Abstract

Introduction:

Robot-assisted training (RAT) exhibits inconsistent efficacy in post-stroke upper limb rehabilitation, with its underlying neural mechanisms remaining unclear. This study aimed to investigate how different therapy modes modulate cerebral activation strategies in distinct subgroups of stroke patients.

Methods:

We utilized functional near-infrared spectroscopy (fNIRS) to investigate differences in cortical activation strategies, specifically the response sensitivity to various robotic training modes, by stratifying forty-one patients based on functional level, disease chronicity, and hemiplegic side. In a single session, each participant underwent four RAT modes (Passive, Assistive-active, Active, and Mirror) while a 48-channel fNIRS system monitored cortical activation.

Results:

Our results revealed no statistically significant differences in global mean activation intensity between any of the subgroups (p > 0.05). Instead, the core finding was a clear dichotomy in neural strategy: low-function, subacute, and left-hemiplegia groups were highly “mode-sensitive,” exhibiting significant changes in brain activation across different training modes (e.g., Active vs. Mirror, p < 0.05). Conversely, high-function, chronic, and right-hemiplegia groups were “mode-consolidated,” demonstrating a stable activation pattern with almost no significant differences among the active modes.

Discussion:

We conclude that the core neural mechanism of post-stroke recovery is characterized not by simple changes in activation intensity, but by a strategic evolution from a flexible, cue-dependent “mode-sensitive” state to a more automated “mode-consolidated” state. This distinction provides a robust neurophysiological rationale for personalizing rehabilitation, enabling clinicians to strategically match therapeutic stimuli to a patient's specific neural profile—applying diversified training to “mode-sensitive” patients and high-load challenges to “mode-consolidated” patients—to break through rehabilitation plateaus.

1 Introduction

Stroke remains a leading cause of long-term disability worldwide, with approximately 80% of survivors experiencing some degree of functional impairment (Feigin et al., 2025; Niu et al., 2025). Beyond the physical toll, the prolonged dependency of survivors exerts a profound impact on modern healthcare systems, necessitating resource-intensive long-term care. Pathophysiologically, the cerebral injury triggers a complex cascade of events, including neuroinflammation and oxidative stress, which collectively drive neuronal cell death and functional deficits (Purrahman et al., 2023). While these molecular mechanisms underpin the initial tissue damage, the restoration of motor function depends heavily on the brain's intrinsic capacity for network reorganization and plasticity (Stinear et al., 2020). Among these sequelae, upper limb hemiparesis is particularly pervasive and debilitating, profoundly compromising an individual's capacity to perform essential activities of daily living (ADL) and consequently diminishing their independence and quality of life. The prolonged and often incomplete recovery of motor function not only imposes a significant psychological and physiological burden on patients but also creates substantial economic and caregiving challenges for their families and society (Langhorne et al., 2011). Therefore, optimizing therapeutic strategies to enhance upper limb motor recovery is a paramount objective in the field of neurorehabilitation.

The foundation of motor rehabilitation is built upon a spectrum of training modes, each with a distinct theoretical basis and clinical application. These range from Passive movement, which relies on an external force to maintain joint mobility and provide sensory input, to Active movement, which requires the patient's own volitional effort to drive motor commands and engage cortical networks. Intermediary modes like Assistive-Active training bridge this gap, pairing a patient's intent with external support, while other approaches like Mirror Therapy utilize visual illusion to modulate cortical activity. The principle of neuroplasticity—the brain's intrinsic capacity for reorganization—is the common target of all these modes (Wu et al., 2021). It is now well-established that high-dosage, high-repetition, and task-oriented training are essential to effectively drive this plasticity and promote functional recovery (Kwakkel et al., 2008).

However, rigorously comparing the neurophysiological effects of these different modes in a clinical setting presents a significant methodological challenge. Conventional manual therapy, while crucial, is inherently variable and subject to therapist fatigue, making it difficult to deliver the standardized, high-intensity dosage required for robust scientific inquiry. It is in this context that robotic platforms serve as indispensable scientific tools (Park et al., 2020). By providing precisely controlled, repeatable, and quantifiable delivery of various training modes—from purely passive motion to active resisted tasks—robotic systems allow us to isolate the variable of interest: the training mode itself (Aguirre-Ollinger et al., 2024). This enables a systematic investigation into how the brain responds to these distinct sensorimotor inputs, an endeavor that would be unfeasible with conventional methods alone.

