Efficacy of a soft robotic exoskeleton for upper-limb rehabilitation in subacute stroke inpatient rehabilitation: a randomised controlled trial Authors: Kai Wen Elvina Tay, Pui Kit Tam, Qi Xie, Hwa Sen Lai, Wai Kin Yuen, Xiao Min Lilian Lim, Shu Ling Ng, Max Lim, Lee Mei Ong, Weiying Amelia Tan, Syamil Rosli, Jeong Hoon Lim Published: 2026 ABSTRACT AbstractBackgroundSoft robotic devices may be a useful adjunct for stroke rehabilitation by providing higher intensity training, while allowing comfort and portability compared to rigid exoskeletons. We evaluated the efficacy of a sensor-assisted soft robotic glove (SSR) during inpatient stroke rehabilitation.MethodsThis assessor-blinded, randomised controlled trial was conducted in 2 inpatient rehabilitation units. Adults aged 21–90 years old, with ischaemic or haemorrhagic stroke and unilateral upper-limb impairment were recruited between July 2023 and May 2025. Participants were randomised to either [1] SSR or [2] supervised Graded Repetitive Arm Supplementary Programme (GRASP) training for up to 15 sessions, 30–45 min/session; 5 days/week for up to 3 weeks. The primary outcome was Fugl-Meyer Assessment of Upper Extremity (FMA-UE). Secondary outcomes included Stroke Upper Limb Capacity Scale, the DuruÖz Hand Index, and Stroke Impact Scale-16 (SIS-16). Outcomes were assessed at baseline (T1), post-intervention (T2) and follow-up at 3 months (T3).ResultsRecruitment ceased before the planned sample size was reached due to funding constraints. Therefore, the trial was underpowered to detect the prespecified between-group difference. Of 323 screened individuals, 61 were randomised. 3 participants did not complete T2 assessment resulting in a modified intention-to-treat population of 58 (SSR n = 28, control n = 30). Both groups showed improvement across primary and secondary outcomes, with no statistically significant differences observed between them. In a post-hoc exploratory subgroup analysis with baseline FMA-UE < 35 and ≥ 10 sessions (n = 38), a statistically significant difference in FMA-UE distal sub-score at T2 was observed, favouring the SSR group (mean difference 3.06, 95% confidence interval [CI] 0.02 to 6.09, p = 0.049) but not at T3. A numerically higher proportion of participants in the SSR group achieved minimal clinically important difference value for SIS-16 (77.8% vs. 55.0%, adjusted odds ratio 6.52 (95% CI 1.00 to 42.58, absolute risk difference 22.8% [95% CI -7.2% to 47.4%], p = 0.050) at T2, but did not reach statistical significance.ConclusionSensor-assisted soft robotic glove training did not demonstrate superiority over supervised GRASP for upper limb function post-stroke. Exploratory subgroup analyses identified a signal suggesting improved distal hand motor recovery in participants with greater baseline impairment with sufficient therapy exposure.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12984-026-02133-0.Keywords: Upper-limb rehabilitation, Stroke, Soft robotic glove, Stroke rehabilitation, Fugl-Meyer Assessment, Randomised controlled trial Introduction Stroke is a major public healthcare problem worldwide and is a significant cause of morbidity and mortality. It is the third leading cause of death and disability globally with over 160 million disability-adjusted life-years lost annually [1]. The global prevalence of stroke has been increasing, and it has been estimated that approximately 85% of stroke survivors suffer from upper-limb impairment [2]. Persistent upper-limb impairments can significantly affect activities of daily living and negatively impact quality of life in stroke survivors [3]. Neurorehabilitation stroke principles for motor recovery emphasize on repetitive, high intensity, task specific practice to promote neuroplasticity [4]. The subacute stroke phase represents a critical window for recovery, during which targeted interventions can lead to greater functional gains [5]. However, delivering rehabilitation at required intensity and task specificity remains challenging in clinical settings. These constraints are driven largely by limited manpower resources and limited clinical time. Robotic-assisted technology has emerged as a strategy to augment therapy intensity [6]. While upper-limb rigid robotics are more common and more established in clinical and research settings, soft robotic devices are an emerging approach due to their comfort, safety, portability and potential for home-based use [7]. Rigid upper-limb exoskeleton devices are typically driven by rigid linkages or joints, which may force participants into unnatural movement patterns, leading to discomfort [8]. Additionally, the rigid exoskeleton devices are typically huge, bulky and difficult to carry, not suitable for bedside use. There are stroke studies on soft robotic gloves [9–11] with earlier designs providing passive motion only [9], including recent randomised controlled trials providing evidence to support the use of robotic gloves for hand rehabilitation in stroke participants. Lee et al. studied robot-assisted glove training in individuals with subacute to chronic stroke and reported improvements in upper-limb motor impairment and activities of daily living compared with conventional therapy [12]. However, in this study, the authors employed a crossover design without long-term follow-up and enrolled mainly chronic participants and did not assess health-related quality of life (HrQOL). Other randomised controlled trials to date demonstrated short-term improvements in hand motor outcomes compared with therapist-assisted training immediately post-intervention. Longer term outcomes and quality of life measures were not included [12–14]. This randomised controlled trial (RCT) was designed to evaluate the efficacy of a sensor-assisted soft robotic (SSR) glove with passive and active-assisted training modes in upper-limb function and HrQOL in a subacute inpatient stroke population, with outcomes assessed at post-intervention and