Can wearable real-time biofeedback gait training devices improve gait speed, balance, functional mobility and activities of daily living (ADL) in individuals post-stroke? A systematic review and meta-analysis of randomized controlled trials

Abstract

Background

Stroke remains a leading cause of long-term disability, impairing gait, balance, and mobility, which critically reduces independence and increases fall risks. Wearable biofeedback devices have been developed and widely applied for gait rehabilitation, by providing real-time monitoring and adaptive feedback to enhance motor recovery. This systematic review and meta-analysis aimed to synthesize the existing evidence on effects of wearable real-time biofeedback gait training devices on gait parameters and functional abilities in stroke survivors, to guide future clinical practice and research exploration.

Method

Databases of PubMed, EMBASE, MEDLINE, Web of Science, Cochrane Library, CINAHL, PsycInfo, PreQuest, and PEDro were searched up to Sep 13th, 2024. Randomized controlled trials (RCTs) investigating and comparing the effects of wearable real-time biofeedback gait training devices, with general rehabilitative gait training or other controls, in stroke survivors were included. The data including subject/participant characteristics, biofeedback device design/set-up, dosage of interventions, and outcome measures were extracted.

Result

A total of 13 RCTs involving 304 participants were included in this systematic review, and 11 RCTs involving 272 participants were included in the meta-analysis. Seven studies measuring gait speed showed statistically significant differences that favored biofeedback gait training over the controls (SMD = 0.41, P = 0.02, n = 204). Subgroup analyses on the efficacy of pressure sensing technology with auditory feedback showed non-significant results, although the P value was close to reaching statistical significance (SMD = 0.30, P = 0.05, n = 166). The pooled data also showed that biofeedback gait training significantly further improved stroke patients’ balance and functional mobility comparing with controls, as evaluated by the Berg Balance Scale (SMD = 0.44, P = 0.03, n = 95) and Timed Up and Go Test (SMD=-0.36, P = 0.01, n = 190), respectively. The meta-analysis showed that biofeedback training was not significantly better than the control treatment in improving activities of daily living, as measured by the Modified Barthel Index (SMD = 0.21, P = 0.38, n = 74).

Conclusions

This review provides moderate quality evidence that wearable real-time biofeedback gait training can improve balance and functional mobility in post-stroke individuals. While a positive overall trend was observed for gait speed, the most prevalent intervention type (pressure sensing with auditory feedback) did not yield a statistically significant effect. No significant benefit was found for activities of daily living. These findings suggest that biofeedback may serve as a useful adjunct to conventional therapy for improving specific aspects of motor function, including balance, functional mobility, and gait speed. Future research should focus on high-quality implementation trials with larger samples, “sham” conditions, and direct comparisons of feedback modalities.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12984-025-01863-x.

Introduction

Stroke is a leading cause of long-term disability worldwide, significantly impacting individuals and healthcare systems [1]. According to the latest report from the World Stroke Organization (WSO), approximately 12.2 million people suffer from a stroke each year globally, with 6.5 million resulting in death [1]. Technological advances have led to reduced stroke-related deaths, enabling a greater number of stroke survivors living with disability. Studies have shown that stroke often leads to significant functional impairment [2], contributing to a loss of 143 million disability-adjusted life years (DALYs) globally [3]. Stroke survivors commonly experience substantial declines in balance and gait, which not only affect their activities of daily living (ADL), but also increase the risk of falls and recurrent strokes [4, 5]. The improvement of gait pattern or walking ability has been ranked among the most desired rehabilitation goals [6]. Gait training is a common component of stroke rehabilitation programs.

Physiotherapy treatment methods following a stroke can be broadly classified into various approaches, based on neurophysiological, motor learning, and orthopedic principles [7]. Neurophysiological approaches, such as the Bobath technique, focus on adjusting the patient’s gait patterns in a more passive manner [8]. Meanwhile, approaches based on motor learning, such as those guided by therapists or assisted using biofeedback devices, emphasize active participation and adaptation from patients [9, 10]. To refine gait patterns, different motor-learning based interventions are utilized [11], including robot-assisted gait training [12, 13], weight support treadmill walking [14, 15], rhythmic auditory stimulation (RAS) [16], and various biofeedback systems [17]. While these interventions have demonstrated efficacy in several studies [12–14, 16] and some are commonly employed in clinical practice, most of them require the therapist’s guidance or large assistive devices to implement. For instance, the application of robot-assisted gait training is often limited by its substantial infrastructure requirements, including large devices and professional operation, confining its use primarily to clinical settings and thereby restricting availability for potential community application [12]. The reliance on clinical infrastructure may also pose a barrier to the generalization of training effects to patients’ daily life and activities. Thus, developing and validating continuous, portable, and cost-effective training methods/approaches for use in home and community settings represents a valuable direction for future stroke rehabilitation.

The biofeedback approach operates by delivering augmented visual, auditory, and/or tactile information; based on measurements of an individual’s biomechanical and/or physiological parameters, including body movements and plantar forces, cardiovascular function, and/or neurological activity; without providing any mechanical or motorized assistance and/or resistance to users [17]. In gait rehabilitation for stroke survivors, some wearable technologies have also been employed to provide real-time biofeedback targeting these specific parameters [18]. These include gait parameters such as stance time [19] and weight bearing adjustment [20]; balance and coordination performance [21]; as well as some functional mobility tasks, such as overground walking [22, 23] and transferring from chairs [24]. This approach may be adapted for daily activities and allows for deployment in both home-based [25] and laboratory settings [24]. In practice, the implementation of this biofeedback training strategy involves the use of wearable sensors such as inertial measurement units (IMUs) [26–28], pressure or force sensors [29–33], and electromyography (EMG) sensors [34]. A recent review reported that these wearable sensors can be attached to shoes, insoles, canes, clothing, gloves, watches, or directly to the skin surface, facilitating a seamless integration into ADL [18]. From the sensed data, some tailored biofeedback signals can be delivered in various modes (visual, auditory, and vibrotactile), content (focused on raw performance or processed results), frequency/intensity (constant, adjustable, or fading), and timing (concurrent or terminal) [35]. These biofeedback cues can be employed for either positive or negative reinforcements during gait training [36]. Specifically, positive reinforcement utilizes various feedback signals to encourage and increase the frequency/intensity of desired movement patterns, to improve the movement repetitions and consolidation [37]. Conversely, the negative reinforcement aims to reduce or even eliminate the undesired movement patterns or behaviors in terms of occurrence or frequency [38]. Both approaches aim to improve gait patterns and functional recovery in patients.

