Introduction

Stroke is a cerebrovascular disease with high morbidity and mortality rates, imposing a significant economic burden on families and society [1]. Research indicates that annually, nearly 4 million individuals in China Suffer a new-onset stroke, with approximately 60% experiencing lower limb motor dysfunction, severely impacting their quality of life [2, 3]. Furthermore, enhancing lower limb motor ability and gait function are primary rehabilitation needs for these patients [4], improving lower limb motor dysfunction a critical and challenging issue in contemporary rehabilitation medicine.

In recent years, the advancement of rehabilitation robots has led to the adoption of wearable exoskeleton robots as viable alternatives and supplements to manual rehabilitation [5].Among rehabilitation robots for the post-stroke population, hip-wearable exoskeletons have emerged as a prominent research focus. The basic principle of hip joint assistance is that these devices operate by continuously monitoring the wearer’s gait via integrated sensors; the drive system then delivers timely power assistance during the appropriate phase of the gait cycle, thereby augmenting hip joint motion function [6]. Biomechanically, targeting the hip joint is advantageous because it simultaneously bears body weight and directly drives lower limb movement-providing powered assistance at this segment enhances overall lower limb movement efficiency [7]. Several studies [8,9,10] suggested that, compared to conventional walking training, exoskeleton robot-assisted training could provide greater intensity training, with demonstrated benefits for motor function improvement. Moreover, hip exoskeletons have been shown to reduce metabolic energy expenditure more effectively than exoskeletons focusing on other joints [11,12,13]. Collectively, these factors indicate the considerable potential of hip-wearable exoskeletons for facilitating stroke rehabilitation.

Several investigations [14,15,16] have demonstrated that hip-wearable exoskeletons can deliver continuous, precise, task-oriented robot-assisted gait training(RAGT) to stroke survivors, resulting in measurable improvements in gait and motor function. However, most existing hip-wearable exoskeletons employ a single control strategy and exhibit limited capability to recognize or predict the user’s movement state (“low motion intention recognition rate”) [17, 18]. Consequently, these devices cannot adapt assistance in real time to each patient’s unique needs (i.e., not being able to achieve adaptive real-time control) [17, 18]. To address these limitations, we have developed a novel real-time assistive hip-wearable exoskeleton that integrates multi-sensor data fusion for enhanced motion-intent recognition and prediction, coupled with an adaptive control strategy that obviates the need for therapists to manually adjust control parameters [19]. Instead, the system dynamically modifies its control algorithm in response to the user’s gait performance, thereby delivering truly individualized walking training [20]. Preliminary trial in healthy volunteers has confirmed the control system’s robust stability and resilience to external disturbances. Nonetheless, its clinical efficacy in the stroke population has yet to be evaluated [20].

Accordingly, the present trial was designed to assess the impact of our novel hip-wearable exoskeleton on lower-extremity rehabilitation in individuals with subacute stroke, with a primary focus on gait function and secondary assessments of lower limb motor impairment and balance. The subacute phase of stroke recovery is widely regarded as critical for functional restitution [21], and prior studies [10, 22] indicate that patients in this phase derive particularly pronounced benefits from RAGT. Therefore, we specifically enrolled subacute stroke patients in this study. We hypothesized that, compared to therapist-assisted gait training, use of the hip-wearable exoskeleton would yield superior improvements in gait performance, lower limb motor recovery, and balance in this population.

Methods

Study design

This study was a single-blind, randomized controlled trial involving patients with Subacute stroke from the Department of Rehabilitation Medicine. Participants were randomly assigned in a 1:1 ratio to either the RAGT group or the control group. To maintain the integrity of the randomization process, an independent researcher, who was not involved in study implementation, used SPSS version 22.0 to generate a random number sequence. These numbers were placed in consecutively numbered, opaque envelopes to ensure allocation confidentiality. A physiotherapist, blinded to the random sequence, then assigned participants to either the RAGT group or the control group based on the envelope contents. Both groups underwent their respective interventions five days per week over a four-week period. The primary outcome was gait function, while secondary outcomes comprised the Fugl-Meyer Assessment for Lower Extremity Scale(FMA-LE), the Berg Balance Scale(BBS), and the Timed Up and Go Test(TUGT). All outcome measures were collected at baseline and again at four weeks post-intervention.To minimize measurement bias, a second physiotherapist, who was also blinded to group assignments, conducted all assessments.

