Wearable Devices for Biofeedback Rehabilitation: A Systematic Review and Meta-Analysis to Design Application Rules and Estimate the Effectiveness on Balance and Gait Outcomes in Neurological Diseases

Wearable devices are used in rehabilitation to provide biofeedback about biomechanical or physiological body parameters to improve outcomes in people with neurological diseases. This is a promising approach that influences motor learning and patients’ engagement. Nevertheless, it is not yet clear what the most commonly used sensor configurations are, and it is also not clear which biofeedback components are used for which pathology. To explore these aspects and estimate the effectiveness of wearable device biofeedback rehabilitation on balance and gait, we conducted a systematic review by electronic search on MEDLINE, PubMed, Web of Science, PEDro, and the Cochrane CENTRAL from inception to January 2020. Nineteen randomized controlled trials were included (Parkinson’s n = 6; stroke n = 13; mild cognitive impairment n = 1). Wearable devices mostly provided real-time biofeedback during exercise, using biomechanical sensors and a positive reinforcement feedback strategy through auditory or visual modes. Some notable points that could be improved were identified in the included studies; these were helpful in providing practical design rules to maximize the prospective of wearable device biofeedback rehabilitation. Due to the current quality of the literature, it was not possible to achieve firm conclusions about the effectiveness of wearable device biofeedback rehabilitation. However, wearable device biofeedback rehabilitation seems to provide positive effects on dynamic balance and gait for PwND, but higher-quality RCTs with larger sample sizes are needed for stronger conclusions.


Introduction
People with neurological diseases (PwND) show mobility disorders including balance and gait deficits leading to a slowing of gait speed and an increased risk of fall [1][2][3] which consequently impact the quality of life [4][5][6].
Gait and balance characteristics of patients with Parkinson's disease (PD) or following a stroke have been largely documented. In patients with PD, the progressive loss of dopamine in the basal ganglia can lead to gait patterns such as festination, shuffling steps, and freezing of gait [7]. Poststroke hemiplegic gait is characterized by asymmetry and reduced weight-bearing on the paretic limb, as a result of the residual functions after upper Finally, a previous review of randomized controlled trials (RCT) [34] showed a positive effect of WDBR on gait and balance outcomes for older adults and patients. Although these results were promising, a greater number of RCTs were needed and have been included in this review as the most robust study design to estimate the feasibility and effectiveness of WDBR in clinical practice [34].
Thus, the objective of this systematic review of RCTs is twofold: First, to analyse the state-of-the-art technology on WDBR and present an overview of wearable devices in terms of their sensor configurations, biofeedback components, and training paradigms to provide practical design rules. Second, to evaluate the feasibility, usability, and estimated effectiveness of WDBR on balance and gait outcomes in PwND.

Data Sources and Searches
The review protocol was registered in PROSPERO (CRD42020162957) and reported according to the PRISMA statement [35]. The following databases were searched: MEDLINE, PubMed, Web of Science, PEDro, and the Cochrane Central Register of Controlled Trials. Additional relevant papers were found by hand in trial registers or other grey literature sources. Studies published in the English language from inception to January 2020 were included. Relevant search terms were combined with boolean operators (OR/AND) as reported in Table 1 and Table S2.

Eligibility Criteria
Eligibility criteria according to the PICOs were: (P) Participants: adults with any neurological diseases or disorders; (I) Intervention: wearable devices biofeedback rehabilitation for balance or gait; (C) Comparators: balance and gait rehabilitation without wearable devices or biofeedback, conventional rehabilitation, or usual care, no training; (O) Outcomes: balance and gait; (s) Study design: randomized controlled trials (RCTs).
To have been included in the systematic review, studies had to be focused on adults (age >18 years) with any type of central neurological disease or disorder. Moreover, studies should have focused on rehabilitation intervention using wearable devices and biofeedback principles. In addition, only studies that evaluated walking, ambulation, postural balance, equilibrium, motor activity, and recovery of function using spatiotemporal gait parameters, instrumental indexes, and clinical scales have been included.
Studies were excluded from this review if they did not use wearable devices or did not provide biofeedback. Moreover, studies not written in English, with relevant missing information, or that were out of topic with respect to the aims of the present study were excluded. Finally, we excluded all study designs except for RCTs.