Several previous systematic reviews [18, 39–41] have indicated significant benefits of the real-time biofeedback and wearable devices in gait training for healthy individuals and those with neurological disorders. Two previous reviews conducted by Chamorro et al. on technology-based biofeedback for patients with abnormal gait [41] and by Gordt et al. on wearable sensor-based training for balance and gait of general health and patient individuals [40] have reported some positive results. However, neither of these reviews focused on the effects of such training for stroke survivors. A narrative review conducted by Spencer et al. recommended the biofeedback-based gait re-training as a promising intervention for post-stroke gait rehabilitation [18]. However, the evidence with higher level or strength, preferably based on some systematic/standardized qualitative and/or quantitative analyses, was lacking. The systematic review and meta-analysis by Bowman et al.—which included 34 studies and 1002 participants across a variety of neurological diseases—provided more robust, quantitative evidence [39]. Their analysis demonstrated that wearable biofeedback devices could generate significant improvements in gait performance as measured by the Timed Up and Go test (TUG) (MD = − 3.43 s; P = 0.03) [39]. While showing a positive trend in stroke survivors (115 stroke survivors from 4 studies), their subgroup analysis was not sufficient to draw firm conclusions for this specific population. The review authors attributed this limitation to “the current quality of the literature” and called for further higher-quality randomized controlled trials (RCTs) with larger sample sizes [39]. Among all the neurological impairments, post-stroke gait is uniquely characterized by asymmetrical patterns due to the unilateral motor impairment, which differs from the bilateral deficits typical of other conditions like Parkinson’s disease [6]. This has made it reasonable to anticipate that the biofeedback intervention outcomes from other neurological patient groups may be different and should not be generalized directly to stroke survivors. To guide the future clinical practice and research directions, a systematic synthesis of the current evidence is necessary to consolidate findings and assess the specific effects of biofeedback gait training for the stroke population.

To address the above-mentioned issues, this systematic review and meta-analysis aimed to synthesize evidence from RCTs on the effectiveness of wearable real-time biofeedback gait training devices for improving post-stroke rehabilitation outcomes, including gait, balance, functional mobility, and ADLs.

Methods

This systematic review and meta-analysis study was undertaken in accordance with the guidelines of PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analysis) statements [42], and was registered in International Prospective Register of Systematic Reviews (PROSPERO) (Reference number: CRD42024588926; https://www.crd.york.ac.uk/PROSPERO/).

Literature search and study identification

Electronic literature searches were conducted in databases of PubMed, EMBASE, MEDLINE, Web of Science, Cochrane Library, CINAHL, PsycInfo, PreQuest, and PEDro, from the inception to September 13th, 2024. The database searches were restricted to human beings and RCTs. The following keywords were used to search potential articles: stroke, wearable, biofeedback, gait, therapy, and rehabilitation. The keywords “stroke” and “hemiplegia”, along with related terms, were used to identify the target population; while “wearable”, “biofeedback”, and “gait”, among the others, were used to describe the interventions being studied. Additionally, terms like “therapy” and “rehabilitation” broaden the context of post-stroke recovery strategies. The full search strategy is available in Supplemental Appendix 1. The reference lists of identified articles were screened to further identify additional potential eligible papers. This involved manually reviewing the bibliographies of all included studies and relevant systematic reviews for any additional RCTs that met the inclusion and exclusion criteria.

Study selection

Reviewer 1 (F.-Y.W.) searched the nine databases and removed the duplicates. To identify eligible studies, reviewer 2 (Y.X.) and reviewer 3 (L.Y.-Y.L.) independently screened the titles, abstracts, and full text of all articles that were generated from the electronic database search. For any persistent disagreements between the two reviewers, reviewer 1 was consulted for the final decision.

The eligibility criteria were formulated according to the population, intervention, comparators, outcomes, and study design framework: (1) Population: stroke survivors; (2) Intervention: gait training utilizing wearable technologies that include sensing components (such as IMUs, pressure sensors, accelerometers, or other sensor-based technologies) and feedback cues (including but not limited to visual, auditory, or haptic feedback); (3) Comparators (or controls): traditional gait training without real-time biofeedback, including but not limited to the usual care or standard rehabilitation practices, and placebo or sham interventions; (4) Outcomes: gait parameters (e.g., gait speed, step length), balance performance (e.g., Berg Balance Scale, BBS), functional mobility (e.g., TUG, 6-Minute Walk Test), and scores of ADL (e.g., Barthel Index); and (5) Study design: RCTs. For the current review, “biofeedback” was defined as a technology that provided sensory feedback based on sensor’s data, without providing any motorized/mechanical assistance or resistance to the user. This excluded systems that mechanically/voluntarily interfered and/or influenced user’s movement, including robotic exoskeletons, to isolate and focus on the intervention effect from the biofeedback component only.

Data extraction

Reviewer 2 and reviewer 4 (H.-B.L.) independently extracted data from the included studies using a previously validated standard data extraction form, which was developed based on Cochrane guidelines [43]. Discrepancies were resolved through discussion with reviewer 1. The corresponding authors of the relevant studies were contacted via email when additional information was required. The following data from each included study were extracted: first author; year of publication; sample sizes, subject characteristics such as sex, age (mean and standard deviation), and post-stroke duration; biofeedback gait training intervention of the experimental group; intervention of the control group; outcome measures; and evaluation time-points.

Risk of bias and quality appraisal

The Cochrane Risk of Bias (RoB 2) tool was used to evaluate the risk of bias of the included studies [44]. This scale contained seven parts: sequence generation and allocation sequence concealment (selection bias), masking of participants, personnel and outcome assessment (detection bias and performance bias), incomplete outcome data (attrition bias), selective reporting (reporting bias), and other bias (e.g. funding assistance). Each part was evaluated to grade low, unclear, or high risk of bias. The risk of bias and applicability was analyzed using RevMan 5.4 (Cochrane Collaboration’s Information Management System).

The Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach [45] was applied to assess the confidence of the effect estimates on the basis of the following criteria: presence of study limitations, indirectness of evidence, inconsistency of results/unexplained heterogeneity, imprecision of results, and high probability of publication bias. The PEDro scale [46] scores were also extracted and reported. These scores, ranging from 0 to 10, represent a valid and reliable measure of an RCT study’s methodological quality. This process was independently performed by reviewer 2 and reviewer 3, with the participation of reviewer 1 when discrepancy occurred.

Data synthesis

The RevMan 5.4 was employed to perform the meta-analysis. A minimum of three studies that been judged clinically homogeneous enough, in terms of the comparator and outcome measures, were considered sufficient for pooling data for the meta-analysis. If the data reported in articles could not be used for data pooling, the authors of the articles were contacted to request the necessary data. For continuous outcome measures of the intervention, effects were evaluated using the Mean Difference (MD) or the Standardized Mean Difference (SMD) along with the 95% Confidence Interval (CI), of post-intervention values. The Q-test and the I² values were employed to rate the heterogeneity of the combined studies. The I² values of 25%, 50%, and 75% indicated low, moderate, and high heterogeneity, respectively. The random effects model was conducted to calculate the pooled effect size due to its ability to account for variability among studies [47]. The SMD values of larger than 0.8, ranging from 0.5 to 0.8, and ranging from 0.2 to 0.5 indicated the large, moderate, and small effect sizes, respectively [47]. Subgroup analyses were used to investigate the potential differential effects between the different experimental biofeedback training and the control interventions. The synthesized results were presented in forest plots.

Results

Study selection

A total of 2,192 studies were identified from the nine electric databases and manual searches from bibliography lists. After removing duplicates, 1,983 studies were screened for titles and abstracts. A total of 31 studies were eligible for full-text screening. Finally, 13 studies [32, 48–59] were included in the review. The PRISMA flowchart illustrates the study searching and selection process at each stage (Fig. 1). The list of the excluded studies after full-text checking is shown in Supplemental Appendix 2.

Fig. 1.

PRISMA flow diagram

Risk of bias and methodological quality appraisal

Details of RoB assessment are presented in Fig. 2. All 13 studies were graded as having high RoB for the lack of blinding of the participants. This was due to the nature of the exercise intervention, which could not be concealed. Three studies [50, 56, 59] were rated as high in RoB, due to the incomplete outcome data and without any imputation for missing data. Additionally, two studies [52, 59] were graded with a high RoB, for the lack of blinding of the outcome assessors. The reporting bias of all studies was assessed as unclear risk. One study [53] was assessed as having a high risk for other bias, due to the same reported age (mean ± SD) of the control group as that of another previous study [54]. The methodological quality of the RCTs, as assessed using the PEDro scale, ranged from fair to good, with scores ranging from 4 to 8 points (Appendix 3).

Fig. 2.

Review authors’ judgments about each risk of bias item presented as percentages across all included studies: A ROB summary; B ROB graph

Characteristics of the included studies

The characteristics of the included studies are summarized in Table 1. The 13 included studies were published between 1994 and 2021, including a total of 304 stroke survivors. The sample size of the included studies ranged from 12 to 45 participants. Subjects’ mean age ranged from 52.1 to 66.2 years in the intervention group, and 53.5 to 65.7 years in the control group. The duration of gait intervention ranged from 2 weeks to 2 months, and the frequency of exercise ranged from 2 times per week to 5–7 times per week.

Table 1.

