**Sensors for Gait, Posture, and Health Monitoring**

**Volume 1**

Special Issue Editor **Thurmon Lockhart**

MDPI • Basel • Beijing • Wuhan • Barcelona • Belgrade

*Special Issue Editor* Thurmon Lockhart Arizona State University USA

*Editorial Office* MDPI St. Alban-Anlage 66 4052 Basel, Switzerland

This is a reprint of articles from the Special Issue published online in the open access journal *Sensors* (ISSN 1424-8220) from 2017 to 2019 (available at: https://www.mdpi.com/journal/sensors/special issues/Gait Recognition).

For citation purposes, cite each article independently as indicated on the article page online and as indicated below:

LastName, A.A.; LastName, B.B.; LastName, C.C. Article Title. *Journal Name* **Year**, *Article Number*, Page Range.

**Volume 1 ISBN 978-3-03936-342-1 (Pbk) ISBN 978-3-03936-343-8 (PDF)** **Volume 1-3 ISBN 978-3-03936-348-3 (Pbk) ISBN 978-3-03936-349-0 (PDF)**

c 2019 by the authors. Articles in this book are Open Access and distributed under the Creative Commons Attribution (CC BY) license, which allows users to download, copy and build upon published articles, as long as the author and publisher are properly credited, which ensures maximum dissemination and a wider impact of our publications.

The book as a whole is distributed by MDPI under the terms and conditions of the Creative Commons license CC BY-NC-ND.

### **Contents**




### **About the Special Issue Editor**

**Thurmon Lockhart** is Professor in the Biomedical Engineering and Biological Design program in the School of Biological Health and Systems Engineering at Arizona State University, Tempe, AZ. He is also Adjunct Professor at the Barrow Neurological Institute, Research Affiliate at the Mayo Clinic College of Medicine, Division of Endocrinology, and Guest Professor at Ghent University in Belgium. His research focuses on the identification of injury mechanisms and the quantification of sensorimotor deficits and movement disorders associated with aging and neurological disorders on fall accidents. His academic grounding in biomechanical modeling, nonlinear dynamics, human postural control, gait mechanics, and wearable biosensor design underscore a fundamental capacity to provide unique clinical solutions to injury prevention utilizing both engineering and biomedical principles. He has translated research findings into practice by reaching a significant number of external organizations and individuals. His outreach efforts have impacted several organizations, including UPS, the US Navy, Los Alamos National Security, the DOE, GE, and BP. In recognition of these scientific achievements, Prof. Lockhart and co-workers were awarded the Alexander C. Williams, Jr., Design Award from the Human Factors and Ergonomics Society in 2008. His research was recently featured on the PBS NOVA ScienceNow and Good Morning America programs and in the Fortune, AgingWell, Men's Health, and Discover magazines.

### **Preface to "Sensors for Gait, Posture, and Health Monitoring"**

The acquisition of gait and postural characteristics during active and passive movements provides important information about limb propulsion and postural control strategies and provides insight into performance and risk of injury. These measures were traditionally assessed by utilizing motion capture systems and force plates. Although modern motion capture laboratories collect precise gait and posture data, they are expensive and immobile and require serial (single person at-a-time data capture) and clustered data collection, limiting the use of motion capture in the field to obtain more realistic motion profiles that may be applicable to various interventions.

As such, in recent years, many technologies for gait and posture assessments have emerged. Wearable sensors, active and passive in-house monitors, and many combinations thereof all promise to provide accurate measures of gait and posture parameters. The objective of this Special Issue is to address and disseminate the latest gait and posture monitoring systems as well as various mathematical models/methods that characterize mobility functions.

This Special Issue explores the core scientific issues associated with the use of custom-designed, wearable, wireless sensor nodes for continuous, non-invasive gait–posture–activity monitoring and analysis in orderto accurately study the relationship between these monitoring variables and physical and psychological health conditions to predict adverse medical events in a variety of populations. This type of assessment will dramatically expand the clinical usefulness of these analyses and pave the way for identifying potential adverse health conditions and appropriate interventions for those most at risk.

> **Thurmon Lockhart** *Special Issue Editor*

### **Technology-Based Feedback and Its Efficacy in Improving Gait Parameters in Patients with Abnormal Gait: A Systematic Review**

