the unimpaired group, and in enhancing cortical network connectivity across different motor-related regions (Fig. 5). These results are consistent with those of previous studies indicating that VME-driven NMES-robots outperform conventional robotic or passive assistance systems in terms of upper-limb motor function recovery in stroke survivors [9, 46]. Overall, our findings comprehensively elucidate the process of dynamic neuromodulation in interventions involving VME-driven NMES-robots, offering key insights into the mechanism of neuroplasticity. Specifically, our between-group comparisons reveal that chronic stroke survivors exhibit varying cortical remolding patterns across different tasks. Alterations in cortical plasticity in stroke survivors may not lead to improvements in motor function, potentially resulting in detrimental cortical remodeling, a phenomenon referred to as maladaptive plasticity [47]. Severe damage and excessive training may cause neural plasticity to progress toward maladaptive pathways, leading to conditions such as phantom pain and muscle tension [48]. Maladaptive plasticity can further diminish post-stroke motor function and impede movement recovery [49]. Therefore, promoting the beneficial cerebral cortex reorganization after a stroke is essential. The current work also highlights the importance of selecting appropriate rehabilitation strategies and stimulation parameters in rehabilitation programs aimed at addressing upper-limb motor function impairments in stroke survivors. Combining MS and NMES together with VME may facilitate cortical reorganization in chronic stroke survivors. Prefrontal dynamic engagement in sensorimotor modulation with VME In this study, the chronic stroke survivors exhibited distinct activation patterns in their prefrontal and sensorimotor cortices during somatosensory processing with VME (Fig. 5). The prefrontal cortex, a higher-order neocortical region characterized by extensive reciprocal connections, - exhibits functionally crucial coactivation with primary and associated sensory cortices, motor planning and execution networks, and subcortical structures [50]. This widespread connectivity pattern indicates that PFC engagement during VME tasks may facilitate the top-down modulation of sensorimotor integration, error monitoring and movement correction, and cognitive-motor resource allocation. In addition, the PFC plays a crucial role in sensory integration and decision-making, coordinating cortical responses to guide motor execution depending on the incoming information [51]. Stroke-induced dyskinesias trigger specific PFC network reorganization, with functional adaptation affecting both cognitive and motor function [52, 53]. Active motor execution elicits stronger sensorimotor activation compared with passive movement [54], reflecting PFC-mediated processes such as attention allocation, working memory engagement, and cognitive-motor integration [55, 56]. During the tasks with VME involvement, cortically-generated motor intentions produce electromyographic signals that elicited multimodal stimulation, indicating that PFC reorganization may influence both motor execution and somatosensory integration in chronic stroke. In the present study, correlation analysis revealed VME interacts with the type of somatosensory input to recruit distinct cortical control strategies after stroke (Fig. 7). The stronger \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{iPFC\to\:cPFC}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{dEC}_{iSMC\to\:cPFC}$$\end{document} connections associated with higher UE-FMA scores may be attributed to natural proprioceptive feedback and VME during Task 4, which likely engaged prefrontal and frontoparietal executive resource. This finding was consistent with evidence that attention to reliable proprioceptive input strengthened sensorimotor executive coupling [57, 58], and that proprioceptive information was integrated via parietal–prefrontal pathways during voluntary movement [58]. The negative correlation observed between UE-FMA and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d{EC}_{cPFC\to\:cSAC}$$\end{document} during Task 5 indicated that electrically evoked, non-physiological cutaneous input may reduce the reliability of somatosensory evidence, thereby increasing the compensatory recruitment of prefrontal top-down modulation of sensory cortex, a pattern consistent with findings reported in previous studies [59, 60]. Multimodal afferent stimulation during Task 6 appeared to amplify cross-hemispheric compensatory routing within the sensorimotor network. The negative correlation between UE-FMA and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d{EC}_{cPSC\to\:iPMC}\:$$\end{document}suggested that when ipsilesional somatosensory processing was compromised, survivors rely more on contralesional sensory-driven pathways for action selection and motor planning, a pattern consistent with previous evidence [61]. Information flow between ROIs may serve as a speculative indicator of upper-limb motor function, suggesting that cortical network metrics are associated with motor function status. Although standardized clinical scales [62] are widely used to assess post-stroke motor function, such scales have certain limitations [63]. These include inconsistent severity estimation, variable predictive accuracy across demographics, and limited sensitivity to functional neuroplasticity. Multimodal biomarker assessments combining structural and functional measures [64] produce promising results but are associated with clinical implementation challenges. Our findings provide preliminary evidence toward identifying precise, clinically feasible indicators associated with upper-limb motor function following stroke. Multimodal somatosensory inputs facilitating cortical