Study Selection and Data Extraction
The studies retrieved using the search strategy and additional sources were screened independently by two main reviewers (TB and EG), based on their titles and abstracts, and considering inclusion and exclusion criteria. Then, the full text of eligible studies was further analysed and independently assessed for eligibility. Any disagreement between the two reviewers over the eligibility of studies was resolved through discussion with a third reviewer (DC).
After inclusion, the study characteristics, research goals, and main findings were extracted and summarized. Extracted information also included: study setting, study population, participant characteristics, wearable device, biofeedback characteristics, details of the intervention and control conditions, study methodology, outcomes and times of measurement, indicators of acceptability, and feasibility.

Data Synthesis
To evaluate the effects of WDBR on balance and gait outcomes, we merged in a meta-analysis study with common characteristics of intervention. Firstly, we considered if wearable devices had been used as "add-on" therapy, or "not add-on", in the experimental group compared to the control group. In studies where devices had been applied as "add-on", both groups performed the same exercises for balance and gait with the only difference being the addition of the device providing biofeedback information in the experimental group. In all other cases, wearable devices have been classified as "not add-on". Secondly, studies have been grouped according to the key components of balance rehabilitation exercises proposed by Horak et al. [36]. The following 6 components were considered: (1) biomechanical constraints, defined as the size, quality of the base of support and any limitations in strength, range of motion, and pain of the feet; (2) movement strategies, defined as the strategies used to return the body to equilibrium in a standing position; (3) sensory strategies, defined as the sensory information from somatosensory, visual, and vestibular systems that are integrated to interpret complex sensory environment; (4) orientation in space, defined as the ability to orient the body parts with respect to gravity; (5) control of dynamics (gait and proactive balance), defined as the control of balance during gait and while changing from one posture to another; and (6) cognitive processing, defined as the cognitive resources required during exercises and postural control.
Studies using the wearable devices as "add-on" therapy and exercises with similar rehabilitation components were pooled into the meta-analysis. First, we determined the overall effect of WDBR versus control intervention on different balance and gait outcomes at postassessment and follow-up (FU) in PwND. Second, to assess the effectiveness of different wearable devices on specific patient's populations, a subgroup analysis has been performed. Studies were grouped according to the wearable sensor type (e.g., IMU sensors and pressure sensors) and the neurological disease (e.g., stroke and Parkinson's).
All meta-analyses have been performed using random effects model and calculating the mean differences (MD) and 95% of confidence interval (CI) to acknowledge the methodological and clinical differences among studies [37]. Heterogeneity of the studies was assessed using the inconsistency test (I 2 ), whose values could be interpreted as follows: from 0% to 40% low heterogeneity; from 30% to 60% may represent moderate heterogeneity; from 50% to 90% may represent substantial heterogeneity; and from 75% to 100%: considerable heterogeneity [38].
Meta-analyses were calculated using Review Manager 5.3 (the Nordic Cochrane Centre, Copenaghen, Denmark). Alpha level was set at 0.05 to test for overall effect.
All studies not included in the meta-analysis have been considered in a qualitative synthesis of the results (see Supplementary Materials).

Risk of Bias Assessment
The risk of bias for all included RCTs was assessed with the six domains defined by the Cochrane Collaboration's tool [39]. These six domains are: (1) selection bias, due to random sequence generation and allocation concealment; (2) performance bias, with blinding of participants and personnel as a possible source of bias; (3) detection bias, due to blinding of outcome assessment; (4) attrition bias, evaluating possible incomplete outcome data; (5) reporting bias, due to selective outcome reporting; and (6) other bias, evaluating any important concerns about bias not covered in the other domains. Each domain was judged as "low risk of bias" ("green"), "high risk of bias" ("red"), or "unclear risk of bias" ("yellow").