The study adhered to the Consolidated Standards of Reporting Trials (CONSORT) guidelines and the Helsinki Declaration. Ethical approval was granted by the Ethics Committee of Beijing Jishuitan Hospital Guizhou Hospital (approval number: LW20230807-2), and the trial was registered at www.chictr.org.cn (registration number: ChiCTR2300074562, dated09/08/2023). Informed consentwas obtained from all the subjects.

Sample size

Sample size estimation was conducted using G*Power version 3.1.9.7.The effect size calculation was derived from the FMA-LE score post-intervention in a previous study [23].We utilized the statistical method of an independent samples t-test, which indicated an effect size of 0.97.Referencing Kuo et al. [24] for additional parameters, we set the alpha level at 0.05 and the statistical power at 0.8.Based on these criteria, it was estimated that a minimum of 18 patients per group would be necessary. To account for a potential 10% dropout rate, we planned to include 20 patients in each group.

Participants

A total of 40 patients with Subacute stroke were recruited between August 2023 and July 2024 in this study. Inclusion criteria: (1)the patients met the diagnostic criteria for stroke established by the World Health Organization (WHO) [25]; (2) first-ever unilateral stroke; (3) aged 40 to 80 years old; (4)14 days to 180 days after stroke onset; (5) the range of Functional Ambulatory Category(FAC) scale was 3 to 4 [26], and the patients were able to walk more than 6 m independently or under supervision without assistive device; (6) no severe spasticity in the paretic lower limb(Modified Ashworth Scale < 3); (7) no cognitive impairment(Mini-Mental State Examination score ≥ 24 [24]).

The exclusion criteria were: (1) bilateral impairment; (2) visual or auditory impairments that could interfere with rehabilitation; (3) severe cardiovascular, endocrine, or tumor diseases; (4) severe musculoskeletal disorders that could impact mobility.

Hip-wearable exoskeleton robot device

The prototype of the hip-wearable exoskeleton robot and description of robot functions are illustrated in Supplementary Fig. 1.This robot is equipped with two active degrees of freedom, specifically designed for hip flexion and extension, with the aim of assisting patients in walking. The robot’s system architecture is divided into three main subsystems: the sensor perception system, the main controller decision system, and the joint motor drive system. The sensor perception system acquires real-time data on patient posture, plantar pressure, and joint angles during ambulation via an array of embedded sensors. The main controller decision system processes these inputs to recognize and predict the patient’s movement state, inferring motion intention, and then calculates the necessary assistive torque before sending precise commands to the drive module. In response, the drive system delivers the calculated torque, up to a maximum peak of 60 N·m, to the hip joint.

Notably, our system employs a hybrid oscillator-based algorithm for gait-phase estimation, enabling adaptive, phase-specific assistance during both level-ground walking and stair negotiation [20]. By dynamically adjusting support according to the patient’s instantaneous gait characteristics, this approach minimizes reliance on therapist intervention. The integration of motion-intention recognition and real-time adaptive control distinguishes our exoskeleton from conventional designs and directly contributes to improved functional outcomes such as enhanced gait symmetry and coordinated motor patterns.

Interventions

Conventional rehabilitation therapy

Conventional rehabilitation therapy, utilized in this study, was comprised of two distinct components: physical therapy and occupational therapy. Physical therapy sessions involved a variety of exercises such as stretching of the lower limbs, range of motion (ROM) training, balance exercises, and muscle strength training. Meanwhile, occupational therapy focused on enhancing the patients’ daily living skills through activities of daily living training, occupational skills exercises, and functional task training. Each therapy type was administered for 30 min per session, twice daily, five days a week, over a period of four weeks.

Therapist-assisted gait training

Therapist-assisted gait training was administered for 35 min per session, once daily, five days a week, over a span of four weeks. During each session, a physiotherapist provided either lateral or posterior Support while patients performed overground walking exercises for 20min, followed by stair-climbing activities for 15 min. Stair training was conducted on a Handrail-equipped staircase comprising 10 steps, each 150 mm in height. Participants completed four sets of stair training, with each set consisting of one ascent and one descent.