follow-up. We hypothesized that intervention with this robotic device would enhance upper-limb recovery and functional outcomes, resulting in clinically meaningful improvement in upper-limb motor recovery, upper-limb activity and health-related quality of life compared with the active control intervention. Materials and methods Study design and setting We conducted an assessor-blinded, parallel-group randomised controlled trial with a 1:1 allocation ratio designed to assess superiority. The study was conducted in inpatient rehabilitation units of Alexandra Hospital and Jurong Community Hospital in Singapore. Ethics statement The study obtained institutional ethical approval from the National Healthcare Group- Domain Specific Review Boards (DSRB reference number: 2020/01105). All participants provided written informed consent prior to enrolment. All research was conducted in compliance with the WHO Declaration of Helsinki. Participants Participants were recruited from inpatient rehabilitation units of the two participating hospitals between July 2023 and May 2025. Participants were recruited during the subacute phase of stroke recovery, defined as within 6 months of stroke onset [15]. Eligible participants were adults aged 21–90 years, with a diagnosis of haemorrhagic or ischaemic stroke, who were medically stable and presented with unilateral upper-limb impairment. Additional inclusion criteria include participants with Fugl-Meyer Assessment of Upper Extremity (FMA-UE) [16] between 8 and 47 out of 66, or between 48 and 56 provided that the hand and wrist subscale was not at ceiling. Participants were required to have the ability to provide informed consent, maintain upright sitting balance for at least 45 min and be able to comprehend and follow commands. Exclusion criteria included recurrent stroke, significant cognitive or communicative impairments, history of severe depression or active psychiatric disorders or unstable medical conditions (such as active infection, recent myocardial infarction, or uncontrolled blood pressure) or an anticipated life expectancy of less than one year, poor skin integrity over the affected hand area (e.g. open wounds), finger amputation at the affected hand, known allergies to adhesive materials or severe spasticity of affected distal upper-limb with Modified Ashworth Scale [17] greater than 3. Randomisation and blinding Randomisation was performed with a computer-generated random sequence using Sealed Envelope (www.sealedenvelope.com). Participants were randomly assigned in a 1:1 ratio to either the experimental or control group using permuted block randomisation with variable block sizes of 2, 4 and 6 balanced by centre with no additional stratification. Allocation concealment was ensured by having the randomisation sequence maintained by an independent research coordinator who was not involved in screening, assessment or therapy delivery. Outcome assessors were blinded to group allocation throughout the study. Due to the nature of the study intervention, therapist and participant blinding was not feasible. Outcome assessments were conducted by the same assessor who had received training in the outcome measures previously. Intervention Interventions The experimental group underwent training with SSR while the control group followed Graded Repetitive Arm Supplementary Programme (GRASP) protocol [18]. The EsoGLOVE™ (Roceso Technologies, Singapore) is a soft robotic assistive glove designed to provide hand rehabilitation through both continuous passive motion and active-assisted modes, accompanied by real-time 3D representation on screen and functional tasks training with real objects. Figure 1 shows the sensor-assisted soft robotic glove and representative training setup used during the intervention sessions. The simultaneous visual effects and proprioceptive feedback (feeling and seeing hand and arm movements) are associated to stimulate the participant’s neuroplasticity. Setup of sensor-assisted soft robotic glove. Participants performing actions such as a grasping b pinching and c release following instructions from the screen The EsoGLOVE™ system consists of a control box, which requires a power input of 36 W (12 V, 3 A). The air pump inside the control box converts the electrical energy to mechanical energy by pressurizing the pneumatic actuators embedded within the textile structure of the glove. The pneumatic actuators assist bi-directional finger and thumb movements, i.e., flexion and extension. The actuators consist of two air channels that inflate upon air pressurization. Upon pressurization of top air channel, the actuators bend, creating a flexion torque to assist finger flexion. Upon pressurization of bottom air channel, the actuators straighten, creating an extension torque to assist finger extension. The amount of the torque is controllable by air pressure within the air channels. In continuous passive motion mode, the glove assists bi-directional finger and thumb movements without participant’s voluntary motor intention, while in active-assisted mode, the glove assists bi-directional finger and thumb movements triggered by participant’s voluntary motor intention (i.e., residual movements of the fingers). Training protocol Participants in the experimental, SSR group were scheduled to receive 10 to 15 sessions of upper-limb rehabilitation sessions using the SSR during their inpatient stay. Training was delivered by the occupational therapy assistants (OTA) five days per week for up to three weeks, with each session lasting 30–45 min. Each session typically comprised a combination of both gross motor cyclic exercises and functional task practice activities. Gross motor exercises commonly included repetitive whole-hand grasp and release, thumb opposition and tripod pinch. Functional task practice was individualised according to the ability of each participant but commonly included activities such as pouring water from a bottle into a cup and grasping