Wearable Device Biofeedback Rehabilitation: Sensors Classification and Configuration
Based on the classification by Giggins et al. [14], we found 13 studies providing biomechanical measurements and six studies providing physiological measurements of the body system, using wearable devices based on different type of sensors ( Figure 2). Considering biomechanical measurements, five studies [43,46,48,52,53] used pressure sensors and one study [57] combined a pressure sensor with a foot switch. Pressure sensors were positioned under the foot of the paretic leg and were activated considering the % body weight loading [53,57] or by the combination of body weight loading and specific gait cycle phase ( Figure 3) [46,48]. In Jung et al. [52], pressure sensors were embedded into a cane used on the nonparetic side and activated by body weight %. Only El Tamawy et al. [43] used pressure sensors in both shoes to detect the push-off phase during gait.
Conversely, only three studies [42,44,58] used IMU sensors alone, while Byl et al. [41] combined IMU sensors with wearable pressure sensors, and the other two studies [45,55] combined IMU sensors with a force platform. IMU sensors were always positioned in the lower limbs, with different configurations involving thighs, shanks, and shoes. In the study by Carpinella et al. [42] IMU sensors were positioned on the upper trunk and the lower trunk, as well as on the lower limbs. Similarly, Schwenk et al. [58] placed IMU sensors on the lower trunk and on the lower limbs. In Van der Heuvel et al. [45], IMU sensors were positioned on the trunk and combined with the force platform, while Lupo et al. [55] placed sensors on the trunk, midthigh level and at midtibial level of the affected or the healthy side, depending on the exercise. Finally, only Cho et al. [47] used an electrogoniometer in the lower limb to measure the knee angle.
Considering physiological measurements, two studies [40,54] used EEG sensors and four studies [49][50][51]56] used EMG sensors. In both studies using EEG, electrodes were placed in compliance with the International 10-20 System. Lee et al. [54] secured an electrode to the scalp over the region of the central lobe (Cz), while Azerpaikan et al. [40] attached two electrodes to the left and right occipital (O1, O2) and one to the subject's left earlobe.
EMG signals were recorded with surface electrodes placed over the tibialis anterior muscle belly of the affected leg [49,50] or over the gastrocnemius lateralis muscles according to Seniam guidelines [49,51]. In one study, EMG sensors recorded the gastrocnemius and pretibial muscles activity in combination with an electrogoniometer inserted in the subjects' shoes to measure ankle movements [56].

Modalities of Exercise Interventions
The duration of the whole rehabilitation period ranged from a minimum of 10 sessions to a maximum of 30 sessions over 2 months. Single session dedicated time varied from 20 to 90 mins of training; in only two studies, treatment's time was not specified [53,56]. With respect to treatment frequency, the majority of the studies reported a frequency ranging from 2/week to 5/week, and in only two studies information was not reported [41,53].
Characteristics of the balance and gait training components of the biofeedback signal were different depending on the studies.

Components of Balance and Gait Training
According to the framework of Horak et al. [36], most of the studies (16 out of 19) involved the control of dynamics during stepping, body weight shift, changing from one posture to another, and walking exercises. Sensory strategies (e.g., balance activities on stable, unstable, and moving surfaces with eyes closed and eyes open) have been practiced in two studies [41,42]. Movement strategies (e.g., practicing balance with feet together) were applied only in the study by Byl et al. [41]. Two studies [41,43] took into account exercises aimed at improving orientation in space by standing in an upright position in front of a mirror to maintain a good postural alignment; moreover, eight studies [41,43,45,47,49,50,56,57] spent treatment's time to improve biomechanical constraints (e.g., increasing muscles strength and training limits of stability). Noteworthy, all the studies involved cognitive processing since cognitive resources are essential during exercises and for the postural control required for balance and gait training and to elaborate information provided by the feedback itself [59].
Treatment progression was described in 6 out of 19 studies, and in some of these studies, the physiotherapist progressively adjusted training complexity by changing the reference values of the task or exercise, according to the ability of each patient [41,42,45,51].
In the study by El-Tamawy et al. [43], the treadmill walking time was increased gradually from 6 to 25 mins, and walking speed progression was self-selected by each subject, while Intiso et al. [50] described two progressive phases of the training increasing the needed threshold to provide biofeedback when the patient made errors of less than 20% during the session.