#### **Gema Chamorro-Moriana 1, Antonio José Moreno <sup>1</sup> and José Luis Sevillano 2,\***


Received: 11 October 2017; Accepted: 2 January 2018; Published: 6 January 2018

**Abstract:** This systematic review synthesized and analyzed clinical findings related to the effectiveness of innovative technological feedback for tackling functional gait recovery. An electronic search of PUBMED, PEDro, WOS, CINAHL, and DIALNET was conducted from January 2011 to December 2016. The main inclusion criteria were: patients with modified or abnormal gait; application of technology-based feedback to deal with functional recovery of gait; any comparison between different kinds of feedback applied by means of technology, or any comparison between technological and non-technological feedback; and randomized controlled trials. Twenty papers were included. The populations were neurological patients (75%), orthopedic and healthy subjects. All participants were adults, bar one. Four studies used exoskeletons, 6 load platforms and 5 pressure sensors. The breakdown of the type of feedback used was as follows: 60% visual, 40% acoustic and 15% haptic. 55% used terminal feedback versus 65% simultaneous feedback. Prescriptive feedback was used in 60% of cases, while 50% used descriptive feedback. 62.5% and 58.33% of the trials showed a significant effect in improving step length and speed, respectively. Efficacy in improving other gait parameters such as balance or range of movement is observed in more than 75% of the studies with significant outcomes. Conclusion: Treatments based on feedback using innovative technology in patients with abnormal gait are mostly effective in improving gait parameters and therefore useful for the functional recovery of patients. The most frequently highlighted types of feedback were immediate visual feedback followed by terminal and immediate acoustic feedback.

**Keywords:** feedback technology; gait; rehabilitation; motor control

#### **1. Introduction**

The basic motor functions of the human being, such as gait, can be altered because of a wide range of traumatalogical, neurological, rheumatic, etc. pathologies [1,2]. Hip arthrosis [3], knee osteoarthritis [4], strokes, hemiparesis [5–7], or lower-limb amputations [8], all produce important alterations to gait patterns.

Developments in technology and information technology (IT) have enabled the development of new techniques for gait re-training based on feedback supplied by electronic devices. This has been demonstrated by authors such as Druzbicki et al. [5], Basta et al. [9], Zanoto et al. [10] and Segal et al. [11].

The basic principle of feedback is the ability to voluntarily control and change certain bodily functions or biological processes when information is provided about them [12]. The main advantage of feedback is the supply of information about a specific biological process about which the patient does not consciously have information [13].

Currently, technology is developing towards facilitating the functional recovery of the patient, sometimes even without the physiotherapist. These treatments incorporate: robot assisted movement [10,14–16], virtual reality technology [17] and inertial monitoring devices [18,19] amongst others. Some of these systems use visual [5,11,20], acoustic [15,21] and/or haptic [22,23] feedback in a coherent and detailed way, adapted to each user's individual needs [24]. New technologies based on feedback are extremely useful in the area of rehabilitation for re-educating an altered function or teaching a new one [2,25]. These aspects represent the main objectives of physiotherapy [13,25].

However, technological systems are frequently adopted in clinical practice without their efficacy having been proven. Researchers need to focus on providing clinical findings [24]. Therefore, the effects of these novel devices need to be measured [26,27] on different study populations, considering gait parameters, therapeutic guidelines adopted, clinical results obtained, systems of assessment used, etc. Similarly, we need to analyze the efficacy of different types of extrinsic feedback, in other words, that coming from an external source [28]. In this case, electronic devices will provide concurrent or immediate feedback, that is, feedback received simultaneously with the action (for example, during the foot support phase, the patient knows the amount of vertical reaction force of the floor on the limb or during walking the patient knows his/her speed); terminal or retarded feedback, or feedback received when the action is finished (for example, at the end of a tour the patient knows information about his/her progress, length of the steps, speed, kinematic of the knee, etc.); acoustic (e.g., beep, oral, etc.), visual (e.g., video cameras, displays, etc.) or haptic information (usually vibrations in some body area such as the soles of the feet) [29]; etc. Finally, this study also considers whether extrinsic feedback offers knowledge of performance (KP), in other words, characteristics of performance (e.g., if the foot bears the right direction, if the trunk remains erect during the action, etc.); or knowledge of result (KR) [30] (correct or incorrect action, score, etc.); whether this is descriptive (description of errors) or also prescriptive (how to correct errors) [24] (for example, we describe an error in walking saying that the patient is dragging the foot during the swing phase of the step. However, to correct it, we ask the patient to flex the hip and knee more when taking the step, so that the foot does not touch the ground).

Hence, the need to review, synthesize and analyze clinical findings related to the use of different kinds of technology-based feedback and their effectiveness in improving certain parameters in functional gait recovery.

#### **2. Materials and Methods**

The method was based on the PRISMA protocol [31].

#### *2.1. Data Sources and Search Strategy*

An electronic search of PUBMED, PEDro, WOS, CINAHL, and DIALNET was carried out from January 2011 to December 2016. In addition to this, we checked the reference lists of the included studies. Mesh terms (Medical Subject Headings) for English language or Decs Terms (Descriptores en Ciencias de la Salud) for Spanish database and search strategies are shown in Tables 1 and 2.


#### **Table 1.** Mesh and Decs Terms put into groups by mean.


#### **Table 2.** Search strategy.

#### *2.2. Study Selection and Inclusion Criteria*

The papers included in this review had to meet the following criteria:


The titles and abstracts of the search results were screened to check if a study met the pre-established inclusion criteria. We obtained the full text article of those studies which met the criteria, and documented the causes for any exclusions at this stage.

#### *2.3. Data Extraction*

Data extraction was carried out by one reviewer (A.J.M.) and checked for accuracy by a second reviewer (G.C.M.), using a table designed to detail information on study features, participant characteristics, feedback modality, technology employed (for feedback and assessment), interventions, comparisons, and outcome measurements.