The characteristics of included studies

Study	Sample size n(EG/CG) No. of F	Age of years (mean ± SD)	Time post-stoke (mean ± SD)	Intervention		Outcome measures	Evaluation time point
				EG	CG
Byl et al. 2015 [59]	12(5/7) 8	EG: 66.2 ± 5.0 CG: 60.8 ± 5.4	EG: 10.4 ± 7.8 CG: 6.6 ± 3.6 (years)	TP: as CG with feedback D: as CG WD: Smart shoes with pressure sensor AND wireless joint angle sensors FT: Visual kinematic feedback interpreted by the engineer and the therapist	TP: Progressive and task-oriented integrated gait training activities, exercises for postural alignment, balance, strengthening of hip abduction, coordination stretching (heel cord, hip, and hamstrings) and dual tasking D: 90 min/day, 6–8 weeks, 12 sessions	Gait Speed, Step Length, Tinetti, 6 min walk, Dynamic Gait, FTSTS, TUG, BBS, Strength, ROM	PRE, POST
Cha et al. 2018 [58]	31(21/10) 10	EG1: 64.6 ± 10.6 EG2: 63.0 ± 4.7 CG: 61.8 ± 9.8	EG1: 56.9 ± 27.0 EG1: 67.5 ± 45.6 CG: 78.6 ± 31.9 (months)	TP: as CG + feedback gait training D: as CG WD: Pressure Sensor located on the sole (EG1: in heel; EG2: in forefoot) FT: Auditory feedback during weight bearing	TP: 30 min comprehensive rehabilitation therapy + 20 min gait intervention D: 50 min/day, 3 times/week, 6 weeks, 18 sessions	Gait speed, FGA, TUG, COL path length EO/EC, COL path velocity EO/EC	PRE, POST
Choi et al. 2019 [57]	24(12/12) 8	EG: 62.8 ± 4.8 CG: 59.7 ± 10.2	EG: 67.2 ± 43.6 CG: 70.0 ± 31.9 (months)	TP: as CG + feedback gait training D: as CG WD: Pressure Sensor located on the sole, placed in the forefoot and hind foot FT: Auditory feedback during weight bearing	TP: 30 min comprehensive rehabilitation therapy + 20 min gait intervention D: 50 min/day, 3 times/week, 6 weeks, 18 sessions	Gait speed, FGA, TUG, COL path length EO/EC	PRE, POST
Intiso et al. 1994 [56]	16(8/8) 7	EG: 61.3 ± 12.3 CG: 53.5 ± 18.5	EG: 11.3 ± 12.6 CG: 8.3 ± 6.0 (months)	TP: PT without foot dorsiflexion exercise + feedback gait training D: as CG WD: EMG biofeedback device FT: Auditory EMG feedback of TA activation during the swing phase of the gait cycle	TP: PT, standard exercise bobath, facilitation and inhibition techniques, neurofacilitatory techniques, standard exercises for dorsiflexion of the foot D: 60 min/day, 2 months	Gait speed, Basmajian scale, MBI, Canadian Neurological Scale, Adams scale, Ashworth scale, Step length, Ankle angle (swing), Ankle angle (heel strike)	PRE, POST
Jonsdottir et al. 2010 [55]	20(10/10) NR	EG: 61.6 ± 13.1 CG: 62.6 ± 9.5	EG: 5.9 ± 10.5 CG: 1.8 ± 0.9 (years)	TP: Task-oriented gait training with feedback D: as CG WD: EMG biofeedback device FT: Auditory EMG feedback of gastrocnemius lateralis during gait	TP: Therapeutic approaches, like neurodevelopmental and neurofacilitation techniques, task specific training, at least 15 min of gait training in each session D: 45 min/day, 3 times/week, 20 sessions	Gait speed, Ankle power peak at push-off, Stride length, Knee flexion peak	PRE, POST, 6 weeks - FU
Jung et al. 2015 [54]	21(11/10) 7	EG: 56.4 ± 11.1 CG: 56.3 ± 17.1	EG: 6.2 ± 2.5 CG: 7.0 ± 2.5 (months)	TP: PT and OT + feedback gait training D: as CG WD: Cane with pressure sensor FT: Auditory feedback of the peak vertical force underwent gait	TP: PT and OT, gait training with a regular cane (without any auditory feedback) D: 30 min/day, 5 times/week, 4 weeks (gait training); 60 min/day, 5 times/week, 4 weeks (PT and OT)	Gait speed, Vertical peak force of the cane, Muscle activation (Gluteus medius, Vastus medialis oblique), Single support phase	PRE, POST
Jung et al. 2020 [53]	20(10/10) 6	EG: 57.1 ± 11.4 CG: 56.3 ± 17.1	EG: 6.3 ± 2.6 CG: 7.0 ± 2.5 (months)	TP: PT and OT + feedback gait training D: as CG WD: Cane with pressure sensor FT: Auditory feedback of the peak vertical force underwent gait	TP: PT and OT, conventional gait training D: 30 min, 5 times per week, 4 weeks (gait training); 30 min/day, 5 times/week, 4 weeks (PT and OT)	Peak vertical force on cane, Muscle activation, Trunk impairment scale (total/static/ dynamic/ coordination), TUG	PRE, POST
Ki et al. 2015 [52]	25(12/13) 6	EG: 55.3 ± 9.2 CG: 60.1 ± 12.3	EG: 19.1 ± 8.2 CG: 22.0 ± 9.9 (months)	TP: Neurodevelopmental treatment with feedback D: as CG WD: Pressure gauge PedAlertTM120 FT: Auditory feedback of subject’s weight bearing exceeding 50%	TP: Neurodevelopmental treatment D: NR about details, 4 weeks	Stance phase duration, Single limb stance duration, TUG	PRE, POST
Kim et al. 2020 [50]	24(12/12) 16	EG: 54.8 ± 5.1 CG: 55.0 ± 4.9	EG: 8.7 ± 2.0 CG: 8.8 ± 1.9 (months)	TP: PT as CG + feedback gait training (shuttle the start and endpoints of the path in forward and backward walking) D: as CG WD: Insole pressure sensors FT: Visual feedback of actual load on both sides	TP: PT (general conditioning exercise, mat exercise, and therapist-guided neuromuscular facilitation training), 5-m walking path for gait training D: 60 min/day, 6 weeks (PT); 30 min, 3 times/week, 6 weeks,18 sessions (gait training)	Gait speed, Step length, Stride length, Single support time, Double support time, Step length ratio, Stride length ratio, Single support time ratio, TUG	PRE, POST
Kim et al. 2021 [51]	45(23/22) 18	EG: 61.8 ± 8.1 CG: 64.2 ± 0.6	EG: 14.6 ± 7.2 CG: 16.9 ± 7.2 (months)	TP: PT and OT as CG + feedback gait training D: PT and OT as CG, 60 min, twice per week, 4 weeks for feedback gait training WD: Smart insole with pressure sensors FT: Auditory feedback of subject’s weight bearing exceeds the threshold	TP: PT (neurodevelopmental therapy, balance training, FES) and OT (upper limb function improvement training), general gait training D: 60 min/day, 5 times/week, 6 weeks (PT and OT); 30 min/day, 5 times/week, (general gait training)	Gait speed, Cadence, Stride time/length, Affected side step time/length, Unaffected side step time/length, USST, ASST, DLS, AGL, TUG, BBS, MBI	PRE, POST
Lupo et al. 2018 [49]	15(9/6) 4	EG: 52.6 ± 13.9 CG: 65.7 ± 9.6	EG: 42.7 ± 31.9 CG: 82.0 ± 67.8 (days)	TP: 6 different exercises with feedback, latero-lateral load shift/antero-posterior load shift/latero-lateral load with knee flexion on stable and oscillating platform D: as CG WD: Inertial measurement units placed on the chest by elastic bands and a force platform FT: Visual feedback of posture and foot force	TP: Conventional training, propaedeutic for the recovery of independence in ADL and based on the use of stable surfaces and unstable ones D: 20 min/day, 3 times/week, 10 sessions	BBS, Canadian Neurological Scales, MBI, Rivermead Mobility Index, National Institutes of Health Stroke Scale, COP EO/EC	PRE, POST, 1 month -FU, 50 days -FU
Owaki et al. 2021 [48]	16(8/8) 3	EG: 56.9 ± 10.8 CG: 56.6 ± 12.0	EG: 1767 ± 2428 CG: 1937 ± 1547 (days)	TP: PT and OT + Feedback gait training D: as CG WD: Prosthesis with pressure sensors FT: Auditory feedback during weight bearing	TP: PT and OT + General gait training D: 30 min/day, 2 weeks, 7 sessions	WBAM during one gait cycle	PRE, POST
Sungkarat et al. 2011 [32]	35(17/18) 11	EG: 52.1 ± 7.2 CG: 53.8 ± 11.2	EG: 3.9 ± 4.8 CG: 4.7 ± 5.8 (months)	TP: PT as CG + Feedback gait training D: as CG WD: Insole shoe wedge, time and pressure sensors FT: Somatosensory and auditory feedback during weight bearing	TP: PT (neuromuscular facilitation techniques, therapeutic exercises, balance and functional training) + General gait training D: 60 min/day (30 min for PT, 30 min for general gait training), 5 times/week, 15 sessions	Gait speed, Step length, Single support time, BBS, TUG, Loading on paretic leg during stance	PRE, POST