Biofeedback Components
In Table 3, we reported biofeedback components as mode, content, frequency, and timing. Moreover, we have considered if the authors provided any explanation about the progression of training in line with motor-learning principles. Regarding timing, all the devices but one [41] provided concurrent, or 'real-time', biofeedback. On the other hand, Byl et al. [41] provided terminal biofeedback while the subject was standing or sitting after performing a few walking trials. In the same way in Carpinella et al. [42], a terminal feedback rating performance was given to the patient after each exercise. Biofeedback content was related to knowledge of the performance (e.g., information about movement coordination or muscle activity during the movement) in all the studies included. In addition, studies [41,42,45,55] also provided knowledge of results, presenting a scoring point or a number representing the outcome of the performance. In eight of the studies [44,46,48,[50][51][52][53]57], the modality exploited to transmit the signal to users' devices was auditory. Four studies [40,41,45,47,54] used visual signal, while in the other five studies [42,49,55,56,58] a combination of auditory and visual biofeedback was applied. Only El-Tamawy et al. [43] provided vibrotactile biofeedback.
According to principles of motor learning, a fading frequency of the biofeedback signal was provided in two studies [42,44]. Jonsdottir et al. [51] provided a constant signal in the first phase of the training and a fading signal in the second phase. The remaining 17 studies provided a constant frequency (e.g., signal provided every time a biomechanical or physiological variable reached a predefined threshold).
Considering the type of reinforcement, in 12 out of 19 studies, a positive reinforcement signal was given when the variable measured by the device remained within a pre-established therapeutic window or reached a predetermined threshold [40,43,45,46,[48][49][50][51][53][54][55][56][57]. Instead, in the study by Jung et al. [52], negative reinforcement has been provided since stroke patients were instructed to avoid activating the beeping sound from the cane with the aim of increasing load in the paretic limb. In four studies [41,42,44,58], biofeedback signal was positive or negative depending on the task or the activity performed, and in two studies [47,55], information about the type of reinforcement was not specified.
3.6. WDBR Estimated Effectiveness on Balance and Gait Outcomes 3.6.1. Berg Balance Scale (BBS) Three studies [42,55,57] compared the effects of add-on WDBR training on BBS with controls. Our meta-analysis revealed a significant overall effect in favour of add-on WDBR at postassessment (MD = 4.99 [1.79, 8.18]; p = 0.002; Figure 4). The same overall effect was maintained in studies with FU assessment (MD = 5.29 [0.06, 10.51]; p = 0.05; Figure 4). Homogeneity criteria were met for all the analyses (I 2 = 0%; Chi 2 = 0.30, p = 0.84). A qualitative synthesis for studies not included in the meta-analysis (see Supplementary Materials) showed no significant differences between groups in two studies [41,45], instead Azerpaikan et al. [40] found significant differences (p < 0.01) in favour of neurofeedback provided by EEG device.
Three studies [41,45,47] included in qualitative synthesis (see Supplementary Materials) found no differences comparing add-on WDBR to control. Similarly, Intiso et al. [50] found no differences between groups, providing EMG-based biofeedback in addition to standard physical therapy compared to standard physical therapy alone.
Jonsdottir et al. [51] provided task-oriented gait training with EMG, while Cozean et al. [49] performed EMG biofeedback combined with FES during static and dynamic activities. Both studies showed significant improvements (p = 0.04) in favour of the experimental group compared to the conventional physical therapy group, and results were maintained at FU assessment (p = 0.02). Similarly, Mandel et al. [56] found significant differences (p = 0.04) in the experimental group (EMG biofeedback followed by rhythmic positional biofeedback) compared to no treatment, and results were maintained at FU assessment (p = 0.035). Finally, El Tamawy et al. [43] found significant differences (p = 0.001) in favour of a treadmill with biofeedback group compared to control, and Lee et al. [54] found significant differences (p = 0.05) in favour of Neurofeedback group compared to sham therapy.

Qualitative Synthesis
Due to heterogeneity between studies, the remaining outcome measures were not pooled into the meta-analysis but have been considered in a qualitative synthesis of the results (see Table S1 in Supplementary Materials).