Despite the standardization offered by such platforms, clinical outcomes remain highly variable (Rodgers et al., 2019; Xie et al., 2022). This inconsistency strongly suggests that the central issue is not the delivery system, but rather a “one-size-fits-all” application of these training modes to a profoundly heterogeneous stroke population. Patients differ vastly in their functional status, stage of recovery (chronicity), and the neuroanatomical characteristics of their lesion (Mehrholz et al., 2020). It is highly probable that a training mode that is beneficial for a patient in the subacute phase may be ineffective or even counterproductive for a patient in the chronic phase. This highlights an urgent need to move beyond monolithic treatment paradigms and toward a precision rehabilitation framework, where therapeutic strategies are tailored to specific, neurobiologically-defined patient subgroups (Xie et al., 2022).

To build such a framework, we must first understand the distinct neural processes that different training modes induce in different types of patients. Functional near-infrared spectroscopy (fNIRS) is an ideal neuroimaging modality for this purpose. As a non-invasive, portable, and motion-tolerant technology (Scholkmann et al., 2014), fNIRS allows for the real-time monitoring of cortical hemodynamics during the performance of functional tasks in ecologically valid settings (Rahman et al., 2020). This provides a direct window into the brain's strategic response to therapy as it unfolds, making it perfectly suited for integration with robotic rehabilitation paradigms.

This study proposes that the critical difference among patient subgroups lies not in the overall magnitude of brain activation, but in the nature of their underlying activation strategy. We introduce the concept of “mode sensitivity” to define this strategic difference, characterizing it as the degree to which an individual's cortical activation patterns are modulated by changes in the therapeutic training mode. This conceptual framework is grounded in established principles of post-stroke neural recovery and hemispheric specialization, leading to two primary hypotheses.

First, based on the dynamic evolution of neural recovery—from a diffuse, bilateral state to a more focal, efficient network (Feydy et al., 2002; Takeda et al., 2007)—we hypothesize that patients in phases of higher neuroplasticity (i.e., the subacute stage and those with lower functional levels) will exhibit greater “mode sensitivity.” Their brains are in an active state of exploration, highly dependent on external cues to guide network reorganization. Second, drawing from the principles of hemispheric specialization (Zemke et al., 2003; Riecker et al., 2010), we hypothesize that patients with right hemisphere lesions (left hemiplegia) will demonstrate higher “mode sensitivity,” reflecting a more extensive, strategic exploratory process orchestrated by the intact, dominant left hemisphere.

Therefore, the primary objective of this study is to utilize a robotic platform as a tool for standardized delivery to systematically investigate how cortical activation strategies are modulated by distinct training modes (Passive, Assistive-Active, Active, and Mirror) in subgroups of stroke patients stratified by hemiplegic side, functional level, and disease chronicity. By testing the concept of “mode sensitivity,” we aim to uncover fundamental neural signatures of the recovery process, thereby laying the neuroscientific groundwork for a more precise and effective paradigm of personalized rehabilitation.

2 Materials and methods2.1 Participants

A total of 41 individuals with post-stroke upper limb motor dysfunction were recruited for this study. Inclusion criteria were: (1) a first-ever unilateral stroke confirmed by computed tomography (CT) or magnetic resonance imaging (MRI); (2) an upper limb functional level rated between 3 and 6 on the Functional Test for the Hemiplegic Upper Extremity, Hong Kong version (FTHUE-HK); (3) a Mini-Mental State Examination (MMSE) score of ≥20, ensuring the ability to understand and follow experimental instructions; and (4) right-handedness as determined prior to the stroke event. Exclusion criteria included a history of other neurological or psychiatric disorders, severe cognitive or aphasic impairments, and any contraindications for fNIRS measurement, such as cranial defects or significant head skin lesions.