#### *2.4. Quality Appraisal*

Apposite studies were assessed for methodological quality using the Physiotherapy Evidence Database (PEDro) critical appraisal tool [33]. This method was valid and reliable for assessing the internal validity of a study (criteria 2–9). We also evaluated the adequacy of the statistical information for interpreting the results (criteria 10–11) [34–36]. PEDro consists of 11 criteria overall; although criterion 1 refers to the external validity of the trial and is not included in the final score [34]. Each criterion could be Yes (one point) or No (0 points), with a maximum score out of ten. Only "fair" (scores 4/5) and "high" (scores ≥ 6/10) quality studies [32] were included in this review.

#### **3. Results**

#### *3.1. Search Results*

We found 884 articles in the electronic databases. Most of them in Pubmed (404), and the rest in PEDro (61), WOS (16), Cinahl (339) and Dialnet (64). Following the removal of duplicates, 776 articles were screened by title, abstract and full-text, due to: not including feedback technology, not applying the feedback directly to the patient, not being RCT, not using feedback for gait functional recovery, not having ≥4 score in PEDro Scale. After the screening, 20 studies were left for inclusion in this review.

Figure 1 shows the search and study selection process, which was based on PRISMA [37] guidelines.

**Figure 1.** Research method of this study.

#### *3.2. Characteristics of Included Studies*

A detailed summary of the features and results of each selected study is shown in Table 3.


**Table 3.** Characteristics of included studies.






**Table 3.** *Cont.*



(Limit of Stability, LOS) (%)





1 "Significant and effective" means that the outcomes show a significant effect of the

means significantly not effective in improving the parameters indicated.

technology-based

 feedback in improving the parameters indicated. "Significant and not effective"

#### *3.3. Quality Assessment*

The results of the PEDro scoring are shown in Table 4. All the selected papers rated "fair" and "high" quality (≥4 points).


**Table 4.** Completed PEDro quality appraisal.

Criteria: <sup>1</sup> Eligibility criteria were specified (not used for score); <sup>2</sup> Subjects were randomly allocated to groups; <sup>3</sup> Allocation was concealed; <sup>4</sup> Groups were similar at baseline regarding the most important prognostic indicators; <sup>5</sup> There was blinding of all subjects; <sup>6</sup> There was blinding of all therapists who administered the therapy; <sup>7</sup> There was blinding of all assessors who measured at least one key outcome; <sup>8</sup> Measures of at least one key outcome were obtained from more than 85% of the subjects initially allocated to groups; <sup>9</sup> All subjects for whom outcome measures were available received the treatment or control condition as allocated or, where this was not the case, data for at least one key outcome was analyzed by 'intention-to-treat'; <sup>10</sup> The results of between-group statistical comparisons are reported for at least one key outcome; <sup>11</sup> The study provides both point measures and measures of variability for at least one key outcome). ᑙ = criteria met; X = criteria not met.

The item "Subjects were randomly allocated to groups" (2) was scored by all papers because it was an inclusion criterion. Besides, the items "Eligibility criteria were specified" (1) and "The results of between-group statistical comparisons are reported for at least one key outcome" (10) were scored in all studies apart from 2.

Although the studies were considered to be of "fair" and "high" quality, there were two items with 0 scores: "Blinding of all subjects" (5) and "Blinding of all therapists who administered the therapy" (6).

#### *3.4. Participant Characteristics*

Relative to the population in this review, neurological patients were found in 15 out of 20 papers (75%). That is: 8 of stroke [5,7,16,19,21,39,41,42]; 1 of cerebral palsy [17]; 2 of hemiparesis [14,18]; 4 of Parkinson's [19,23,38,40]; and 1 with incomplete spinal cord injury [15]. Byl et al. [19] include stroke and Parkinson's in the same research. Besides, 2 studies were found with patients in the orthopaedic area [11,20]; and 3 more with healthy subjects [10,22,26].

All participants were adults bar one [17].

#### *3.5. Feedback Technology*

Four studies [10,14–16] stood out due to their use of exoskeletons, although only 2 of them produced feedback, Alex II [10] and Lokomat [16]. The others used complementary technology which only assists gait: Gar [14] and Lokomat [15] in this case without feedback.

Six studies were based on load platforms [5,14,18,22,40,42], such as Smart Equitest® [40], Gait Trainer® [5,18] and Functional Trainer System® [42]; and 5 on pressure sensors [11,19,22,26,39] for example Emed-Q100® [39] or Ped-Alert TM120® [21].

The feedback technology was supplemented with other tools in 8 papers: treadmills [5,11,14,16,23,40], exoskeletons [14,15], forearm crutches [15], and metronome [18]. Figure 2 summarizes the use of technologies.

**Figure 2.** Feedback technologies.

#### *3.6. Feedback Modalities*

The studies used different types of feedback: visual, acoustic and haptic; terminal/retarded and concurrent/immediate; descriptive and prescriptive; with both KR and KP. Visual feedback was used in 60% of the papers, acoustic in 40% and haptic in 15%. Terminal/retarded feedback was used in 55% and concurrent/immediate in 65%. Descriptive feedback was used in 50% of cases, with prescriptive in 60%. KP was featured in 45% and KR in 70% (Table 5).