an apple from a bowl. The same occupational therapists (OT) who were overseeing participants’ conventional therapy prescribed the device’s training modes, levels and duration, functional task selection for each participant. Training mode was prescribed according to participants’ baseline motor impairment and progression throughout rehabilitation. Participants demonstrating voluntary finger movement progressed to active-assisted modes while participants with minimal voluntary finger movement generally commenced in passive mode. Progression between modes was determined by OT based on clinical assessment, including baseline and interval FMA-UE scores together with observation of voluntary movement during therapy sessions. The OTA administered the session based on the OT’s prescription and provided feedback to the OT following each session on the participant’s performance and on whether the prescribed programme was too easy or too difficult. OTs reviewed participant performance and modified the prescribed training mode, functional task selection and exercise difficulty according to each participant’s clinical progress. Participants in the control group received 30–45 min of GRASP session each day with a neoprene wrist support, in addition to their 1-hour conventional occupational therapy session. The neoprene wrist support was used to provide wrist positioning during GRASP activities and to create a more comparable training setup between groups. The neoprene wrist support did not provide active assistance for finger or hand movement, unlike the SSR intervention. The OT prescribed GRASP activities according to participants’ baseline upper limb impairment using the published GRASP programme. Commonly prescribed GRASP activities included therapeutic exercises, ball and towel exercises, with progression to more complex fine motor tasks at higher GRASP levels. OTA provided supervision to ensure compliance and hands-on facilitation where required and reported participant performance to the OT for subsequent adjustment of exercise prescription. During the study period, participants also received conventional occupational therapy focused on individualised therapy determined by the treating occupational therapist and could address a range of upper-limb, self-care or functional activities. The study intervention represented an additional structured 30–45 minutes programme delivered according to the allocated intervention protocol (SSR or GRASP). Outcome measures The primary outcome was Fugl-Meyer Assessment of the Upper Extremity. FMA-UE scale ranges from 0 to 66 with a higher score indicating better function. It has 4 sub-sections: proximal (maximum score 36), wrist (maximum score 10), hand (maximum 14), and coordination (maximum 6). The wrist and hand sub score is often combined to form a distal sub-score (maximum 24) [19–21]. Secondary outcomes include Stroke Upper Limb Capacity Scale (SULCS) [22], a measure of upper-limb activity limitation, activity DuruÖz Hand Index (DHI) [23], a self-reported questionnaire on hand related activities of daily living (higher score indicates greater disability), and Stroke Impact Scale-16 item version (SIS-16) [24], a measure of HrQOL. All these were measured three times: before intervention at baseline (T1), after intervention (T2) and 3 months post-intervention (T3). T2 was designated as the primary endpoint as it most directly reflected the immediate effects of the allocated intervention, while T3 was included to evaluate the persistence of treatment effects over time. Adverse events were monitored prospectively throughout the study and recorded by therapists and investigators at each session. Sample size and power calculation From our unpublished pilot data between-group difference of 6 points in FMA-UE improvement was anticipated. The standard deviation of change in FMA-UE was assumed to be 11 points, based on published data from an inpatient stroke rehabilitation trial evaluating GRASP [18]. Using a two-sample t-test with a two-sided significance level of 0.05 and 80% power, a sample of 53 participants per group (total n = 106) was required. To account for lost to follow-up and withdrawals, the recruitment target was increased to 130 participants (65 per group). Statistical analysis A modified intention-to-treat (mITT) analysis was performed including all participants with available primary outcome data (FMA-UE) at T2, analysed according to their randomised allocation. For baseline demographics, independent sample t-tests were used for continuous variables and chi-square tests for categorical ones for comparison between SSR and control group. To evaluate between-group difference in outcome measures at different time points we used linear mixed models with baseline score of the corresponding outcome as covariate, and the model included group (SSR vs. Control), Time (T2 and T3), group x time interaction and centre as fixed effects. Unstructured covariance matrix was used to account for the relationship between the time points. Satterthwaite approximation was used to estimate degrees of freedom. Pairwise comparisons were adjusted for multiple comparisons using the Bonferroni correction. Normality of residuals and homoscedasticity were checked to ensure model assumptions were not violated. Missing data were checked using Little’s MCAR (missing completely at random) test. Missing data were handled using restricted maximum likelihood estimation. Effect sizes (Cohen’s d) were calculated at each timepoint using the pooled standard deviation (SD) derived from the standard errors (SE) of the least-square means, SD=(SE×√n ) [25]. We performed a sensitivity analysis including all the randomised participants (n = 61) using the same linear mixed model with baseline observation carried forward for missing post-intervention scores. We also compared proportion of participants who achieved the reported minimal clinically important difference (MCID) in each group. MCID value was