Participants were stratified into three pairs of subgroups for analysis: (1) hemiplegic side: left hemiplegia (n = 21) vs. right hemiplegia (n = 20); (2) disease chronicity: subacute phase (≤6 months post-stroke, n = 28) vs. chronic phase (>6 months post-stroke, n = 13); and (3) functional level: high-function (FTHUE-HK score 5–6, n = 12) vs. low-function (FTHUE-HK score 3–4, n = 29). The study protocol was approved by the Ethics Committee of the First Affiliated Hospital of Nanhua University (Approval No. 2024KS-KF-28-02) and was registered with the Chinese Clinical Trial Registry (ChiCTR2500096992). All participants provided written informed consent prior to their inclusion in the study. Flowchart of the experiment was elaborated in Figure 1.

Flowchart of the experiment.

2.2 Rehabilitation robot and training modes

The study utilized the Wisebot-X5, a three-dimensional (3D) upper limb exoskeleton rehabilitation robot (Shenzhen Huaquejing Medical Technology Co., Ltd., China). This device is designed to enable multi-joint arm movements within a gamified 3D workspace, simulating functional tasks. The robot was configured to deliver four distinct training modes, each with precise operational parameters. The general range of motion for the tasks included shoulder horizontal abduction up to 90° and adduction to 40°, with shoulder forward flexion from 45 to 130°.

The four training modes were defined as follows: (1) passive mode: the robot fully drove the participant's affected limb through a predefined trajectory at a constant angular velocity of 15°/s, requiring the participant to remain relaxed. (2) Assistive-active mode: the participant initiated the movement, and the robot provided an assistive torque of 2 N·m only when the participant's voluntary effort was insufficient to complete the task. (3) Active mode: the participant volitionally drove the robotic arm to complete the task against a constant resistance of 1 kg. (4) Mirror mode: a body-sensing camera captured the real-time motion of the participant's unaffected limb, and the robot drove the affected limb to perform a symmetrical, mirrored movement at a maximum velocity of 60°/s.

2.3 Experimental design and procedure

This study employed a cross-sectional, repeated-measures design in which each participant underwent all four robotic training modes within a single experimental session (Figure 2). The presentation order of the four modes was pseudo-randomized across participants to mitigate potential order effects. For each training mode, a classic block design paradigm was implemented. This paradigm consisted of five repeated cycles, with each cycle comprising a 30-s task execution period followed by a 30-s resting period. Prior to the first task block, a 30-s baseline recording was acquired to serve as a reference for subsequent activation analysis.

Experimental design.

To ensure anatomical accuracy for group-level analysis, the 3D coordinates of each source and detector were recorded for every participant using a 3D digitizer. These coordinates were subsequently transformed into the Montreal Neurological Institute (MNI) standard coordinate space, allowing for precise anatomical localization of the measured cortical activity. Signal quality was visually inspected prior to each recording to ensure good optode-scalp contact.

2.4 fNIRS data acquisition

Cortical hemodynamic responses were continuously monitored using a 48-channel continuous-wave fNIRS system (NirSmart, DanYang HuiChuang Medical Equipment Co., Ltd., China). The system was configured with 21 light sources and 16 detectors, utilizing two wavelengths of near-infrared light (730 and 850 nm) at a sampling rate of 10 Hz. The optodes were arranged with a standard source-detector separation of 30 mm and were securely mounted in a specialized cap positioned according to the international 10–20 system. The cap placement ensured comprehensive coverage of bilateral motor-related cortices (Table 1), including the primary motor cortex (M1), premotor cortex (PMC), supplementary motor area (SMA), primary somatosensory cortex (PSC), dorsolateral prefrontal cortex (DLPFC), frontopolar area (FP), frontal eye fields (FEF), and somatosensory association cortex (SAC). The arrangement of the channels and the detailed annotation of the brain's regions of interest (ROI) based on Brodmann area were elaborated in Figure 3.