**Table 5.** Outline of the types of feedback used in each study.


The combination of types of feedback used in descending order was: 55% visual, concurrent/immediate and prescriptive feedback [5,10,11,14,16–20,40,42]; 30% acoustic, terminal/retarded and descriptive [5,7,15,17,21,41]; 10% haptic, terminal/retarded and descriptive [22,26], acoustic, concurrent/immediate and descriptive [10,38] and visual, terminal/retarded and descriptive feedback [39,40]; 5% combined haptic, concurrent/immediate and prescriptive feedback [23] (Figure 3).

**Figure 3.** Types of feedback.

#### *3.7. Assessment Technology*

The technology used to assess gait in the selected studies was as follows: 3D movement analysis systems [5,18,20,23]; platform or treadmill force sensors [10,11,22,26,40]; pressure sensors in insoles [19], platforms [26] and parallel bars [7,15,21,22,40,41], pulsometer and ergospirometry [16]; functional training system [42]; exoskeleton [15]; and Gaitway [11].

#### *3.8. Interventions and Comparators*

In six studies the application of the feedback systems lasted 20 min [5,11,14,15,18,40], although some took up to 90 min [19]. Results also included some complementary treatments to technological feedback, such as balance [5], strength training [19], postural correction [23], stretching [7,40], speech therapy [16] and medications [11].

#### *3.9. Outcome Measures and Results*

The measurements taken in the studies were in descending order of frequency: speed, 75% [5,7,14,15,17–19,22,23,38,40,41]; step length, 50% [17–19,23,38,40,42]; Up and Go Test, 20% [19,21,22,41]; cadence, 20% [5,15,18,23]; ROM, 10% [18,23]; 10MWT 10% [5,42]; Berg Scale 10% [19,41] and 2MWT 10% [5,38]. Other parameters approached to a lesser degree were: IQR [5], peak respiratory rate [16], peak heart rate [16], etc.

For the most frequently considered parameters (speed, step length, Up and Go Test, Cadence, ROM, 10MWT and Berg Scale) the studies with significant outcomes were: 58.33% for speed [7,15,17,19,23,40,41]; 62.5% for step length [17,19,23,40,42]; 75% for TUG [21,22,41], 50% for cadence [15,23], 100% for ROM [18,23], 50% for 10MWT [42] and 100% for Berg Scale [19,41]. The clinical interventions of these studies with significant outcomes, except one [18], were effective in improving the parameters indicated. Table 6 summarizes these studies.

*Sensors* **2018**, *18*, 142

10MWT = 10 meters Walk Time; 2MWT = 2-min test; ROM = Range Of Motion.

#### **4. Discussion**

The aim of this review was to synthesize clinical findings regarding the effectiveness of technological feedback in assisting functional gait recovery. Studies defending such effectiveness versus non-technological feedback include: Baram et al. [17], Ki et al. [21], El-Tamawy et al. [23] and Sungkarat et al. [41] amongst others. The authors of this study defend the use of technological feedback but not at the cost of usual care such as: mirror therapy [7], assisted gait [7] or verbal feedback [19], etc. In other words, technological feedback and traditional physiotherapy complement each other in assisting the functional recovery of the patient. To a lesser degree, other authors such as Brasileiro et al. [18], Byl et al. [19] or Hunt et al. [20], state that technological feedback did not obtain positive, or at least significant, results, in relation to other treatments.

In Physiotherapy, the current trend is to improve treatments using new technologies adapted as much as possible to the user needs. Furthermore, it is not only the system that must be individualized, but also the type of feedback used. To exemplify this trend, consider the GCH Control System [27], an instrumented forearm crutch that controls the loads exerted on the crutch when the patient has to partially discharge his/her affected limb. It includes a feedback mechanism to send information about these loads to both the physiotherapist and the patient. When the patient has deficiencies in their coordination skills, the first sessions are usually started with indirect feedback. That is, the therapist receives feedback from the system and verbalizes it to the patient. The patient finds it easier to understand the information through the physiotherapist, who verbally adapts it to their individual conditions (e.g., "Load a little more", "Try to keep that same load", "Be careful that you load more with the right stick than with the left", etc.). The system also has the possibility of adapting the type of feedback (immediate, delayed, visual, auditory, etc.) according to the user's needs. For instance, based on our experience, the use of immediate feedback is easier for the patient and leads to a faster but less lasting result, so it is used when the patient has fewer skills. The delayed feedback is, on the contrary, more complex for the patient and the results come later, although they are more durable [43]. On the other hand, in the case of the GCH System the visual feedback is much simpler than the auditory feedback, which can only be used when the user completely dominates the former.

The articles analyzed in this review highlight how the feedback used when the subject is healthy is more complex [10,22] than when he/she is sick [7,15,17]. Also, in the present review, it is observed how there are parameters such as the cadence that can be easily corrected by means of a sound signal such as that emitted by a digital metronome or a more complex one by means of an exoskeleton [15,21,41]. On the other hand, deviations from the center of gravity are better worked by means of images [11,25].