available for FMA-UE total (12.4), FMA-UE distal (4.9) and SIS-16 (9.4) [20, 26]. Binary logistic regression was used with group and centre as independent variables and was adjusted with the baseline score concerned. Absolute risk differences (ARD) were calculated from the observed events with the 95% confidence intervals (CI) computed using the Newcombe-Wilson method. Since ARDs are primarily provided to complement the adjusted odds ratios (OR) and aid clinical interpretability, in cases of discrepancy between the ARD confidence interval and the p-value, the regression-derived p-value takes precedence. An exploratory post-hoc subgroup analysis for participants who completed at least 10 intervention sessions and whose baseline FMA-UE total score < 35 was performed using the same linear mixed-effects and logistic regression models as the mITT analyses, except that centre was not included. Given the exploratory nature of the subgroup analyses and the reduced sample size, a more parsimonious modelling approach was adopted to reduce model complexity. Data were analysed by SPSS version 27. Statistical significance was set as p < 0.05. Study registration and funding The study was registered at clinicaltrials.gov (NCT05995665). The study was sponsored by Enterprise Singapore and Roceso Technologies, Singapore. This study is reported in accordance with the CONSORT 2025 guidelines [27]. The completed checklist is provided in Supplementary Table S1. No important changes to methods or outcomes were made after trial commencement. Patient and public involvement No patients or members of the public were involved in the design, conduct, reporting, or dissemination of this research. Results Participant flow and baseline characteristics A total of 323 participants admitted for inpatient stroke rehabilitation were screened. Out of these participants, 61 met eligibility criteria and were randomised to SSR group (n = 31) or the control group (n = 30). Recruitment ceased before the planned sample size was reached due to the end of the funding period. Figure 2 shows participant flow and recruitment. Three participants did not complete the T2 assessment for FMA-UE and were therefore excluded from the mITT analysis due to missing primary outcome data. Two participants, despite meeting clinical inclusion criteria, were unable to achieve adequate fit with the device because of hand size. This represents a device-specific usability limitation. The other participant dropped out due to an intercurrent medical event unrelated to study intervention. This resulted in a final sample of 58 participants for mITT analysis (SSR n = 28, control n = 30). CONSORT flow diagram of participant recruitment and retention. *sensitivity analysis. SSR sensor-assisted soft robotic, T2 after intervention, T3 3-month follow-up, FMA-UE Fugl-Meyer Assessment of Upper Extremity Table 1 shows the baseline demographics and clinical characteristics, and these were comparable between groups, including sex, stroke type, age, time since stroke and baseline upper-limb motor impairment. There were no statistically significant differences between the two groups. Table 1Baseline demographic and clinical characteristics of the study participants in the modified intention-to-treat analysisSSRn = 28Controln = 30 p Age, years#61.9 (8.8)63.9 (10.8)0.436Number of intervention sessions#11.9 (2.5)11.9 (2.1)0.966Days since stroke onset#18.2 (9.9)21.3 (10.1)0.246Ischaemic stroke, n(%)22 (78.6%)25 (83.3%)0.644Sex (male: female)20 : 821 : 90.905FMA-UE#Proximal19.7 (6.0)19.0 (8.6)0.700Distal8.9 (8.0)9.7 (6.9)0.679Coordination0.7 (1.1)0.6 (1.0)0.681Total29.3 (13.7)29.2 (15.6)0.990SULCS#3.9 (2.8)3.6 (3.2)0.747SIS-16#42.9 (11.2)39.0 (9.4)0.159DHI#68.6 (28.3)72.5 (21.3)0.559#Data are expressed as mean (standard deviation)SSR sensor-assisted soft robotic exoskeleton, FMA-UE Fugl-Meyer Assessment of Upper Extremity, SULCS Stroke Upper Limb Capacity Scale, SIS-16 Stroke Impact Scale-16, DHI DuruÖz Hand Index Baseline demographic and clinical characteristics of the study participants in the modified intention-to-treat analysis #Data are expressed as mean (standard deviation) SSR sensor-assisted soft robotic exoskeleton, FMA-UE Fugl-Meyer Assessment of Upper Extremity, SULCS Stroke Upper Limb Capacity Scale, SIS-16 Stroke Impact Scale-16, DHI DuruÖz Hand Index Intervention exposure Participants in the SSR and control group completed a mean of 11.9 (SD 2.5) and 11.9 (SD 2.1) intervention sessions, respectively. For the SSR group, participants spent 56.8 (SD 37.3)% and 43.2 (SD 37.25)% of time on active-assisted mode and passive mode, respectively. Based on device-generated training logs, participants completed 53.9 (SD 37.0) and 42.4 (SD 40.0) repetitions per session on active-assisted and passive mode, respectively. Six participants completed fewer than the pre-specified minimum of 10 intervention sessions (four were discharged early due to social reasons and two had rehabilitation interrupted by medical events, all deemed unrelated to study intervention) but these participants were all retained in the mITT analysis. At the 3-month follow-up (T3), outcome data was unavailable for eight participants (four in each group). One participant in each group experienced a recurrent stroke and was excluded, while the remaining six were lost to follow-up. Little’s MCAR test indicated that data were missing completely at random (χ2 = 37.0, dF = 32, p = 0.249), supporting the validity of the linear mixed-model approach used to handle missingness. Overall attrition at the primary endpoint (T2) was low and less likely to introduce significant bias. Primary and secondary outcomes In the mITT analysis, all four outcomes (FMA-UE, SULCS, SIS-16 and DHI) improved significantly from T1 to T3 (time main effect p < 0.005 for all outcomes, Supplementary Table S2). There was no statistically significant between-group difference for all the outcomes concerned at T2 and T3 (group x time p > 0.05 for