EG: experimental group; CG: control group; No. of F: number of females; SD: standard deviation; TP: training program; D: dosage; WD: wearable device; FT: feedback training; FTSTS: Five Times Sit to Stand(s); ROM: range of motion; FGA: functional gait assessment; COL: center of loading; EC: eyes closed; EO: eyes open; NR: not report; PT: physical therapy; EMG: Electromyographic; TA: tibialis anterior; FU: follow up; OT: occupational therapy; FES: functional electrical stimulation; USST: unaffected side sing limb support time; ASST: affected side sing limb support time; DLS: double limb support; AGL: affected side gait line; COP: center of pressure; WBAM: whole body angular momentum;

Three distinct types of detection methods—kinetic [32, 48, 50–54, 57, 58], EMG [55, 56], and kinematic [49, 59]—were employed in the included studies to enhance rehabilitation outcomes in stroke survivors. Kinetic biofeedback relied on force and weight distribution signals during gait, EMG-based biofeedback focused on muscle activation patterns, and kinematic biofeedback provided real-time information based on joint and limb movements. Various wearable devices were utilized to collect information during gait training in experimental groups, including the wireless joint angle sensors [59], EMG electrodes [55, 56], and pressure sensors located at the shoe sole [57, 58], insole [32, 50, 51], cane [53, 54], or prosthesis [48]. Pressure sensors were placed on the shoe sole in two included studies, which provided auditory feedback during weight-bearing gait training [57, 58]. Two studies used canes with pressure sensors to provide auditory feedback, when the sensor detected the peak vertical force during walking [53, 54]. Two studies incorporated insole pressure sensors to provide visual [50] and auditory [51] feedback on the actual load during gait training. One study also used the same type of sensors in an auditory feedback system [32]. For the EMG-based biofeedback systems, one study developed an auditory feedback system to improve the tibialis anterior muscle activation [56], whereas another study employed similar biofeedback to target the gastrocnemius lateralis muscle activity during walking [55]. Only two studies used a fading progression strategy in their intervention designs [55, 56]. Two studies employed pressure sensors combined with IMUs to provide visual biofeedback of kinematic information for patients [49, 59]. The control interventions were primarily conventional rehabilitation therapies, such as gait training, weight shifting exercises, and balance training.

The included studies used diverse outcome measurements. A wide range of gait parameters were adopted, including gait speed [32, 50, 51, 54–59], step length [32, 50, 51, 55, 56, 59], stride length [50, 51, 55], and dynamic gait indices [48–51, 57, 58]. Physiological parameters such as muscle strength [59], range of motion (ROM) [59], muscle tone [56], and ankle peak power during push-off [55] were also evaluated. Additionally, assessments for balance and functional mobility, such as the BBS [32, 49, 51, 59], 6-minute walking test [59], or TUG [32, 50–53, 57–59], were used. Furthermore, the Neurological Scales [49, 56] and Barthel Index [49, 51, 56] were used to comprehensively evaluate stroke recovery and ADL. All included studies incorporated pre-/baseline and post-intervention assessments, and two of them further reassessed the outcomes 4–6 weeks after the completion of intervention [49, 55].

Qualitative synthesis of included studies

Most of the included studies reported that compared to stroke survivors in the control groups, those undergone gait training with wearable biofeedback systems demonstrated significantly greater improvements in gait-related outcomes. These improvements were observed across various domains, including muscle activation, gait parameters, balance, and functional mobility. Specifically, four studies [53–56] showed significant improvements in muscle activation and strength. Ten studies [32, 50–55, 57–59] reported significant increases in gait speed, step length, and balance controls as assessed by the BBS and TUG. Eight studies [32, 48, 50–52, 54, 57, 58] observed significant improvements in walking symmetry. Moreover, three studies [57–59] demonstrated that real-time biofeedback enabled users to immediately recognize and correct gait deviations, leading to improved outcomes.

The effectiveness of biofeedback approaches in stroke rehabilitation is supported by findings from 13 studies that employed various sensing components and feedback formats. For the sensing components, one study [59] demonstrated that smart shoes equipped with pressure and wireless joint angle sensors improved mobility, balance, and strength during gait training. Similarly, two studies [57, 58] utilized pressure sensors on the sole, which contributed to the enhancements in walking performance and functional gait assessments. Additionally, while the application of IMUs significantly improved balance skills [49], the EMG-based biofeedback system significantly increased muscle strength [56], foot-drop recovery [56], peak ankle power [55], and gait velocity [55]. Two studies [53, 54] reported significant improvements in stroke survivors’ muscle activation and gait speed when using canes with pressure sensors. For the feedback formats, two studies highlighted that real-time visual [50] and auditory [51] feedback, based on insole sensor measurements, led to significant improvements in various gait parameters. Additionally, one study [52] showed that auditory feedback during gait training increased stance phase duration at the paretic side, while another study [48] revealed that the auditory feedback promoted walking balance, as evidenced by a reduction in the range of the frontal whole body angular momentum in the experimental group. Lastly, the combination of somatosensory and auditory feedback significantly improved gait speed and balance control [32]. These findings illustrate the diverse application and strategies of various sensing components and feedback cues in stroke rehabilitation, and their potential contributions to gait recovery.