Feasibility and Usability of WS Training
Only four studies [42,44,45,58] evaluated the feasibility and usability of wearable devices. In the study by Carpinella et al. [42], the Tele-healthcare Satisfaction Questionnaire-Wearable Technology showed that all patients, but one found the wearable device beneficial. Moreover, it was considered reliable, safe, and easy to use by all the patients, and comfortable by 15 out of 17 subjects. Among all, 65% found that using the wearable devices required effort and that such effort was worthwhile for them. Physiotherapists appreciated the wearable devices, but they suggested reducing the number of sensors and to simplify calibration procedures. In the study by Ginis et al. [44], participants were very positive about the system, and scores on user-friendliness were on average above 4 on a 5-point Lik-ert scale. Further, it was observed that participants with previous smartphone experience had the least problems using the system. In the study by Schenk et al. [58], participants described their experience using the technology with an adapted questionnaire. Most participants stated that it was fun. Likewise, most participants rated the usage, form, and design of the technology positively. They felt safe while using it, never experiencing fear of falling, and without the need for balance support during the therapy. For most participants, the balance exercises were not difficult to perform and were not too fast. Finally, Van Der Hovel et al. [45] did not use outcome measures to assess patients' perspective, but feasibility and usability were reported by the therapists involved in the training. They confirmed that the device-based therapy was well accepted by most participants, with the element of scoring being appreciated. They also observed that less-disable patients could operate the workstations independently, conversely patients with higher disability required more assistance. Furthermore, the system was considered suitable for use in a group setting where continuous one-on-one supervision is not needed.

Risk of Bias (RoB) Assessment
Risk of bias graph is reported in Figure 7. Most of the studies (more than 75%) present unclear risk of bias in allocation concealment (selection bias), blinding of participants and personnel (performance bias), blinding of outcome assessment (detection bias), incomplete outcome data (attrition bias), and selective reporting (reporting bias). Considering "other bias", all the studies present unclear risk of bias. The risk of bias summary (Figure 8) reveals that only 8 out of 19 studies presented a low risk of bias and four [41,43,46,48] high risk of bias in "random sequence generation". Conversely, 3 of 19 studies showed high risk of bias in "blinding of outcome assessments". Two studies [41,43] had high risk of bias in "blinding of participants and personnel", and one study [41] had high risk of bias in all the remaining bias.

Discussion
To the best of our knowledge, our systematic review with meta-analysis of randomized controlled trials represents the most comprehensive synthesis to date on the type and configuration of wearable devices for rehabilitation purposes. We also considered their feasibility, usability in a clinical setting, and attempted to estimate the effectiveness of biofeedback rehabilitation using wearable devices on balance and gait outcomes in PwND.

Training Paradigm: Type and Configuration of Sensors
This review identified a great variety in terms of sensor type and configuration and biofeedback components used in the rehabilitation of PwND.
A direct comparison between sensor configurations is difficult due to the different methodologies and lack of information. Nevertheless, we described the most frequent paradigm used to apply biomechanical and physiological sensors for different conditions or functional disorders. Most of the studies [41][42][43][44][45][46]48,52,53,55,57,58] reported the use of biomechanical sensors as pressure and inertial sensors. As previously stated, pressure sensors were prevalently placed under patients' feet to measure the ground reaction force generated by the body and were used to give biofeedback about weight-bearing or centre of pressure (COP) position during gait cycle phases.
The use of pressure sensors combined with auditory biofeedback has primarily been used in poststroke rehabilitation, to increase weight-bearing on the paretic leg, with promising results on gait speed improvement, as reported in the section about training effectiveness. In this review, only 2 studies out of 19 applied pressure sensors in PD. Even if PD patients present specific alterations of gait parameters easily detected and trained providing biofeedback with pressure sensors, References [60,61] further RCTs should explore this training paradigm on PD. On the other hand, IMU sensors provide sensitive measures of postural sway [14]. They can be used to estimate three-dimensional information of a body segment, such as orientation, velocity, and gravitational force, and to identify COM movement during balance training. They were positioned on the trunk and lower limbs (thighs, shanks, and feet) with different configurations depending on the study. It is noteworthy that for balance and gait training purpose, 5 out of 7 studies using IMU placed the sensors on the great mass body part (e.g., chest or lower back; Figure 3), often combined with IMU positioned on the smaller part of the lower limb (e.g., tight or shank; Figure 3).
Training paradigm using IMU has been applied in most of the studies involving PD patients. Even if further evidence on the effectiveness of this subgroup of patients should be provided, IMUs can assess specific postural problems typical of PD population. In PD subjects, postural deficits are easily measured using wearables devices to control axial segments (trunk and pelvis) and the relative position of the limbs. [42] In this review, physiological sensors mostly involved the measurement of EMG and EEG signals. EMG sensors (surface electrodes) were mainly placed over the distal muscle (tibialis anterior or gastrocnemius) belly of the affected leg of stroke subjects. Based on our results, training paradigms using EMG sensors are mainly used in this population; in fact, no studies using EMG on PD or other neurological populations have been included. Two studies [50,56] used EMG configuration to increase the production of voluntary dorsiflexion through activation of the anterior tibialis.
Similarly, Jonsdottir et al. [51] recorded the EMG signal from the gastrocnemius lateralis muscle to provide biofeedback about performance to increase the power production of the ankle during gait. Only Cozean et al. [49] used EMG signal to increase tibialis anterior recruitment and to simultaneously induce relaxation of the gastrocnemius in those patients with moderate or severe spasticity. Only two studies [40,54] used EEG sensors as a component of a device (Procomp Infiniti system) for neurofeedback training. EEG surface electrodes were placed on the scalp in compliance with the International 10-20 System to measure brain waves. During neurofeedback training, sensorimotor rhythm wave and β wave, which are activated when focusing, were set to reward threshold; conversely, the Delta wave, which is activated when sleeping, and the Gamma wave, which is activated when nervous, were set to 'inhibit' threshold.