RAGT

The RAGT encompassed overground walking training and exercises involving ascending and descending stairs, each session lasting 35 min, conducted once daily, five days a week, for a total Duration of four weeks. Initially, the walking speed was set at a low pace of 0.33 m/s [27], and the control parameters of the exoskeleton robot were pre-set, requiring no adjustments During the training sessions. Patients were required to undergo adaptive walking training for 3 days before enrollment.

During the first 20 min of each session, patients walked freely according to their individual capabilities. Throughout this period, the exoskeleton robot actively collected gait data from the patients, enabling it to provide predictive and adaptive assistance based on the analyzed gait patterns. The final 15 min consisted of stair-climbing exercises on a Handrail-equipped staircase with 10 steps (each 150 mm in height). Participants completed four sets, with each set comprising one ascent and one descent. A physiotherapist remained alongside each patient to provide assistance and ensure safety throughout the training.

Outcome measurements

Primary outcomes

Gait function was assessed in the biomechanics laboratory equipped with a BTS motion-capture system and four force plates. The motion-capture system comprised eight optical infrared cameras (BTS SMART DX4000, Italy, sampling rate: 250 Hz). Additionally, these force plates (BTS P-6000, Italy, sampling rate: 500 Hz [28]) were strategically positioned in the middle of a 6-meter walkway.

Twenty-two retro-reflective markers were positioned on the skin surface following the “Helen Hayes with Medial Markers” protocol [29, 30]. The exact locations of these markers are outlined in Supplementary Table 1. For formal assessments, patients initially stood on force plates to collect static kinematic and kinetic data. Subsequently, they were instructed to walk along the walkway at their normal gait speed for three trials, each capturing more than five valid gait cycles. Data analysis was conducted using BTS SMART Analyzer software. The kinematic and kinetic raw data from both lower limbs were processed through a low-pass fourth-order Butterworth filter, applying cutoff frequencies of 6Hz for kinematic data [31] and 50 Hz for kinetic data [32]. Kinematic parameters included peak hip flexion/extension, peak knee flexion/extension, and peak ankle dorsiflexion/plantarflexion on both the paretic and non-paretic sides, measured in degrees (°).Positive (+) signs denoted flexion and dorsiflexion angles, whereas negative (−) signs denoted extension and plantarflexion angles. Kinetic parameters featured the peak vertical ground reaction force (vGRF) on both the paretic and non-paretic sides, standardized by the subject’s body weight and expressed in N/kg. Additionally, asymmetry in gait kinematics and kinetics for both lower limbs was quantified using the ASI formula [33], where a higher value indicated a greater degree of asymmetry. The formula for ASI, denoted as X for both kinematic and kinetic variables, is as follows:

$$\:\:\:\text{A}\text{S}\text{I}\left(\text{\%}\right)=\left|\frac{{100\times\:(\text{X}}_{\text{P}\text{a}\text{r}\text{e}\text{t}\text{i}\text{c}\:\text{s}\text{i}\text{d}\text{e}}-{\text{X}}_{\text{N}\text{o}\text{n}-\text{p}\text{a}\text{r}\text{e}\text{t}\text{i}\text{c}\:\text{s}\text{i}\text{d}\text{e}\:\:})}{0.5\times\:(\:{\text{X}}_{\text{P}\text{a}\text{r}\text{e}\text{t}\text{i}\text{c}\:\text{s}\text{i}\text{d}\text{e}}+{\text{X}}_{\text{N}\text{o}\text{n}-\text{p}\text{a}\text{r}\text{e}\text{t}\text{i}\text{c}\:\text{s}\text{i}\text{d}\text{e}\:\:})}\right|$$
(1)

In this study, spatiotemporal parameters were further calculated from the kinematic data of the foot, including gait speed (m/s), cadence (steps/min), step length (m), and step width (m).

Secondary outcomes

The Fugl-Meyer assessment for lower extremity scale (FMA-LE)

The FMA-LE is a well-validated and reliable tool used to assess lower limb motor impairment in stroke patients [34]. The scale comprises 17 items, yielding a total possible score of 34. Higher scores on this scale indicate better motor impairment.