all outcomes, Table 2a, Fig. 3, and Supplementary Figure S1). When comparing the proportion of participants achieving MCID, there was no significant difference between two groups (Table 3a). A sensitivity analysis comparing the whole randomised cohort (n = 61) with the mITT cohort was performed (Supplementary Table S3), showing generally consistent results between the two models. Table 2Linear mixed model analysis for all outcome measuresOutcome measureTimeGroup LS mean (SE)Between-group difference (SSR vs. Control)Group x TimeSSRControlMean difference (95% CI) p d p a) All participants (modified intention-to-treat) FMA-UE ProximalT226.8 (0.9)25.8 (0.9)0.98 (-1.61 to 3.56)0.4510.200.443T328.5 (1.1)26.3 (1.1)2.21 (-0.98 to 5.40)0.1700.39 FMA-UE DistalT215.7 (0.9)14.8 (0.8)0.86 (-1.59 to 3.30)0.4860.180.866T318.4 (1.1)17.3 (1.0)1.12 (-1.91 to 4.16)0.4600.21 FMA-UE CoordinationT22.2 (0.3)1.8 (0.3)0.45 (-0.42 to 1.33)0.3040.270.123T32.8 (0.3)3.0 (0.3)-0.22 (-1.10 to 0.67)0.623-0.14 FMA-UE TotalT244.7 (1.8)42.4 (1.7)2.27 (-2.64 to 7.18)0.3570.240.776T349.7 (2.2)46.6 (2.1)3.10 (-2.95 to 9.15)0.3080.29 SULCST26.0 (0.4)5.7 (0.4)0.32 (-0.72 to 1.37)0.5370.160.846T37.4 (0.5)7.0 (0.5)0.46 (-0.99 to 1.92)0.5260.18 SIS-16#T257.5 (1.7)55.7 (1.7)1.72 (-3.19 to 6.63)0.4850.190.465T364.5 (2.7)65.0 (2.6)-0.58 (-8.27 to 7.10)0.879-0.04 DHI#T247.2 (3.9)45.9 (3.7)1.38 (-9.37 to 12.13)0.7980.070.585T335.3 (4.8)30.6 (4.6)4.63 (-8.65 to 17.91)0.4860.20b) Subgroup: number of intervention sessions ≥ 10 and baseline FMA-UE total < 35 FMA-UE ProximalT224.3 (1.4)23.2 (1.3)1.11 (-2.91 to 5.14)0.5780.190.475T325.8 (1.7)23.1 (1.6)2.69 (-2.11 to 7.48)0.2630.38 FMA-UE DistalT214.0 (1.1)10.9 (1.0)3.06 (0.02 to 6.09) 0.049 0.660.804T316.6 (1.6)14.1 (1.6)2.50 (-2.07 to 7.08)0.2730.36 FMA-UE CoordinationT21.9 (0.4)1.2 (0.4)0.73 (-0.45 to 1.92)0.2170.410.179T32.3 (0.4)2.3 (0.4)-0.03 (-1.18 to 1.11)0.952-0.02 FMA-UE TotalT239.8 (2.4)35.7 (2.3)4.08 (-2.75 to 10.90)0.2340.390.955T344.2 (3.2)39.9 (3.1)4.31 (-4.76 to 13.38)0.3420.31 SULCST25.1 (0.5)4.2 (0.5)0.89 (-0.44 to 2.22)0.1840.440.468T36.3 (0.7)6.1 (0.7)0.23 (-1.86 to 2.32)0.8240.07 SIS-16T254.8 (2.0)49.7 (1.9)5.08 (-0.63 to 10.79)0.0790.590.886T363.7 (3.5)58.0 (3.4)5.67 (-4.22 to 15.56)0.2520.37 DHIT255.9 (5.3)59.3 (5.0)-3.42 (-18.22 to 11.38)0.642-0.150.519T343.8 (6.9)42.0 (6.6)1.72 (-17.58 to 21.02)0.8570.06#SIS-16 and DHI were missing for T2 and T3, therefore the LMM model only included n = 57 for these 2 outcomesLS mean least-square mean, SE standard error, CI confidence interval, SSR sensor-assisted soft robotic exoskeleton, FMA-UE Fugl-Meyer Assessment of Upper Extremity, SULCS Stroke Upper Limb Capacity Scale, SIS-16 Stroke Impact Scale-16, DHI DuruÖz Hand Index Linear mixed model analysis for all outcome measures #SIS-16 and DHI were missing for T2 and T3, therefore the LMM model only included n = 57 for these 2 outcomes LS mean least-square mean, SE standard error, CI confidence interval, SSR sensor-assisted soft robotic exoskeleton, FMA-UE Fugl-Meyer Assessment of Upper Extremity, SULCS Stroke Upper Limb Capacity Scale, SIS-16 Stroke Impact Scale-16, DHI DuruÖz Hand Index Between-group differences for selected outcome measures. Baseline scores are presented as unadjusted means, while post-intervention and follow-up values are least-square from the linear mixed model. Error bars represent 95% confidence intervals. *p = 0.049. SSR sensor-assisted soft robotic exoskeleton, FMA-UE Fugl-Meyer Assessment of Upper Extremity, SULCS Stroke Upper Limb Capacity Scale, SIS-16 Stroke Impact Scale-16, DHI DuruÖz Hand Index Table 3Proportions of participants achieving minimal clinically important difference (MCID)OutcomeTimeSSRn/N (%)Control n/N (%)Adjusted OR (95% CI)ARD* (95% CI) p (a) All participants (modified intention-to-treat) FMA-UE DistalT218/28 (64.3%)17/30 (56.7%)1.36 (0.46 to 3.99)7.6% (-16.8% to 30.7%)0.579T320/24 (83.3%)19/26 (73.1%)1.61 (0.36 to 7.10)10.3% (-13.0% to 31.9%)0.532FMA-UE TotalT217/28 (60.7%)16/30 (53.3%)1.32 (0.43 to 4.03)7.4% (-17.2% to 30.7%)0.623T320/24 (83.3%)20/26 (76.9%)1.20 (0.25 to 5.69)6.4% (-16.2% to 27.9%)0.816 SIS-16T221/27 (77.8%)19/29 (65.5%)2.26 (0.62 to 8.19)12.3% (-11.3% to 33.8%)0.217T321/24 (87.5%)21/26 (80.8%)1.47 (0.29 to 7.36)6.7% (-14.7% to 27.1%)0.638(b) Subgroup: number of intervention sessions ≥ 10 and baseline FMA-UE total < 35 FMA-UE DistalT214/18 (77.8%)11/20 (55.0%)3.83 (0.81 to 18.02)22.8% (-7.2% to 47.4%)0.089T316/17 (94.1%)12/18 (66.7%)7.81 (0.82 to 74.22)27.5% (0.3% to 50.9%)0.074 FMA-UE TotalT214/18 (77.8%)13/20 (65.0%)1.70 (0.38 to 7.56)12.8% (-15.7% to 38.2%)0.489T314/17 (82.4%)13/18 (72.2%)1.74 (0.34 to 8.86)10.1% (-17.8% to 35.9%)0.505 SIS-16T214/18 (77.8%)11/20 (55.0%)6.52 (1.00 to 42.58)22.8% (-7.2% to 47.4%)0.050T315/17 (88.2%)13/18 (72.2%)3.41 (0.52 to 22.39)16.0% (-11.2% to 40.6%)0.202SSR sensor-assisted soft robotic exoskeleton, CI confidence interval, OR odds ratio, ARD absolute risk difference, FMA-UE Fugl-Meyer Assessment of Upper Extremity, SIS-16 Stroke Impact Scale-16*ARDs are provided for clinical interpretability; in cases of discrepancy between the ARD CI and the p-value, the regression-derived p-value takes precedence Proportions of participants achieving minimal clinically important difference (MCID) SSR sensor-assisted soft robotic exoskeleton, CI confidence interval, OR odds ratio, ARD absolute risk difference, FMA-UE Fugl-Meyer Assessment of Upper Extremity, SIS-16 Stroke Impact Scale-16 *ARDs are provided for clinical interpretability; in cases of discrepancy between the ARD CI and the p-value, the regression-derived p-value takes precedence Subgroup analysis This subgroup analysis was post-hoc and exploratory. We analysed participants who completed at least 10 intervention sessions as planned and whose baseline FMA-UE total score was < 35, indicating moderate to severe impairment [28]. This subgroup analysis included 38 participants (18 in SSR and 20 in control). There was no significant overall