Training Paradigm: Biofeedback Components
Wearable devices in the included studies mostly provided real-time biofeedback signals about performance based on positive reinforcement, with only 4 out of 19 studies providing negative reinforcement. Previous studies [22,62] suggested that positive reinforcement can be successfully used during patient's rehabilitation resulting in greater improvement in outcomes and increased retention of the motor memory. Moreover, it has been already demonstrated that the knowledge of good performance can activate the striatum, a key region of the reward system and highly relevant for motivation [23].
In this context, the rehabilitation process can be considered as a learning environment in which real-time positive biofeedback stimulates motor learning, and wearable devices providing biofeedback should be promoted to maximize learning effects.
Only three studies [42,44,51] reported a fading progression of the biofeedback, highlighting that motor-learning principles were not properly described in most of the treatments' protocols. Indeed, a proper progression is required to maximize the treatment's effects phases.

Feasibility and Usability of Sensor-Based Training
Only 4 studies out of 19 have provided information on the feasibility and the usability of wearable sensors. The two most often cited factors influencing the acceptability of the wearable sensors were safety and comfort, collected by the patients and therapists. Most of the studies failed to evaluate the feasibility and the usability-this is a major limitation. This means that we have few clues from the literature on strategies to reduce the effort required to use wearable devices by both therapists and patients, to reduce the time spent setting up, to regulate biofeedback threshold, and to improve data extraction. In this regard, the implementation of new technology in rehabilitation requires a meticulous approach, and different questions should be investigated (e.g., Is the device capable of reducing the workload for clinicians? Have all the factors associated with patient comfort and safety been evaluated?). Further studies that specifically focus on improving these aspects may help to foster the implementation of wearable sensors technology into rehabilitation settings.

Training Effectiveness
The results from our meta-analysis should be carefully interpreted because the majority of the studies present an "unclear" risk of bias in most of the established domains. Even if higher than in previous works, the number of studies included is probably not enough to determine a more conclusive statement about the effectiveness of WDBR. Further, due to the high variabilities in terms of type and configuration of sensors, outcome measures, and biofeedback, we performed the overall analysis on multiple pathologies limiting the possibility to apply the results on a specific population; only one subgroup analysis on stroke patients has been performed.
However, our meta-analysis showed possible positive effects on balance with add-on WDBR for PwND compared to a control treatment, where subjects performed the same activities with the only difference being the use of wearable devices. In specific, a higher effect on balance was found in dynamic postural stability (e.g., TUG) outcomes. Similarly, a clinically significant improvement in functional standing balance (e.g., BBS) was noted after intervention and maintained at follow-up in Parkinson and stroke population. It is noteworthy that improvements from baseline in the WDBR group seem to reach the minimally clinically important change (MCIC) in the BBS outcome, highlighting that WDBR could be a promising approach to stimulate clinically significant improvements in PwND. According to Tomlinson et al. [9], a five-point improvement in BBS is needed to reach the MCIC for people with Parkinson's disease, while a six-point difference represents the amount of change needed to conclude that a "true" clinical change in balance has occurred in stroke population [63].
Even if considerations about MCIC are clinically relevant, this evidence should be taken carefully and verified in future RCT to confirm the suggested effects.
Subgroup analysis has only been provided for four studies sharing similar sensor type (pressure sensors), biofeedback components (auditory signals with constant frequency related to performance), and population (stroke) [46,48,52,57]. This analysis suggested that the add-on WDBR provides statistically significant effects in gait speed compared to controls. Moreover, all four studies reported an improvement in gait speed higher than 0.1 m/s corresponding to the MCIC in stroke population.
Conversely, subgroups analysis on stroke population suggested no effects on TUG, probably because in these studies wearable devices were predominantly used during gait activities to increase active weight-bearing on the paretic feet, while the TUG test involves more complex activities such as posture transitions and turning.