Berg balance scale (BBS) and the timed up and go Test(TUGT)

The balance ability of stroke patients was evaluated using the BBS and the TUGT. The BBS is a well-validated tool specifically for assessing balance in stroke patients [35, 36]. It consists of 14 items, each rated from 0 to 4 points, with higher scores indicating better balance. The TUGT assesses dynamic balance and involves a sequence of actions [37, 38]: the patient starts by sitting in a chair with a backrest, then stands up, walks 3 m at a normal pace, turns around, walks back to the chair, and sits down. The duration, measured in seconds, from the moment the patient’s back leaves the chair to when it touches the backrest again, is recorded. Longer times suggest a lower balance ability.

Statistical analysis

Statistical analysis was conducted using SPSS V22.0 software. Continuous data were expressed as mean ± standard deviation, and the normality of the data distribution was assessed using the Shapiro-Wilk test. Demographic characteristics and baseline outcome measures were compared by Chi-square test or independent- samples t-test between groups.

Within-group differences between pre-and post-intervention were analyzed using the paired-samples t-test [39, 40]. In addition, analysis of covariance (ANCOVA) was utilized to assess between-group differences after intervention, using baseline values as covariates [39, 40]. The level of significance was set at P < 0.05 [41].

Results

Participant characteristics

Figure 1depicts the flow chart of the study. A total of 63 stroke patients were assessed for eligibility, with 23 patients being excluded from the study (15 didn’t meet inclusion criteria, 4 declined to participate, and 4 other reasons). Supplementary Table 2 provides a patient-by-patient listing of exclusion reasons for the 15 cases not meeting inclusion criteria. Consequently, 40 stroke patients were ultimately enrolled and equally randomized into the RAGT group and the control group. Furthermore, no significant differences were found in the demographic characteristics between the two groups of patients (Table 1).

Fig. 1
Fig. 1
Full size image

Flow chart of the study

Table 1 Demographic characteristics of patients
Full size table

Primary outcomes

Table 2 shows the results concerning kinematic parameters. The RAGT group demonstrated significant enhancements in peak hip flexion/extension and peak knee flexion across both lower limbs before and after the intervention(P<0.05). Although peak hip flexion on the paretic side in the control group also showed significant improvement post-intervention (P<0.05), the values for peak hip flexion/extension and peak knee flexion in both lower limbs of the RAGT group were significantly superior post-intervention compared to those of the control group(P<0.05).In addition, the ASI of peak hip flexion/extension and peak knee flexion in the RAGT group after intervention were significantly reduced compared to those before intervention(P<0.05). Compared with the control group, the RAGT group had a lower ASI value in peak hip flexion/extension and peak knee flexion after intervention(P<0.05).

Table 2 Comparison of kinematic data and kinetic data before and after intervention in both groups
Full size table

Table 2shows the results of kinetic parameters. Compared to before intervention, peak vGRF on the paretic side was significantly higher in both groups after intervention, while the ASI of the peak vGRF was lower(P<0.05). A significant difference was observed in peak vGRF on the paretic side (F = 7.11, η2P = 0.30, P = 0.016) and ASI of the peak vGRF (F = 9.61, η2P = 0.34, P = 0.006) between the two groups after intervention.

Figure 2 shows the results of spatiotemporal parameters, revealing that both gait speed and cadence improved significantly in both groups post-intervention compared to pre-intervention metrics. Specifically, the RAGT group post-intervention exhibited notable increases in gait speed (F = 6.99, η2P = 0.27, P = 0.016), step length (F = 26.99, η2P = 0.59, P<0.001), and cadence (F = 24.55, η2P = 0.56, P<0.001) compared to the control group. However, there was no significant change in step width between the groups.

Fig. 2
Fig. 2
Full size image

Comparison of spatiotemporal parameters before and after intervention in both groups. A Gait speed; B Step length; C Cadence; D Step width. *P<0.05, **P<0.01, ***P<0.001

Notably, the findings including hip/knee kinematics, vGRF on the paretic side, and gait asymmetry may be more clinically or mechanistically relevant.

Secondary outcomes

Figure 3A displays the results of the FMA-LE. After the intervention, significant improvements were observed in the FMA-LE scores for both groups compared to their baseline scores (RAGT group: t=−4.221, P = 0.002; Control group: t=−4.302, P = 0.002). Additionally, the post-intervention FMA-LE score for the RAGT group was significantly higher than that of the control group (F = 13.21, η2P = 0.41, P = 0.002).