ROIChannelFunctionLeft M1CH29, CH31Motor execution: generates the final motor commands for direct control of fine movements of the contralateral side of the bodyRight M1CH20, CH21Left SMACH30, CH32Motor planning and sequencing: internally plans, prepares, and sequences complex movements, especially crucial for bimanual coordinationRight SMACH19, CH22Left PMCCH10, CH46, CH48Movement preparation and guidance: prepares for movements guided by external (e.g., visual, auditory) cues and transforms sensory information into motor commandsRight PMCCH34, CH35, CH38Left PSCCH33, CH36, CH37Primary sensory processing: receives and processes basic sensory information (touch, pressure, temperature, proprioception) from the contralateral (right) side of the bodyRight PSCCH9, CH45, CH47Left FPCH6, CH7, CH16High-Level Cognitive Planning: Responsible for the highest level of goal setting, long-term planning, decision-making, and exploratory behavior, indirectly influencing motor strategy selectionRight FPCH4, CH5, CH15Left DLPFCCH8, CH17, CH18, CH25, CH26Working memory and executive function: maintains and manipulates relevant information (e.g., task rules) during movement, monitors performance, and inhibits inappropriate actionsRight DLPFCCH3, CH13, CH14, CH23, CH24Left FEFCH27, CH28Eye-hand coordination and attentional guidance: controls voluntary eye movements to locate visual targets and directs spatial attention toward the goal of the movementRight FEFCH39, CH40Left SACCH11, CH12Higher-order sensory integration: integrates various sensory inputs from the PSC to form a detailed perception of objects and one's own limb positionRight SACCH41, CH42

Detailed annotation of ROI based on Brodmann area and corresponding functions.

Arrangement of the channels and distribution of ROI.

2.5 fNIRS data analysis

All fNIRS data were processed and analyzed using the NirSpark software (v1.8.9) (DanYang HuiChuang Medical Equipment Co., Ltd., Danyang, China). (Chu et al., 2023). The preprocessing pipeline began with the identification and exclusion of channels with poor signal quality, defined by a coefficient of variation greater than 20%. The raw light intensity data from the remaining channels were then converted into changes in optical density. Subsequently, a band-pass filter with a cutoff range of 0.01–0.2 Hz was applied to remove physiological noise, such as cardiac pulsations and respiratory signals, as well as low-frequency signal drift. Motion artifacts were detected and corrected using a spline interpolation algorithm.

Finally, the preprocessed optical density data were converted into concentration changes of oxyhemoglobin (ΔHbO) and deoxyhemoglobin using the modified Beer-Lambert Law (MBLL). For all subsequent statistical analyses, we focused on ΔHbO signals, as they are considered to have a higher signal-to-noise ratio and a stronger correlation with task-related neural activity.

2.6 Statistical analysis

All statistical analyses were conducted using Python, with the significance level set at α = 0.05. The Shapiro-Wilk test was first applied to all continuous variables to assess for normality, guiding the selection of either parametric or non-parametric tests. To determine task-related cortical activation, the mean ΔHbO values during task blocks were compared against the baseline using one-sample t-tests (for normally distributed data) or Wilcoxon signed-rank tests (for non-normally distributed data) for each ROI. The resulting p-values were corrected for multiple comparisons using the Benjamini–Hochberg false discovery rate (FDR) procedure.

To compare overall activation levels between the stratified subgroups (e.g., high- vs. low-function), independent samples t-tests or Mann–Whitney U tests were employed. Analysis of covariance (ANCOVA) was used to control for the effects of confounding baseline variables, such as age, where significant differences between groups were identified. To assess “mode sensitivity” within each subgroup, a repeated measures ANOVA or a non-parametric Friedman test was conducted to identify significant differences in activation across the four training modes. Significant main effects were followed by post-hoc paired t-tests or Wilcoxon signed-rank tests with FDR correction. To explore the relationship between neural strategies and clinical outcomes, Spearman's rank correlation analysis was used to test the association between a quantified “Mode Sensitivity Index” (defined as the global mean activation difference between the Active and Passive modes) and both the FTHUE-HK score and the time since stroke.

3 Results3.1 Participant characteristics and baseline demographics

The demographic and clinical characteristics of the 41 participants are presented in Table 2. As per the stratification criteria, the chronic group had a significantly longer time since stroke compared to the subacute group (p < 0.001). Furthermore, the high-function group was significantly older than the low-function group (p = 0.016), and the age difference between the subacute and chronic groups approached statistical significance (p = 0.056). An analysis of covariance (ANCOVA), controlling for age, confirmed that there were no significant main effects of functional level (F(1, 38) = 1.563, p = 0.219) or disease chronicity (F(1, 38) = 1.743, p = 0.195) on global mean activation during the active training mode. This finding indicates that the observed differences between subgroups were not attributable to overall neural effort but rather to the underlying activation strategies.