However, it is worth mentioning that, again according to our experience, current technological systems have the tendency to personalize their treatments but without even nuancing the exact needs of the patient. It will be the therapist who makes the decision to use the technology in one way or another, always based on an initial and continuous assessment of the process and taking into consideration the coordinating, proprioceptive abilities of the user. The feedback received by the therapist for decision-making will be not only through technological means, but also through observational analysis. Both assessments, the technological and the visual or manual, are again complementary in the process of functional recovery of gait.

The technological devices, based on feedback, used by the different authors range from the complex to the basic. The complex group would include, for example: Biodex [5,18], Gaitway [11], GAR [14] or LOKOMAT [15]. The specific characteristics of each device means they each have pros and contras in terms of functionality. For example, LOKOMAT requires much more preparation time than GAR [14]. The basic devices include: heel switches [23], virtual glasses (used as computer monitor) and headphones [17], or a cane with a step-counting sensor [7]. The latter has been rendered obsolete as it has been superseded by other canes [27,44] with much more advanced technology and functions. These devices even have their own software designed specifically for functional gait recovery [27].

On the other hand, the high cost of these devices means that their everyday use is unfeasible despite their effectiveness [20]. Many authors [10,20,26], including those writing this article, favour efficiency versus the effectiveness of clinical technology in relation to financial, spatiotemporal and human resources [45]. In other words, clinical professionals require assessment and treatment systems which are feasible for everyday clinical practice, allowing adequate development of a process of functional [1,22] gait recovery. For instance, Quinzaños et al. [15] highlight the efficacy of the acoustic stimulus for re-training gait cadence and symmetry. As a result, a basic metronome [18] can be highly useful for functional gait recovery.

As this paper's introduction shows, there are many different classifications of feedback. For example, depending on the sense used, it will be acoustic, visual or haptic [28]. Relating to the moment of the stimulus, there is immediate/concurrent or retarded/terminal feedback. Finally, if the information provides data about performance or result we would be talking about KP or KR [30]. The results of this review show that authors do not just use one isolated type of feedback, instead they sometimes prefer to combine them. The one used most on its own is visual feedback [5,10,11,14,16–20,39,40,42], which is also concurrent [5,10,11,14,16–20,23,38,40,42]. In contrast, combined, we find four articles with visual and acoustic feedback at the same time [5,10,17,38]: prescriptive and again concurrent visual feedback; and descriptive, concurrent or terminal, acoustic feedback. Summing up, of the RCTs selected in this review, 55% of the articles featured prescriptive and concurrent visual feedback [5,10,11,14,16–20,40,42], and 30% descriptive and terminal acoustic feedback [5,7,10,15,17,21,41]. Although many of the devices used in the clinical trials had more types of feedback available (for example, haptic [23,26]), the authors opted for concurrent feedback, either terminal acoustic or concurrent visual which are the most effective according to Agresta et al. [6]. Thus, it has been demonstrated that concurrent feedback produces the best short-term results [24], while retarded feedback obtains the best results in the long-term [46,47]. However, other authors such as Parker et al. [24] or Salmoni et al. [48] stress that feedback can be counterproductive for learning a complex task if the procedure is applied in too detailed a manner. In other words, detailed feedback can make it more difficult for the participant to understand or process other sensory information.

We must clarify that this statement refers especially to short-term learning, particularly if complex information is offered to patients with limited coordination skills. If we consider a long-term learning the patient has more time to assume complex information although the authors of this study advocate the progression in difficulty based on a continuous assessment of the process. Another handicap of complex and prolonged feedback is the creation of the patient's dependence on receiving feedback. In this sense, the patient responds to feedback automatically in a specific task but does not integrate the learning so it is unable to extrapolate it to other similar tasks [49].

On the other hand, all the information received by the patient can be descriptive (it simply states and describes the error) or prescriptive (it provides data on how to correct the error) [24]. When the correction is simple like in the aforementioned case of the instrumented forearm crutch, just by describing the load exerted the patient knows that he/she must exert more or less force. In other cases, the description and prescription of the correction are not so obvious. When a patient touches the ground with the foot in the swing phase of a step, the correction depends on the cause and this is multifactorial (kinematics, poor coordination, etc.). The patient may not flex the hip, knee or ankle sufficiently, either due to joint limitation or muscle weakness of the tibialis anterior in the case of dorsiflexion of the ankle, hamstrings for knee flexion or iliopsoas and anterior rectus of the quadriceps in the case of the hip. Another cause would be the lack of proprioception of the patient that prevents her/him from making the gesture or even carrying it out simultaneously (step and triple flexion of the lower limb at a time). In this case, the prescription must be offered by the physiotherapist based on the causes, in a progressive and individualized manner. Selective muscle strengthening exercises, manual therapy to gain range of motion in some joint or working the patient's balance independently to the walking session may be prescribed.