Figure 3B shows the results of BBS. BBS scores in both groups after intervention were significantly improved compared to the pre-intervention level (RAGT group: t =−14.807, P<0.001;Control group: t =−5.310,P<0.001). A significant difference was observed in BBS score between the two groups after intervention (F = 16.57, η2P = 0.47, P = 0.001).

Figure 3C shows the results of TUGT. The RAGT group showed significant within-group improvement in TUGT before and after intervention (t = 4.540, P = 0.001).

Fig. 3
Fig. 3
Full size image

Comparison of FMA-LE, BBS and TUGT before and after intervention in both groups. A FMA-LE, B BBS, C TUGT. *P<0.05, **P<0.01, ***P<0.001

A significant difference was observed in TUGT between the two groups after intervention (F = 27.03, η2P = 0.59, P<0.001).

Discussion

Key findings

The primary objective of this trial was to evaluate the impact of a newly developed wearable hip exoskeleton robot on the lower extremity rehabilitation in subacute stroke patients. The key findings indicate that RAGT was more effective than therapist-assisted gait training in enhancing gait function and reducing gait asymmetry in subacute stroke patients. Additionally, RAGT significantly improved lower limb motor impairment and balance in this population. These results validate the hypothesis that training with our novel hip-wearable exoskeleton yields greater improvements in lower-extremity function for subacute stroke patients than conventional therapist-assisted gait training.

Gait improvements

This study demonstrated that RAGT significantly enhanced bilateral hip-knee coordination and gait symmetry in stroke patients, while revealing a critical limitation in ankle function improvement due to the device’s passive ankle joint design. In this study, the spatiotemporal results highlighted that the RAGT and control groups exhibited significant improvements in gait speed and cadence. Kinematic evaluations revealed a comparable pattern: the RAGT group exhibited significant enhancements in peak hip flexion and extension, peak knee flexion, and their associated ASI values, indicating that robot-assisted gait training effectively normalized hip and knee joint symmetry in both lower limbs. In contrast, the control group showed improvement only in peak hip flexion, suggesting that while therapist-assisted gait training can positively influence gait kinematics, its capacity to restore bilateral joint symmetry is limited. Gait is a complex, dynamic process that relies on the coordinated activity of multiple lower-limb joints [42]. Mechanical simulations [43,44,45,46] have demonstrated that inadequate hip flexion shortens the rectus femoris force arm and increases knee-extensor loading, thereby reducing the rectus femoris’s contribution to knee flexion and ultimately impeding knee flexion. In stroke patients, restricted hip flexion alters hip–knee biomechanical coupling, resulting in decreased knee-flexion angles on the paretic side [7, 31]. In the present study, the exoskeleton provided real-time torque assistance to the hip joint during ambulation, directly augmenting hip flexion and concurrently correcting abnormal hip-knee coupling to facilitate knee flexion, thus improving gait symmetry. However, RAGT did not enhance the ankle ROM, likely because the device lacks active control of the ankle joint. This absence of active ankle engagement may have limited improvements in ankle function; future designs should therefore incorporate ankle-control strategies.

vGRF and compensatory strategies

vGRF analysis demonstrated that RAGT effectively mitigated paretic limb unloading and improved dynamic weight-bearing symmetry in stroke patients, suggesting superior gait rehabilitation outcomes compared to conventional therapy. Research indicates that stroke patients often adopt compensatory strategies during walking, which typically results in higher peak vGRF on the non-paretic side compared to the paretic side [7, 47]. This compensatory behavior not only leads to vGRF asymmetry between the lower limbs but also heightens the risk of falls [48].The results reveal that both the RAGT and control groups exhibited a significant increase in peak vGRF on the paretic side and a reduction in peak vGRF asymmetry; however, the improvement was more pronounced in the RAGT group. These improvements can be attributed to RAGT enhancing balance and promoting more symmetrical movement of the hips and knees in both lower limbs [49]. This, in turn, strengthens the ability of the paretic limb to transfer and bear body weight more effectively [49].Furthermore, RAGT training may boost the confidence of stroke patients in walking independently, reducing their reliance on compensatory strategies [50]. This leads to a more balanced state of weight-bearing across both lower limbs, thus promoting a more normalized gait pattern. Overall, RAGT offers substantial benefits for the recovery of gait function in stroke patients.