Meta-analyses

Meta-analyses were conducted on four outcome measures: gait speed, TUG, BBS, and Modified Barthel Index (MBI).

Gait speed

Seven included studies [32, 50, 51, 54, 56–58] measuring gait speed were included in a meta-analysis, involving a total of 204 stroke participants. The meta-analysis revealed that the gait speed was significantly faster in the biofeedback gait training group than that of the control group (SMD = 0.41; 95% CI = 0.006 to 0.77; P = 0.02; I² = 34%; n = 204) (Fig. 3). Of the seven studies, one study utilized pressure sensors with visual cues [50], one study used EMG for sensing with auditory cues [56], while the remaining studies adopted pressure sensors with auditory cues [32, 51, 54, 57, 58]. The subgroup analysis was conducted to assess the efficacy of the pressure sensing technology with auditory feedback [32, 52, 55, 58, 59], and the pooled results did not reveal statistically significant between-group differences in gait speed (SMD = 0.30, 95% CI: -0.01 to 0.61, P = 0.05; I² = 0%, n = 166). Notably, the study utilizing pressure sensors with visual cues demonstrated a significant improvement compared to the control group (SMD = 1.63, 95% CI: 0.68 to 2.57, P = 0.0007, n = 24) [51], while the study using EMG electrodes with auditory cues showed no statistically significant difference between the groups (SMD = 0.00, 95% CI: -1.06 to 1.06, P = 1.00, n = 14) [57].

Fig. 3.

Forest plot of biofeedback gait training (BGT) versus control intervention for gait speed

TUG

The meta-analysis included six studies [32, 51–53, 57, 58] involving 190 participants to evaluate the effectiveness of biofeedback in improving the TUG test results. The results demonstrated that the functional mobility of the participants in the experimental biofeedback group was significantly better than that of the control group (SMD=-0.36; 95%CI=-0.65 to -0.08; P = 0.01; I² = 0%; n = 190) (Fig. 4(A)).

Fig. 4.

Forest plot of biofeedback gait training (BGT) versus control intervention for functional abilities: A TUG as measurement; B BBS as measurement; C MBI as measurement

BBS

This meta-analysis included three studies [32, 49, 51] to evaluate the effect of biofeedback on BBS, involving 95 participants. The results showed that the balance ability of participants in the biofeedback group was significantly better than that of the control group (SMD = 0.44; 95%CI = 0.03 to 0.85; P = 0.03; I² = 0%; n = 95) (Fig. 4(B)).

MBI

Three studies [49, 51, 56] were included in the meta-analysis to determine the effectiveness of biofeedback in improving MBI scores, involving 74 participants. The results indicated no statistically significant differences in MBI scores between the two participant groups (SMD = 0.21; 95%CI=-0.25 to 0.67; P = 0.38; I² = 0%; n = 74) (Fig. 4(C)).

Certainty of evidence

As outlined in Supplemental Appendix 4, the quality of the evidence was evaluated using the GRADE system [60]. There was moderate-quality evidence to support that the real-time biofeedback gait training has moderate effects on BBS and small effects on TUG, and low-quality evidence to support its moderate effect on gait speed. The certainty of evidence was downgraded primarily due to inconsistency and imprecision. Only for gait speed (using visual and auditory feedback) outcome that, the evidence showed moderate statistical heterogeneity (I²=34%), and the direction of effect was consistent across studies. For the rest outcomes, no significant heterogeneity was observed (I²=0%). Therefore, this evidence was not downgraded for inconsistency. The imprecision mainly resulted from small sample sizes across different analyses (gait speed: n = 204; pressure-sensor subgroup: n = 166; BBS: n = 95; TUG: n = 190; MBI: n = 74) and the wide 95% CI in gait speed (pressure sensor-auditory feedback) and ADL analysis.

Discussion

The present systematic review and meta-analysis of 13 RCTs covering 304 stroke survivors comprehensively summarized the evidence regarding the effectiveness of wearable real-time biofeedback gait training on gait patterns, balance performance, functional mobility and ADL, as compared to the traditional rehabilitation approaches. Moderate-quality evidence suggested that wearable biofeedback systems utilizing pressure sensors could significantly further improve the gait speed, while providing auditory biofeedback based on such measurements did not significantly improve gait speed, when comparing with traditional rehabilitation approaches. There was moderate-quality evidence suggesting that wearable biofeedback system had a small effect on improving TUG or BBS results, and was not significantly better than conventional rehabilitation in improving MBI scores of stroke survivors. More details can be found below.

Methodological quality of the included studies

In terms of the RoB assessment results, all included studies exhibited a high risk of bias due to the lack of blinding of participants or instructors. This overarching limitation across the evidence base should be considered when interpreting the reported positive meta-analysis findings for outcomes such as balance. The observed effects could potentially be inflated by participant’s expectations or therapist’s enthusiasm. Addressing this limitation in future research would require the adoption of sham condition/devices to control placebo effects [61]. Additionally, some included studies [28, 44, 46, 48–50, 52, 53, 55] failed to satisfy the criteria for intention-to-treat analysis. To avoid the potential inaccuracies and misleading interpretations, it is suggested that when participants withdraw or discontinue the intervention, future studies should incorporate the intention-to-treat analysis to facilitate an unbiased and reliable assessment of treatment efficacy [44].