Another example is arm movement during gait. Error detection and description can be easily implemented using technology. On the contrary, the prescription for its correction is usually more complex because again the causes are multiple: lack of integration of the arms in the body scheme, lack of dissociation between the scapular and pelvic waists, lack of mobility of the glenohumeral joint, etc. Deepening further, the patient can brace but not fluidly, i.e., without rotation of the shoulder girdle and without transferring the energy from proximal (trunk) to distal (arms), which would be incorrect. Even the patient may not swing arms in an opposing direction with respect to the lower limb, which would lead to an erroneous walking. Again the prescription must be made by the physiotherapist based on the cause and of course on a rigorous initial and continuous assessment.

Other authors such as Sigrist [49] affirm that to provide the idea of a movement, the feedback should be in principle prescriptive. Eventually, when the subject has internalized the action, descriptive feedback may be applied to make the correction more effective. Similarly, Sulzenbruck [50] states that, before the skill is acquired, prescriptive feedback is more effective than descriptive feedback. Still, there are authors such as Ki et al. [21] who use descriptive feedback (a beep to indicate that the weight load has been exceeded in the paretic limb) while others such as Segal et al. [11] opt for prescriptive feedback in his RCT (a graphic representation of the subject by means of a skeleton, on a screen, informs him how the optimal knee movement should be made).

Overall, the selected articles obtained significantly positive results in relation to the use of technological feedback. Even so, it should be noted that some specific parameters were not particularly significant. That is the case of stride speed or time [5,14,17,18,22,40], which can be influenced by complex robotized systems or exoskeletons, treadmills, supports etc., and the focus of the user's attention on other parameters of interest. These show an improvement in overall gait despite not actually increasing speed.

As for the populations covered, most of the technological feedback applications were applied in the neurological field. The results of this review show that 75% came from that area [5,7,14–19,21,23,38–42]. Hence, feedback is capable of changing motor strategies in patients with neurological lesions [18], with the application of this type of treatment being more appropriate during early stages of rehabilitation [24]. As for other clinical areas, this review has only included 2 articles (10%) based on muscular-skeletal lesions [11,20]. They outlined the limitation of traditional physiotherapy in the recovery of lower-limb functions [51]. Only 3 articles (15%) used a sample of healthy subjects [10,22,26]. Despite being an RCT, it is sometimes necessary to perform research with healthy subjects to ascertain the efficacy of a new technological system before using it with patients requiring treatment. Continuing with the study population, it should be noted that 95% of the reviewed articles included samples of adult subjects [5,7,10,11,14–16,18–23,26,38–42]. Only 5% of the subjects were under 18 [17]. For this reason, we believe more scientific findings need to be generated in other clinical areas and in young population samples.

The following gait parameters were assessed in the selected RCTs, in descending order of frequency: speed (cm/s) [7,15,17–19,22,23,38,40,41], step length (m) [15,17–19,22,23,38,41,42], and cadence (steps/min) [5,15,18,23]. These parameters were chosen because the unit of gait is the step and time-space parameters are essential for its assessment [2,52–55]. The measurement devices were in some cases also those providing the feedback [7,10,14,19,26,39–42]. The majority measured short-term effects [5,7,14–16,18–21,23,41,42]. The few which measured long-term effects did not obtain conclusive results [11,39,40], which underlines the need for prospective studies.

As a final reflection, the authors of this study recognize that technological progress has led to the development of highly useful tools in the field of physiotherapy which complement conventional therapy. In no case are these technologies considered substitute media, in contrast to the opinion of Parker et al. [24]. Despite the multiple benefits which new technologies offer, a physiotherapist's face-to-face treatment of a patient cannot be equaled by technological means. The personalized and intuitive adaptation of the health-care professional is the key to successful treatment.

#### **5. Conclusions**

Treatment based on feedback using innovative technology in patients with abnormal gait is mostly effective in improving gait parameters and therefore of use in the functional recovery of a patient.

Concurrent/immediate visual is the most frequently used type of feedback, followed by terminal/retarded acoustic. Also, prescriptive feedback and knowledge of result are the most frequent alternatives.

Most of the systems used are based on force and pressure sensors, normally accompanied by complementary software.

Walking speed is the most frequently evaluated parameter, with the majority of studies reporting significant improvements (in one study the changes were only significant after 3 months). The positive effect on the stride length is also found significant in most cases. In general, the number of studies with significant outcomes for the other parameters (such as balance or range of movement) is too low.

**Acknowledgments:** Part of this work was supported by the Telefonica Chair "Intelligence in Networks" of the Universidad de Sevilla, Spain.

**Author Contributions:** Gema Chamorro-Moriana and José Luis Sevillano conceptualised the idea. Antonio José Moreno and Gema Chamorro-Moriana carried out the study selection, data extraction and manuscript drafting. Gema Chamorro-Moriana, Antonio José Moreno and José Luis Sevillano have been involved in critically revising for important intellectual contents. All authors contributed to the final version and approved the final paper for publication.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **Abbreviations**



#### **References**


**Sample Availability:** All primary data were extracted from the referenced sources. Full search strategy available from the authors on request.