group x time interaction among all 4 outcome measures in the linear mixed model (Table 2b, Fig. 3, and Supplementary Figure S1). Notably, there was a statistically significant difference in FMA-UE distal sub-score at T2 favouring the SSR group (mean difference 3.06; 95% CI 0.02 to 6.09; d = 0.66, p = 0.049), although there was no difference at T3. There was a numerically higher proportion of participants in SSR group than control achieved MCID of FMA-UE distal at both T2 (77.8% vs. 55.0%, ARD 22.8% [95% CI -7.2% to 47.4%], adjusted OR 3.83 [95% CI 0.81 to 18.02], p = 0.089) and T3 (94.1% vs. 66.7%, ARD 27.5% [95% CI 3.0% to 50.9%], adjusted OR 7.81 [95% CI 0.82 to 74.22], p = 0.074) though both not achieving statistical significance (Table 3b). A similar pattern was also observed for SIS-16, where at T2 the SSR group achieved higher score (mean difference 5.08, 95% CI -0.63 to 10.79), d = 0.59, p = 0.079), and a numerically higher proportion of participants achieved MCID than control (77.8% vs. 55.0%, ARD 22.8% [95% CI -7.2% to 47.4%], adjusted OR 6.52 [95% CI 1.00 to 42.58], p = 0.050), although neither was statistically significant. There was no significant difference at T3. No significant difference was observed for FMA-UE total, SULCS and DHI. No adverse events related to the intervention were reported during the entire trial. No device-related complications or safety concerns were identified. Discussion Key findings We evaluated the efficacy of sensor-assisted soft robotic exoskeleton glove as an adjunct for post-stroke rehabilitation in participants with subacute stroke. Using the modified intention-to-treat analysis, SSR did not demonstrate superior outcomes compared to supervised GRASP therapy, as improvements in functional capacity, upper-limb motor impairment and patient-reported outcomes were similar between the two groups. No statistically significant differences were observed between intervention and control group. The exploratory subgroup analyses indicated a possible short term improvement for distal hand motor outcomes among participants with greater initial upper-limb impairment and with adequate therapy exposure. The absence of statistically significant between-group differences between SSR and supervised GRASP should be interpreted in the context of an evidence-based active comparator, rather than usual care alone. GRASP is an established upper-limb rehabilitation programme and has demonstrated efficacy in improving upper-limb function [18]. GRASP can be readily implemented in routine clinical practice. This study intended to evaluate whether SSR could provide additional benefit beyond an established upper-limb rehabilitation approach rather than compare it with usual care alone. Though originally designed as a self-administered exercise, GRASP was supervised by therapy assistants in our study to ensure participants’ adherence to proper practice and compliance. The use of an active comparator, GRASP, likely limited the ability to detect incremental effects of SSR, unless experimental intervention provides significantly greater dose, task-specificity or participant engagement compared to the control group’s programme [29]. This may reflect the strong effect of the active comparator rather than a lack of efficacy of SSR. Another plausible factor that may have contributed to the lack of statistically significant between-group differences between control and SSR groups was that both groups received individualised therapy, whereby occupational therapists adjusted training modes and progression according to each participant’s progression and functional status. This individualised and patient-centered therapy might reflect routine clinical practice in inpatient stroke rehabilitation; however, this pragmatic approach may also reduce the contrast between the two treatment arms. Future studies may study stratified training approaches based on baseline impairment severity and more standardized intervention protocols to better isolate the specific beneficial effects of SSR intervention. Although statistically significant superiority was not demonstrated, the study demonstrates that SSR-based training can be feasibly delivered in routine inpatient rehabilitation workflows, using a model where occupational therapists prescribed training modes and parameters while trained therapy assistants administered sessions. This delivery model demonstrates the potential scalability of SSR delivery in real-world services. SSR features, dosage as a determinant of treatment effectiveness and responsiveness to treatment dependent on severity of initial impairment Many earlier versions of glove robotics provide passive movement and there is less emphasis on multi-mode training or sensor-assisted active modes [30, 31]. The SSR system provides pneumatically driven distal hand exercises such as flexion and extension assistance, grasp, and release practice, with additional training functions including active-assisted range of motion and gamified therapy enabled through performance monitoring and pressure sensing. One possible explanation for the similar outcomes between SSR and control group is the relatively insufficient intervention dosage. While we did not have data for control group, the number of repetitions for participants in the SSR group was approximately 96 repetitions per session. While this was still greater than the 30–50 repetitions/session during typical conventional rehabilitation [32, 33], some studies observed that 300–800 repetitions/session might be required, although the number of repetitions required to generate long lasting neuroplastic changes is still unknown [34]. Moreover, participants in the both groups received a mean of approximately 11 sessions. In comparison, one RCT showing benefit of SSR in subacute stroke rehabilitation had 20 intervention sessions spanning 4 weeks [13]. The importance of intervention dose and training duration is supported by landmark neurorehabilitation