sensitivity Compared with unimodal stimulation, synergistic stimulation with MS and NMES in Task 3 evoked significantly stronger cortical activation (Fig. 6A) and interregional connectivity (Fig. 6B), indicating that integrated multimodal somatosensory inputs were associated with supra-additive cortical facilitation. These neurophysiological findings confirmed that concurrent multimodal somatosensory stimulation induces the coactivation of motor and somatosensory cortical areas [65, 66], suggesting a fundamental mechanism of cross-modal integration in sensorimotor processing. Our neurophysiological findings can be interpreted through Friston’s Cortical Response Theory [67], which provides a unified framework for understanding neural dynamics. According to this predictive coding framework, the amplified cortical activation observed during multimodal stimulation in Task 3 reflects an adaptive strategy in the brain involving minimization of free energy by increasing synaptic efficacy in sensorimotor pathways and optimizing metabolic resource allocation under heightened perceptual uncertainty. This theoretical perspective indicates that combined pneumatic and NMES somatosensory stimulation may facilitate cortical reorganization through two complementary mechanisms: enhancing the precision weighting of somatosensory prediction errors, and second promoting the formation of more efficient hierarchical representations in the sensorimotor system. In the present study, chronic stroke survivors had significantly diminished cortical sensitivity to somatosensory stimulations (Fig. 6), this finding is consistent with those of research documenting impaired afferent processing in this population [68]. Although basic cutaneous tactile perception may remain intact post-stroke, neurophysiological data indicates the process of substantial deficits in proprioceptive and sensorimotor integration capacities. These functional impairments likely originate from the disruption of thalamocortical and corticocortical pathways following stroke, which compromises the synaptic connections between somatosensory and motor networks [69]. In this process, cortical desensitization may be attributable to three interrelated pathological reasons: structural disconnection in sensorimotor loops due to ischemic damage [70], functional decoupling of sensory and motor cortical areas [71], and reduced neural resource allocation for multi-somatosensory processing [72]. As presented in Fig. 5C, in the current study, the primary somatosensory cortex demonstrated robust activation in response to NMES, a finding consistent with the neurophysiological observations reported in previous study [73]. The presence of stimulation-dependent PSC recruitment suggests that NMES-induced afferent volleys effectively engage the lemniscal sensory pathway, and that the intensity of stimulation correlates with the degree of sensorimotor network activation. As a somatosensory stimulation method, NMES evokes multimodal proprioception through cutaneous mechanoreceptors (skin stimulation), muscle spindle Ia/II afferents (muscle depolarization), and Golgi tendon organ Ib afferents (evoked contractions). These afferent signals demonstrate privileged access to sensorimotor circuits through direct thalamocortical projections to the PSC and transcortical pathways linking the PSC with primary motor cortex [74]. Such privileged neuroanatomical specificity explains why NMES produces more pronounced cortical effects compared with cutaneous stimulation alone, particularly in the context of sensorimotor rehabilitation after neurological injury. Limitations This exploratory study has several limitations. The limited sample size and heterogeneity of the study population, particularly the varying severity of motor impairment among stroke participants, may limit the generalizability of the findings. Methodological constraints should also be considered. GCA was applied directly to hemodynamic signals without HRF deconvolution, regional HRF variability may have affected temporal lags and potentially biased the inferred dominant of information flow, limiting interpretation of the findings as direct neural causality. Future studies incorporating HRF deconvolution, region-specific HRF modeling, or simultaneous EEG–fNIRS recordings may help disentangle neural and neurovascular contributions and provide a more comprehensive characterization of effective connectivity. Furthermore, the stroke survivors included in this study were in the chronic stage, limiting our ability to draw insights related to the potential of rehabilitation during the subacute period after stroke. Large-scale studies are needed to further clarify how neuromodulatory effects vary across post-stroke recovery phases and levels of motor dysfunction severity. Conclusion The current work indicated that the combined mechano-electrical somatosensory stimulation activated broader cortical areas than either sole stimulation in both the stroke group and the unimpaired group. The stroke group exhibited enhanced ipsilesional activation in response to NMES compared to the controls, whereas MS elicited comparable ipsilesional activation patterns between the two groups. The involvement of VME evoked distinctive activation in the prefrontal and sensorimotor networks in the stroke group compared to the passive tasks without VME. In particular, the combination of MS and NMES co-stimulation with VME induced more widespread cortical activation than either modality alone. Notably, this combined approach also attenuated intergroup differences in cortical network organization in contrast to the control group, resulting in connectivity patterns post-stroke resembled loser to those observed in unimpaired persons. Supplementary Information Supplementary Material 1. Supplementary Material 1.