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **Flexible Piezoelectric Sensor-Based Gait Recognition**

#### **Youngsu Cha 1,\*, Hojoon Kim 1,2 and Doik Kim <sup>1</sup>**


Received: 2 January 2018; Accepted: 3 February 2018; Published: 5 February 2018

**Abstract:** Most motion recognition research has required tight-fitting suits for precise sensing. However, tight-suit systems have difficulty adapting to real applications, because people normally wear loose clothes. In this paper, we propose a gait recognition system with flexible piezoelectric sensors in loose clothing. The gait recognition system does not directly sense lower-body angles. It does, however, detect the transition between standing and walking. Specifically, we use the signals from the flexible sensors attached to the knee and hip parts on loose pants. We detect the periodic motion component using the discrete time Fourier series from the signal during walking. We adapt the gait detection method to a real-time patient motion and posture monitoring system. In the monitoring system, the gait recognition operates well. Finally, we test the gait recognition system with 10 subjects, for which the proposed system successfully detects walking with a success rate over 93%.

**Keywords:** flexible sensor; gait recognition; piezoelectric material; wearable

#### **1. Introduction**

Recent advancements in wearable sensors have promoted major scientific and technological developments in the field of human activity recognition [1,2]. Wearable sensors, for this purpose, have significant benefits over external sensors in terms of privacy, pervasiveness, and complexity [1]. Most people would be averse to being continuously monitored by external cameras [1,3]. It is also difficult to obtain full body imagery all the time. Moreover, image processing to recognize human activities is computationally complex.

Gait recognition and analysis are one of the most important research areas, because walking is a common human physical activity [2,4]. The gait recognition system can be utilized for therapeutic and diagnostic applications [5,6]. In addition, the sensing part can be combined to gait assistant system [7,8]. Various wearable sensors, such as inertial measurement units (IMU), pressure sensors, force sensitive resistors, accelerometers, and gyroscopes, are utilized for gait recognition and analysis [2,4,9]. These sensors are positioned on hips, thighs, knees, shanks, and feet in the lower body; they measure angles and contacts during walking [4]. For example, a piezoelectric gyroscope on the thigh of one leg has been used for detecting stride length and walking velocity [10]. Piezoresistive accelerometers have been studied in ambulatory monitoring [11]. A real-time gait phase detection system with a gyroscope and force sensitive resistors installed inside a shoe insole has also been presented [12]. Bending type piezoelectric accelerometers have been utilized for footstep detection [13]. A method of measuring joint angle for gait using a combination of accelerometers and gyroscopes has been presented [14]. A step counting method, using tri-axial accelerometers on the ankles, thigh, and waist has been proposed [15]. In most research, wearable sensors have been tightly attached to the body for accurate sensing. However, tight attachment can cause inconvenience for users.

In this paper, we introduce a flexible piezoelectric sensor-based gait recognition system available in loose clothing. The flexible sensor is made of a piezoelectric material: polyvinylidene fluoride (PVDF) [16–18]. PVDF is one of the most flexible piezoelectric materials, outputting voltage signals when it experiences strain [19–21]. Four flexible piezoelectric sensors are attached to the knee and hip parts of the loose clothing. We measure the sensor outputs during human walking by analyzing the raw signals, signal processed values, and fast Fourier transform (FFT) magnitudes. Furthermore, we demonstrate a gait detection system based on the outputs of the flexible piezoelectric sensors. Specifically, we add the gait detection method to the real-time patient motion and posture monitoring system reported in [22]. The integration of the gait detection to the monitoring system presents that the proposed method can be combined as other posture detections or system. In addition, the monitoring system can have an important update about gait detection of patients through the integration. The monitoring system recognizes five postures including walking and eight transitions between postures. We show the operation of the monitoring system in a demo scenario. Additionally, we conduct user tests of gait detection in the monitoring system.

This paper is organized as follows. In Section 2, we introduce the experimental setup of the flexible piezoelectric sensor, the module for collecting the sensor's signal, and walking test on a treadmill. In Section 3, we present an investigation of the sensor outputs during the walking motion. The decision method for gait recognition using the sensor outputs during walking is described in Section 4. In Section 5, we present the demonstration of gait recognition with a real-time patient motion and posture monitoring system. The conclusions are summarized in the final section.

#### **2. Experimental Setup**

We use similar patient cloth with flexible sensors as [22] for our experiments. In this work, the four sensors are embedded in both knee and hip parts on the patient's pants (see Figure 1). The sensors are composed of PVDF, produced by Measurement Specialties, and Mylar [17,22]. A PVDF sheet is glued on a Mylar sheet with 3M DP460 epoxy. The sensor size is (*L*)<sup>75</sup> × (*W*)<sup>25</sup> × (*T*)0.3 mm3. We sense electrical signals from the flexible sensors using conductive adhesive copper tape electrodes, produced by 3M, attached on both surface of PVDF. The copper electrodes are connected to a sensing module through wires. The voltages from the sensors, carried by the wires, are digitized using the analog-digital converters in the sensing module. The signals are then transmitted by Bluetooth wireless communication. The sensing frequency is approximately 100 Hz.