Training paradigm: sensor type and configurations

Kinetic measurements (using pressure sensors) were utilized more frequently than kinematic (using IMUs) and physiological (using EMG) measurements in the reviewed studies. Most of these studies [32, 48–51, 53, 54, 57–59] reported improvements in gait symmetry by increasing the weight bearing of the paretic leg in a gait cycle. Interestingly, the positioning of the pressure sensor does not seem to significantly affect the effectiveness of these interventions. These sensors could be easily installed to accurately measure ground reaction forces, and facilitate achieving real-time feedback with less data processing time and efforts [62]. For the less commonly used kinematic approaches, IMUs were used to capture three-dimensional kinematic data of body segments, to provide insights into user’s postural sway and movement patterns [63, 64]. Only one out of the 13 studies implemented IMU-based biofeedback and observed significant balance improvement on the BBS [49]. Such improvement aligns with findings from other studies on patients with Parkinson’s disease [65–67]. Two included studies utilized EMG to quantify the activation states of specific muscles [55, 56]. This approach could allow for a targeted strategy, in which personalized EMG-based biofeedback can be used to address specific muscle weaknesses in stroke survivors. Jonsdottir et al. [55] demonstrated that targeting gastrocnemius lateralis enhanced peak ankle push-off power, which is crucial for maintaining proper gait dynamics. The improved push-off power can increase propulsion, contributing to faster gait and reduced energy expenditure [68]. Additionally, Intiso et al. [56] highlighted the need of focusing on the tibialis anterior to address foot drops, which can enhance ankle dorsiflexion and reduce the risk of trips and falls that are associated with foot dragging [69]. Collectively, the evidence indicates complementary roles of different sensing modalities in post-stroke biofeedback rehabilitation: kinetic measurements are primarily leveraged to modulate gait symmetry via feedback on limb loading, whereas kinematic and electromyographic data are utilized to address deficits in postural balance and local muscle activation, respectively.

Training paradigm: feedback cues and implementation

Regarding feedback cues and their implementation, two notable issues emerge from the current evidence. First, the evidence is predominantly based on studies employing single-modality approaches, with auditory feedback being the most frequently utilized. Visual cues were less commonly investigated. Visual feedback’s primary advantage lies in its simplicity and directness, making it more comprehensible than the other feedback formats [70]. However, it may potentially disrupt patients’ balance performance by interfering with their visual sensory input, which is one of the three sensory inputs (i.e., visual, vestibular, and proprioceptive) for maintaining balance [71]. Conversely, auditory feedback provides real-time information that can be easily integrated, without affecting the patient’s visual, vestibular or proprioceptive sensory inputs. However, concerning the fact that there has been a lack of comparisons between visual and auditory feedback, the available evidence is insufficient to guide future practice regarding the selection of these two feedback modalities. The observed preference for auditory cues in existing literature may be partially attributable to considerations of technical implementation, and the theoretical concern that visual feedback may interfere with balance control, though the exact reasons have remained unclear. Second, the principle of fading feedback—a strategy grounded in motor learning theory to promote skill retention and transfer [72]—was less frequently applied and poorly reported. Only two studies [55, 56] implemented a systematic fading strategy, despite the theoretical models suggesting its importance for facilitating the transition from associative to autonomous motor control [73, 74]. Consequently, there was inadequate evidence to determine the optimal cue selection, and most of the existing feedback applications did not adhere to the well-established motor learning principles in terms of the progressive withdrawal/fading feature.

Training effectiveness: gait speed

The pooled results revealed low to moderate quality evidence that the wearable real-time biofeedback gait training devices had small to moderate effects on improving gait speed. This finding is consistent with previous research [18, 40], which has highlighted the efficacy of biofeedback interventions in enhancing gait performance among stroke survivors. While an overall positive trend was observed, subgroup analyses based on sensor (pressure sensors) and feedback cue (auditory signals) type indicated that the specific intervention modalities did not yield a statistically significant improvement in gait speed. The sole exception was one study which utilized pressure sensors with visual feedback and demonstrated a significant effect [50], and contributed substantially to the overall pooled result. This suggests that the overall trend may be driven by specific feedback configurations rather than a universal effect. This discrepancy highlights the critical role of cue modality. The notable positive effect of visual feedback may be attributed to its continuous and spatially aligned representation of load distribution [50], potentially providing a more intuitive reference for correcting posture than that of the temporally based auditory cues. The enhancement of gait symmetry on weight bearing may lead to an indirect boost in gait speed, particularly for those associated with reduced energy expenditure as reported in previous studies [75, 76]. However, this observation must be interpreted with caution, as several studies [32, 52, 58] employing auditory feedback also reported improvements in symmetry. Unfortunately, the current research was unable to conduct a meta-analysis on gait symmetry due to the varied measurements that were used across different studies.

Training effectiveness: balance and functional mobility

This review identified moderate-quality evidence for a small to moderate significant benefit of biofeedback training on balance, as measured by the BBS and TUG. This finding aligns with the previous study that was conducted by Bowman et al. [39]. However, these findings must be interpreted with caution, due to some limitations in the available evidence. This observed benefit may be attributable to principles of motor learning theory. Biofeedback provides augmented real-time information on performance, enabling error correction and the reinforcement of appropriate movement patterns, which support the retraining of motor skills, including coordination and postural control [69, 77]. However, the number of studies included in the analysis for these outcomes (BBS and TUG) was relatively small, and the overall sample sizes were modest, which may limit the statistical power and generalizability of the results. Furthermore, the low statistical heterogeneity observed in the BBS and TUG meta-analyses should be interpreted cautiously. It may not robustly demonstrate a consistent treatment effect across all potential clinical scenarios. This apparent consistency could partly be a consequence of the limited number of studies and their methodological similarities, rather than definitive evidence of a uniform effect size.

Training effectiveness: ADL

Moderate-quality evidence suggested that the wearable biofeedback gait training was not significantly different from conventional rehabilitative gait training on ADL, based on MBI scores. It is important to note that the MBI is a comprehensive measure of daily living activities and does not measure gait specifically. It may not be sensitive to the changes in a specific domain of gait or balance performance. Additionally, the number of included studies in this meta-analysis was relatively small, which may have limited the statistical power to detect significant differences in MBI scores.