Suggestions for Design Rules and Implementation for Clinical Practice
Based on the findings reported in this review, we have reported some highlights and suggestions to design effective and user-friendly wearable devices, facilitating their adoption in everyday clinical practice.
Firstly, we found that motor-learning principles were not properly integrated with the wearable devices used to provide biofeedback and were not properly described in most of the treatments' protocols. This is a major concern, as wearable devices should be capable of modulating the biofeedback provided according to motor-learning principles since we considered rehabilitation as a learning environment. Thus, new systems should be designed to provide constant and real-time biofeedback in the first phase of rehabilitation in which repetitive cognitive stimulation is relevant for motor learning, and a fading or reduced biofeedback in the latter phases of rehabilitation in which increased variability of the feedback is required to stimulate motor learning from the associative to the autonomous phase [15,64]. Following these principles, the possibility to regulate biofeedback threshold and intensity according to patient's need and patient's rehabilitation phase should be implemented.
Secondly, the variety in wearable devices (in terms of sensor types and configurations, and biofeedback components) makes it difficult to compare them in terms of effectiveness in biofeedback rehabilitation in specific patients' populations.
As a consequence of this high variability, currently, there is no evidence to claim if one configuration of sensors is superior to another. In our opinion, the configuration of sensors, in terms of number and positioning, should take into account the expected outcome, patient comfort, and clinical practicality (time spent to set up and reproducibility) [65]. In particular, wearable sensors' configurations measuring biomechanical parameters should be tailored to the activity being examined (e.g., standing, turning, and walking) taking into consideration pathology specific impairments (e.g., specific abnormalities of gait phases) to provide an effective biofeedback.
Thirdly, the heterogeneity of the clinical and instrumental outcome measures for balance and gait. Often, the instrumental measurements provided by wearable devices were not comparable. To solve these problems, wearable devices should provide consistent, reliable, and reproducible outcome parameters (e.g., step length, cadence, and single and double support time to assess gait) that clinical trials should report to enable a proper comparison between studies. Moreover, new easy-to-use instrumental indexes should be implemented to get a more objective and detailed assessment. Instrumental indexes should be complementary to the clinical assessment since they are able to detect subtle alterations not always visible from the clinical score to give a more complete portrait, and, consequently, to help clinicians in defining tailored rehabilitation treatments.
Finally, some suggestions are needed to improve the quality of studies on this topic. Due to the lack of evidence, there is a need for comparative studies to define what the best type of feedback and/or sensor configuration is most useful for clinical practice. Caution should be taken when considering the results of the effectiveness of using different types of sensor configuration and biofeedback on different neurological diseases. Future studies should address these issues by implementing higher-quality randomized clinical trials with larger sample sizes to improve the generalizability of the results.

Conclusions
Our systematic review provides a comprehensive picture of the use of wearable sensors in clinical practice for PwND. Biofeedback has mostly been provided in realtime during movement execution, using biomechanical sensors and positive reinforcement through auditory or visual modes. Pressure sensors and EMG were mainly used to improve weight-bearing and muscles recruitment on the paretic leg in stroke patients, while inertial sensors were used to control axial segments and limbs in Parkinson's disease. Nevertheless, the best sensor configuration in terms of number and positioning of sensors is not yet clear.
Due to the current quality of the literature, it was not possible to achieve any firm conclusions about the effectiveness of WDBR from this study. Add-on WDBR seems to provide possible positive effects on dynamic balance for people with neurological diseases.
Specific design rules on motor-learning principles, feedback components, sensor configuration, and clinical practicality should be integrated to improve the effectiveness of wearable devices biofeedback rehabilitation. Higher-quality randomized control trials with a larger sample size are needed to draw any reliable conclusion.