Motor & balance mechanisms

The results of the FMA-LE demonstrated that both RAGT and control groups showed significant improvements in lower-limb motor impairment compared to baseline, confirming the efficacy of both interventions. Notably, the RAGT group produced significantly greater improvements than the control group, with higher post-intervention FMA-LE scores despite equivalent training intensity and duration. This superior therapeutic outcome likely stems from RAGT’s combination of motor training with enhanced sensory stimulation.

Previous research [51,52,53] has shown that pairing repetitive motor practice with augmented sensory feedback promotes cortical reorganization and strengthens motor-network connectivity, thereby enhancing neuroplasticity and reducing motor deficits. In the present study, RAGT delivered continuous, precisely timed torque assistance synchronized to the gait cycle, enabling patients to practice overground walking in a realistic context and facilitating relearning of correct movement patterns. Simultaneously, RAGT encouraged weight-bearing on the paretic limb and decreased lower-limb loading asymmetry (as evidenced by vGRF results), thereby increasing afferent input from that side. These sensory signals travel via peripheral nerves to the spinal cord and sensory cortex, improving patients’ awareness of limb position and movement [54, 55]. Enhanced proprioceptive feedback, combined with repetitive task-specific practice, likely underlies the observed improvements in motor precision and coordination. Consequently, RAGT may more effectively reduce lower-limb motor impairment by leveraging a synergistic effect of motor training and sensory augmentation. In addition, the exoskeleton robot satisfies the patient’s desire to walk on the ground and gives full play to the patient’s subjective initiative, which is also conducive to the patient’s access to positive psychological cues, thereby accelerating the recovery process [56]. Previous studies [2, 57] have reported the positive value of various exoskeleton robots on lower limb motor impairment recovery in stroke patients, aligning with our findings. However, the robot system used in this study did not incorporate virtual reality or varied scenarios. Future enhancements to the robot will aim to offer more diverse training content and formats to stroke patients, potentially augmenting therapeutic outcomes.

The balance function demonstrated that the RAGT group surpassed the control group in enhancing BBS scores and reducing TUGT durations, indicating that RAGT substantially improves the balance capabilities of stroke survivors. The RAGT modality used in this study enhanced patients’ ability to adjust their center of gravity by facilitating mediolateral weight shifts during gait, thereby improving both dynamic and static balance [57, 58]. Furthermore, the exoskeleton robot utilized offers real-time joint assistance, which enables a quicker response time than physiological human reactions [59], and therefore prompt correction of postural disturbances during gait activities. These findings highlight the significant clinical value of the exoskeleton robot in enhancing balance abilities in stroke patients.

Limitations

The study, while providing valuable insights, has several limitations: (1) Despite conducting a sample size estimation, the final sample size remained small; (2) The Duration of the RAGT was limited to only 4 weeks, leaving the long-term efficacy of the intervention unexplored; (3) The study did not account for variations in the type and severity of stroke among participants, factors which could significantly influence the outcomes; (4) This trial was a single-center design, which may limit the generalizability of the findings; (5) The study was limited by the lack of complete blinding and the absence of long-term follow-up and patient-reported outcomes; (6) The exoskeleton robot used in this study lacked active ankle control and detailed ankle ROM assessment.

Clinical implications & future directions

Clinical implications: This study provides preliminary evidence that RAGT delivered by a hip exoskeleton demonstrates superior efficacy over therapist-assisted gait training in improving gait function, lower limb motor impairment, and balance capacity in subacute stroke patients. From both rehabilitation efficacy and time-efficiency perspectives, hip exoskeleton-assisted training represents a viable priority intervention. However, therapist-assisted gait training remains warranted for complex cases (e.g., patients with severe lower-limb spasticity).

Future directions:​​ In terms of robotic design, we will further improve the exoskeleton’s human motion intention prediction capability and integrate active ankle control. Regarding study methodology, we plan to validate the rehabilitation value of the exoskeleton for stroke patients by expanding the sample size, adopting a multicenter design, and extending both the intervention and follow-up periods.

Conclusion

The exoskeleton employed in this study featured motion-intention prediction, real-time adaptive control, hip-specific actuation, and integrated stair-training capabilities. Our findings demonstrate that RAGT with this device effectively enhanced gait recovery and reduced gait asymmetry in subacute stroke patients. Moreover, the robot significantly improved lower limb motor impairment and balance in this population.