CharacteristicAll patients (N = 41)Left hemiplegia (n = 21)Right hemiplegia (n = 20)p-valueaHigh-function (n = 12)Low-function (n = 29)p-valuebSubacute phase (n = 28)Chronic phase (n = 13)p-valuecAge (years), Mean ± SD53.8 ± 12.153.4 ± 10.554.1 ± 13.80.86360.2 ± 9.651.1 ± 12.20.01651.6 ± 13.058.4 ± 8.60.056Time since stroke (days), Mean ± SD162.4 ± 164.1168.4 ± 169.8156.4 ± 162.40.409172.6 ± 203.4158.1 ± 148.30.28183.5 ± 35.8346.6 ± 198.80.000FTHUE-HK, Mean ± SD4.0 ± 1.13.8 ± 1.04.3 ± 1.20.2415.5 ± 0.53.4 ± 0.50.0004.0 ± 1.14.0 ± 1.20.879FMA-UE, Mean ± SD26.6 ± 8.924.4 ± 7.228.8 ± 10.10.13935.9 ± 7.623.1 ± 6.60.00028.0 ± 9.223.2 ± 7.50.108Gender, n(%)Male26 (63.4)13 (61.9)13 (65.0)1.0009 (75.0)17 (58.6)0.48016 (57.1)10 (76.9)0.305Female15 (36.6)8 (38.1)7 (35.0)3 (25.0)12 (41.4)12 (42.9)3 (23.1)Stroke type, n(%)Hemorrhagic19 (48.7)9 (45.0)10 (52.6)0.7526 (54.5)13 (46.4)0.73115 (57.7)4 (30.8)0.176Ischemic20 (51.3)11 (55.0)9 (47.4)5 (45.5)15 (53.6)11 (42.3)9 (69.2)Not specified0 (0.0)0 (0.0)0 (0.0)0 (0.0)0 (0.0)0 (0.0)0 (0.0)Phase, n(%)Subacute28 (68.3)14 (66.7)14 (70.0)1.0008 (66.7)20 (69.0)1.00028 (100.0)0 (0.0)–Chronic13 (31.7)7 (33.3)6 (30.0)4 (33.3)9 (31.0)0 (0.0)13 (100.0)Function level, n(%)High-function12 (29.3)4 (19.0)8 (40.0)0.18112 (100.0)0 (0.0)–8 (28.6)4 (30.8)1.000Low-function29 (70.7)17 (81.0)12 (60.0)0 (0.0)29 (100.0)20 (71.4)9 (69.2)

Baseline demographic and clinical characteristics of participants.

ap-value for comparison between Left hemiplegia (n = 21) and Right hemiplegia (n = 20).

bp-value for comparison between High-function (n = 12) and Low-function (n = 29).

cp-value for comparison between Subacute phase (n = 28) and Chronic phase (n = 13).

3.2 Influence of hemiplegic side on brain activation strategies

The two hemiplegia groups exhibited distinct activation characteristics during the different training modes (Figures 4B, C). For the left hemiplegia group (right hemisphere lesion), the passive mode did not elicit significant activation in any of the measured ROIs (p > 0.05). In contrast, both the assistive-active and active modes induced robust and widespread significant activation across all ROIs, including M1, PMC, and SMA (p < 0.05). The mirror mode also resulted in significant activation in most ROIs, with the exception of the left FP and FEF. For the right hemiplegia group (left hemisphere lesion), the passive mode was similarly non-activating (p > 0.05). The assistive-active and active modes significantly activated all ROIs (p < 0.05), though the active mode showed a slightly more restricted pattern, failing to activate the left FP and SMA. The mirror mode in this group induced a much more localized activation pattern, significant only in the left M1, left FEF, right SMA, and bilateral PSC and SAC (p < 0.05).

(A) Average activation values of brain regions for each training mode in the left/right hemiplegia group; (B) brain activation topographic maps for each RAT training mode in left/right hemiplegic patients; (C) activation status of brain regions for each RAT training mode in left/right hemiplegic patients; (D) comparison of brain region activation differences among each mode in left hemiplegia patients; (E) comparison of brain region activation differences among each mode in right hemiplegia patients.*p < 0.05; **p < 0.01; ***p < 0.001.