trials, demonstrating positive effects when intervention was delivered at high training doses [35, 36]. A higher dose of SSR training including longer session duration, increment of number of sessions or continuation beyond discharge, may be required to maximize between-group differences. In addition, this study included a wide range of participants with mild-to-severe upper-limb impairment. While this reflects real-world generalisability, the wide heterogeneity in baseline severity may reduce statistical power to detect treatment effects with ceiling effects in mild cases and variable recovery trajectories. Heterogeneity in baseline severity has been highlighted as a challenge in rehabilitation trials [37]. Further studies on SSR may aim at targeted recruitment of participants with moderate to severe impairment, where robotic devices may provide greater incremental value and meaningful changes in outcomes. Exploratory subgroup findings and implications for participant selection Primary analysis between SSR and control group did not demonstrate significant superior outcomes for SSR. However, in an exploratory subgroup analysis on participants who completed ≥ 10 sessions and had baseline FMA-UE < 35, a statistically significant difference in FMA-UE distal sub-score at T2 was observed, favouring the SSR group. In addition, a numerically higher proportion of participants in the SSR group exceeded MCID thresholds for FMA-UE distal score and SIS-16 at T2, as well as FMA-UE distal at T3 with though not statistically significant. These findings were associated with medium to large effect sizes. These findings should be interpreted with caution. Given the number of tests in the subgroup analysis, the probability of observing at least one significant result (p < 0.05) by chance is 0.64 [38]. The post-hoc analysis simultaneously involves baseline severity and treatment adherence/exposure and is therefore highly susceptible to selection bias. This analyses should therefore be interpreted as hypothesis-generating and are intended to inform future trial design only rather than to guide clinical decision-making. However, they are clinically plausible and align with literature that benefits of robotic training vary by baseline severity. Mehrholz et al. describes evidence that while robotic devices enable high repetition training, there can be subgroup effects and benefits vary by baseline severity [39]. Veerbeek et al. also describes robotics as a way to increase intensity and repetitions with outcomes influenced by initial impairment level [40]. Soft robotic assistance may benefit participants with greater initial impairment most by assisting to generate sufficient voluntary finger movement for high-repetition practice, allowing greater training volume and more consistent grasp-release cycles. Conversely, participants with milder deficits may not require device-based rehabilitation and may achieve similar benefits from conventional task specific training. One possible explanation is that participants with more severe hand impairment may benefit from the assistance provided by SSR during grasp and release. Participants with limited voluntary finger movement may otherwise have difficulty completing the repeated grasp release movements and the assistance provided by SSR may allow these participants to participate more effectively in repetitive task specific practice. In contrast, participants with milder impairment who already have sufficient voluntary hand movement may be able to perform similar repetitive exercises without robotic assistance and may therefore derive less additional benefit from SSR. It is also possible that wearing a full glove may add some weight to the hand, reduce tactile sensory input during object manipulation, and potentially restrict finger or hand range of motion because of the glove structure. No significant difference was found in the two measures of activity limitation, SULCS and DHI. One meta-analysis found that 16 h of extra therapy is required to obtain significant gain in activities of daily living [41]. The extra therapy time in our study was way below 16 h and therefore even if there is improvement in impairment might not be easily translated to improvement in daily function. This may be particularly relevant for participants with moderate to severe upper-limb impairment, who may require more prolonged therapy exposure before changes in impairment are reflected in functional outcomes and participation. Clinical implications and future directions Although SSR did not demonstrate superior outcomes compared to supervised GRASP in the primary analysis, SSR may still offer some practical advantages in stroke rehabilitation. The lightweight and portable design of SSR may make it easier to incorporate into routine inpatient rehabilitation and may support use of SSR in outpatient or home-based settings, extending therapy beyond the inpatient setting [7]. SSR may also have the potential to increase the intensity of task-specific practice that can be provided, especially in environments with manpower constraints [39, 40]. There is potential to explore implementing SSR training beyond inpatient discharge, at outpatient or home settings, to increase cumulative therapy time and to evaluate sustained gains during the post-stroke recovery phase. Cost-effectiveness of such intervention should be explored. The exploratory subgroup analysis findings suggest that future studies may benefit from focusing on participants with moderate to severe upper-limb impairment. This highlights the importance of a targeted rehabilitation approach to identify participants that will most likely benefit from robotic intervention and ensuring adequate training dose. In addition, given the device is mainly applied to the distal hand, the primary outcome could be a score focusing mainly on distal hand functions such as FMA-UE distal score. Based on the effect size of 0.66 for the FMA-UE at T2 for the subgroup