Recommendations for future clinical practice

Based on the findings of this systematic review and meta-analysis, clinicians could consider utilizing wearable biofeedback devices during the balance and gait training for stroke survivors in future clinical practice. Regarding sensor and modality selection, findings suggest that for improvement of gait speed, biofeedback systems utilizing pressure sensors with visual feedback may be particularly beneficial, as this configuration demonstrated a significant effect, while interventions using auditory cues did not show similar efficacy. For retraining weight-bearing symmetry, both auditory and visual cues appear promising. There are also many options for sensor placement, including under the paretic/nonparetic feet with less/more weight bearing, or at the bottom of walking aids. However, the current evidence does not permit definitive recommendations regarding the optimal placement of sensors, due to the lack of comparative studies. The positive effects observed across the diverse placements support flexibility in clinical practice. While none of the included studies have reported any incidents of patients experiencing falls or near-falls during biofeedback gait training, it is worthwhile noting that adequate protective measures are necessary in clinical practice. This is especially important when conducting training with stroke survivors, who exhibit poor balance control. It is also suggested to consider the cost-effectiveness and necessity of implementing various wearable devices in clinical practice. This could facilitate evidence-based decision-making, especially if the device is very expensive. However, as most of the reviewed initial efficacy RCTs (which are the focus of this study) are typically not designed to answer the health economic questions, the cost-effectiveness and necessity of implementing expensive biofeedback training devices has remained as a critical area that warrants future research and efforts.

Future research directions

Despite the advancements in biofeedback gait training, several critical areas have remained underexplored, warranting further investigation to enhance future clinical practice. Firstly, while providing fading feedback is a recognized motor learning principle, it was implemented in only two of the included studies. Future trials should specifically examine how to optimally structure the fading protocols—e.g., by comparing different schedules for reducing feedback frequency or intensity—and determine the appropriate timing for transitioning from continuous to intermittent feedback during stroke rehabilitation. Secondly, unlike biofeedback-based balance training [78], the real-time biofeedback gait training lacks RCTs that directly compare the different training methods or biofeedback strategies. To ascertain the superiority of one biofeedback method over the other, future studies should be meticulously crafted to directly compare a range of the sensed or measured parameters, the sensor types and integrations, the feedback formats, content, frequency, timing, and its application in either a positive or negative manner, among individuals exhibiting diverse post-stroke sensorimotor impairments. Thirdly, future research should incorporate control groups using “sham” biofeedback devices—which are applied to users but provide either no feedback or just random, non-contingent feedback—to rigorously control for placebo effects, and more precisely isolate the specific therapeutic benefits of the active biofeedback intervention. Fourthly, given the small effect size of observed balance improvement or TUG test results, it is also necessary to conduct some cost-effectiveness and implementation trials, to evaluate whether the addition of such biofeedback devices really is worth the expense clinically. Fifthly, in addition to the visual and auditory feedback, some other feedback approaches, such as tactile and a combination of different feedback modalities, may be utilized to the future wearable real-time biofeedback gait training devices. Some previous crossover studies have observed positive training outcomes of vibrotactile biofeedback training devices in stroke survivors [22]. However, there has still been a RCT validating the effect of such devices with the traditional rehabilitation therapy. Similarly, in addition to pressure sensors, IMUs, and EMGs, future research may also identify if some other wearable sensing components can be utilized into the biofeedback gait training devices. Examples include but are not limited to the wearable ultrasound imaging that enable the stroke survivors to visualize their paretic ankle muscle contraction dynamically [79]. These attempts and developments might help further improve the treatment efficacy of the wearable real-time biofeedback gait training devices in stroke survivors.

Study limitations

This systematic review and meta-analysis has several limitations that should be acknowledged. Firstly, all the included studies had small sample sizes, with the largest sample size of 46 participants, and variable methodological quality, which may impact the generalizability and reliability of our findings. Secondly, there is significant heterogeneity among the included studies in terms of intervention duration, frequency, and participant characteristics, which complicates the comparison and interpretation of the results and potentially explains the low quality of evidence.

Conclusion

The present systematic review and meta-analysis revealed low to moderate quality evidence that the wearable real-time biofeedback gait training devices had small to moderate effects on improving gait speed, balance, and functional mobility of stroke survivors. However, moderate-quality evidence suggested that such training approach was not significantly different from conventional rehabilitative gait training on ADL based on MBI scores. The larger scale well-designed RCTs with longer follow-up time durations are still needed, to determine the efficacy of such wearable biofeedback gait training approaches in stroke rehabilitation. Future research should also consider and optimize the specific characteristics of biofeedback interventions, including but not limited to the type of feedback cue and the duration and frequency of gait training sessions, to optimize the effectiveness of this novel rehabilitation approach. It is also worthwhile to conduct cost-effectiveness and implementation studies, to investigate the effectiveness of such devices in real-world clinical settings in future studies and facilitate evidence-based clinical decision-making.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ADL

Activities of Daily Living

BBS

Berg Balance Scale

BGT

Biofeedback Gait Training

DALYs

Disability-Adjusted Life Years

EMG

Electromyography

GRADE

The Grading of Recommendations Assessment, Development, and Evaluation approach

IMU

Inertial Measurement Units

MBI

Modified Barthel Index

PRISMA

The Preferred Reporting Items for Systematic Reviews and Meta Analysis

PROSPERO

The International Prospective Register of Systematic Reviews

RAS

Rhythmic Auditory Stimulation

ROM

Range of Motion

RoB

Risk of Bias

TUG

Timed Up and Go test

VR

Virtual Reality

WSO

World Stroke Organization

Author contributions

F.-Y.W. contributed to study screening, data extraction, quality assessment, data analysis, data interpretation, manuscript drafting and revision. Y.X., L.Y.-Y.L., H.-B.L., Y.-P.J., Z.-Q.B. contributed to study screening, data extraction, quality assessment, and manuscript revision. M.-Z.H., A.Y.-L.W., L.Y., M.Z., and Y.-H.Y. contributed to data analysis, data interpretation, and manuscript revision. C.Z.-H.M. contributed to study conception, study design, data interpretation, funding acquisition, and manuscript revision. All authors reviewed the manuscript.

Funding

This work was supported by the Hong Kong Research Grants Council (RGC) - Early Career Scheme (ECS) [grant number 25100523]; Chinese Association of Rehabilitation Medicine - Science and Technology Development Project [grant number KFKY-2023-050]; Research Institute for Smart Ageing (RISA), The Hong Kong Polytechnic University [grant number P0050739]; and the Science & Technology Department of Sichuan Province, China [grant number 2024NSFSC0539].

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

All authors have approved this manuscript for publication. This manuscript has not previously been published and is nor pending publication elsewhere.

Competing interests

The authors declare no competing interests.