To assess for differences in overall neural effort, we compared the global mean activation between the left and right hemiplegia groups specifically during the active mode. An independent samples t-test revealed no statistically significant difference in activation magnitude between the two groups (p > 0.05, Supplementary Table 2a) (Figure 4A). Furthermore, we tested the classic hypothesis of a contralesional activation bias. Our intra-hemispheric analysis in both the left and right hemiplegia groups found no significant asymmetry in activation between the unaffected and affected hemispheres, either at the whole-hemisphere level or within corresponding ROIs (p > 0.05, Supplementary Tables 2b, c). This null finding suggests that a simple contralesional dominance model does not adequately describe the activation patterns in this heterogeneous sample.

The primary distinction between the groups was revealed in their response to changing training modes. The left hemiplegia group demonstrated a highly “mode-sensitive” profile. As shown in the heatmap in Figure 4D, there were significant and widespread activation differences between the passive mode and all three active-participation modes across nearly all ROIs. Crucially, significant differences were also observed among the active modes (e.g., Active vs. Assistive-Active; Mirror vs. Active), indicating a dynamic neural strategy that was finely tuned to the specific demands of each mode. Conversely, the right hemiplegia group exhibited a “mode-consolidated” strategy. Although their active modes were significantly different from the passive mode, the number of brain regions showing these differences was considerably smaller (Figure 4E). More importantly, there were almost no significant activation differences among the three active-participation modes. This suggests a more fixed, less adaptable neural activation pattern.

3.3 Influence of functional level on activation strategies

Activation patterns also varied markedly with functional level (Figure 5B). In the low-function group, the passive mode was non-activating (p > 0.05). In contrast, the assistive-active, active, and mirror modes all induced widespread, significant activation across nearly all ROIs (p < 0.05). In the high-function group, the passive mode was also non-activating. However, the active-participation modes recruited a more focused neural network; specifically, the assistive-active mode failed to activate the ipsilesional FP, FEF, and contralateral SMA, the active mode failed to activate the ipsilesional FP and M1, and the mirror mode failed to activate the ipsilesional FEF, DLPFC, and SMA (all p > 0.05).

(A) Average activation values of brain regions for each training mode in the low/high function groups; (B) activation of brain regions in each RAT training mode for patients in the low/high function groups; (C) comparison of brain region activation differences among each mode in the low-function group; (D) comparison of brain region activation differences among each mode in the high-function group. *p < 0.05; **p < 0.01; ***p < 0.001.

Consistent with the findings from the hemiplegic side analysis, there was no statistically significant difference in global mean activation during the active mode between the high- and low-function groups (p > 0.05, Supplementary Table 3a) (Figure 5A). Similarly, intra-hemispheric analyses within both the high- and low-function groups revealed no significant activation asymmetry between the ipsilesional and contralesional hemispheres (p > 0.05, Supplementary Tables 3b, c). These null findings further support that the key distinction between these groups is strategic rather than quantitative.

The core difference was again found in mode sensitivity. The low-function group was highly mode-sensitive, showing significant activation differences between the passive mode and all active modes across a vast array of ROIs (Figure 5C). Significant distinctions were also present among the active modes, particularly between the active mode and the other two (mirror and assistive-active), evident in key motor ROIs such as bilateral M1, DLPFC, and SAC. In striking contrast, the high-function group was “mode-insensitive.” As depicted in Figure 5D, there were no statistically significant differences in cortical activation among any of the three active-participation modes, indicating a stable and automated neural strategy.

3.4 Influence of disease chronicity on activation strategies

Stratification by disease chronicity revealed a similar trend (Figure 6B). In the subacute group, the passive mode was non-activating, whereas all three active-participation modes elicited robust and significant activation across almost all measured ROIs (p < 0.05). In the chronic group, a more efficient activation pattern was observed. While the passive mode remained non-activating, the active modes required the engagement of a smaller set of brain regions. Specifically, during the assistive-active and active modes, the ipsilesional M1, contralateral FP, and SMA were not significantly activated, and during the mirror mode, only bilateral SAC showed significant activation (all p > 0.05).