analysis, a sample size of about 90–100 participants is required for future studies. Notably, there were two participants who were unable to proceed due to inability to achieve adequate device fit. This highlights a potential usability limitation of the device in certain participant subgroups and the importance of accommodating hand morphology variability in future designs. Limitations There are some limitations of this trial. Firstly, the number of intervention sessions achieved was lower than planned, largely due to early discharge (due to social reasons that were unrelated to the intervention), which led to a lower cumulative exposure to the intervention and may have attenuated detectable difference between control and SSR group. Secondly, although planned sample size was 130 participants based on power calculation, the achieved sample size was lower than expected; therefore, the study was underpowered to detect small to moderate group differences in primary outcome and the possibility of a type II error cannot be excluded. Thirdly, participants were recruited with a wide range of baseline upper-limb impairment severity. The heterogeneity in baseline function may increase outcome variability and reduce sensitivity to detect treatment effects, especially for participants with mild impairment due to ceiling effects. Finally, neurophysiological biomarkers as motor evoked potentials were not collected. This information may be relevant for prognostication and could improve treatment stratification in future work. Conclusion In this randomised controlled trial of subacute inpatient stroke rehabilitation, SSR did not demonstrate superior outcomes compared with supervised GRASP in the overall modified intention-to-treat analysis. Exploratory subgroup analyses suggest that SSR may provide an exploratory signal of benefit for distal hand recovery among participants with greater baseline impairment, highlighting the importance of both participant selection and adequate dosage of SSR. Future trials with larger sample sizes, higher rehabilitation doses and stratification by baseline severity will be important to define the optimal clinical role of sensor-assisted soft robotic hand rehabilitation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Acknowledgements The authors thank the participants and their families for their involvement in this study. We acknowledge the occupational therapists and therapy assistants at Alexandra Hospital and Jurong Community Hospital for their support in delivering the intervention and collecting data. We also thank Ms June Leam and Dr. Tang Ning from the Department of Medicine, National University Hospital; therapist assistant Peng Chai Meng and therapy associate Habiban Md Hassan from Jurong Community Hospital; senior therapist assistant Lee Hong Ye and therapist assistant Harris Hidayat from Alexandra Hospital for their contributions to this study. We thank Dr. Yap Hong Kai for providing technical expertise for the manuscript. We also thank Roceso Technologies for providing the soft robotic glove used in this study. Abbreviations Duruöz hand index Fugl-Meyer Assessment of Upper Extremity Graded repetitive arm supplementary programme Health-related quality of life Minimal clinically important difference Modified intention-to-treat Occupational therapist Occupational therapy assistant Randomised controlled trial Stroke impact scale-16 Sensor-assisted soft robotic exoskeleton Stroke upper limb capacity scale Author contributions KWET and PKT conceptualised and designed the study, supervised data collection, performed data analysis, and drafted the manuscript. Both contributed equally to this work and share first authorship. QX, HSL, and WKY contributed to participant recruitment, data collection, study design and interpretation of data. JHL contributed to development of the intervention device and provided technical expertise, and critical revision of the manuscript. XMLL, SLN, ML, LMO, SR and AT contributed to participant recruitment and data collection. All authors read and approved of the final manuscript. Funding This work was supported by Enterprise Singapore (Grant No. 2009TC3N) and Roceso Technologies, Singapore. Roceso Technologies provided the intervention device used in this study. The funders had no role in study design, data collection, data analysis, data interpretation, or manuscript preparation. Data availability The data generated and analysed in this study are not publicly available due to ethical and institutional restrictions related to patient confidentiality. The data can be made available from the corresponding author upon reasonable request and with appropriate approvals. All relevant data supporting the findings of this study are presented within the article and/or in the Supplementary Information file uploaded. Declarations Ethics approval and consent to participate The study was approved by the National Healthcare Group Domain Specific Review Board (NHG DSRB; Reference No. 2020/01105). All participants provided written informed consent prior to enrolment. The study was conducted in accordance with the Declaration of Helsinki and relevant institutional guidelines. Consent for publication Not applicable. Competing interests JHL is a named inventor on patents related to the soft robotic glove technology evaluated in this study. Roceso Technologies was involved in funding the study and provided the device used in the intervention. The remaining authors declare that they have no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement The data generated and analysed in this study are not publicly available due to ethical and institutional restrictions related to patient confidentiality. The data can be made available from the corresponding author upon reasonable request and with appropriate approvals. All relevant data supporting the findings of this study are presented within the article and/or in the Supplementary Information file uploaded.