**Figure 1.** Patient clothing for four flexible sensors.

A KTS 7500TS treadmill is utilized to test walking at standard speed. In the tests, a male student (age: 25 years, height: 170 cm, weight: 70 kg) wears the patient clothes with flexible sensors and walks on the treadmill. The test walking speed varies from 0.5 to 6 km/h, which, if exceeded, requires running. When the student walks on the treadmill, we record the sensor outputs via the sensing system.

#### **3. Sensor Output for Walking**

We analyze the sensor outputs during walking. Figure 2 shows the voltage outputs from the sensors during walking at 4 km/h on the treadmill. The signal amplitudes at both knee sensors are bigger than those from the hip sensors because of several peaks. However, all signal wave forms of the hip sensors are more periodic than the knee sensors. The differences can be attributed to the loose patient cloth. During walking, the knee has a larger bending angle motion than the hip [23]. With tight clothes, the electric signal from the piezoelectric sensor attached to the knee will be larger than the signal at the hip [24]. With the loose clothing of this test, the knee sensors do not perfectly capture the periodic motions of the knees during walking. Conversely, the hip signals are sensed well during walking, because the waist band of the pants fits more tightly. Thus, the hip sensors are attached more closely than the knee sensors.

**Figure 2.** Sensor voltage outputs during walking at the (**a**) left hip; (**b**) right hip; (**c**) left knee; and (**d**) right knee. The test walking velocity is 4 km/h.

The sensor voltages correlate to the angular velocities of their bending [22]. We process data to acquire the bending motions from the sensor voltages, following the methods of [22]. Briefly, the sensor voltage outputs are processed through removing their voltage offsets and integrating as the time. We sense clearer periodic signals during walking (see Figure 3). Similarly, the processed values at the hip are more periodic than at the knee. To clarify the periodicity between the processed values at the hip and knee, we conduct FFT in MATLAB using the processed values. Figure 4 displays the FFT results from the signal processed values of Figure 3. The magnitudes of the main walking frequency in all FFT results are the largest. However, the results at the knee show that the higher harmonic frequencies still have big amplitudes.

**Figure 3.** Signal processed values from the sensor outputs during walking at the (**a**) left hip; (**b**) right hip; (**c**) left knee; and (**d**) right knee. The test walking velocity is 4 km/h.

**Figure 4.** FFT magnitude spectra of the signal processed values from the sensor outputs during walking at the (**a**) left hip; (**b**) right hip; (**c**) left knee; and (**d**) right knee. The test walking velocity is 4 km/h.

#### **4. Gait Recognition**

We decide to use only left and right hip sensors to recognize walking motion, based on the findings in Section 3. We perform additional tests at various walking speeds (0.5, 1, 2, 3, 4, 5 and 6 km/h) on the treadmill. Figure 5 displays the FFT magnitude spectra at 0.5, 1, 3 and 6 km/h. We can detect the main walking frequency by finding the biggest magnitude in each plot. We observe that the main walking frequencies increase with the walking speeds. The walking frequency as the speed of the treadmill is presented in Table 1. Herein, we focus on normal walking speed [25], which is approximately 1.1 m/s (∼ 4 km/h). At that walking speed, the main frequency is between 0.9 and 1 Hz.

For detecting the walking frequency in real-time system, we use discrete time Fourier series instead of FFT to reduce computing time. In the same context, we obtain the series coefficients of the frequency, 1 Hz, and the second harmonic. Specifically, the coefficients are obtained every half second. When the average of the two coefficients is over a decision value, we make a decision about the gait recognition. Herein, we select the decision value of 0.8 from preliminary tests. When we use the gait recognition method with the decision value, the detection rate is over 90% at 3–6 km/h.

**Figure 5.** FFT magnitude spectra during walking of (**a**) left and (**b**) right hip at 0.5 km/h, (**c**) left and (**d**) right hip at 1 km/h, (**e**) left and (**f**) right hip at 3 km/h, and (**g**) left and (**h**) right hip at 6 km/h.

**Table 1.** Walking speeds, frequencies, and detection rate on the treadmill.


#### **5. System Integration**

We adapt the gait detection to the patient monitoring system of [22]. The patient monitoring system was developed for motion and posture recognitions of patients; the gait detection method is also programmed for the system. Figure 6 displays the test condition for the patient motion and posture monitoring system. Specifically, the flexible sensor data is collected and processed in an external computer for monitoring; the monitoring program shows the decision about the motion and posture of the patient. In the monitoring program, we can detect a total of eight transitions and five postures. The five postures are (i) walking, (ii) standing, (iii) sitting, (iv) sitting knee extension, and (v) supine. We also predict the motions during the transitions between the postures (see Figure 7). To avoid incorrect operations, the walking is only transited from and to the standing pose. We note that the total four sensors' data are monitored in the system, the two hip sensors' data are utilized for the gait detection, and the two left sensors' data are used for capturing the transitions between standing, sitting, sitting knee extension, and supine. We comment that the system follows [22] about the detailed procedures for the decision between the postures of standing, sitting, sitting knee extension, and supine.

Figure 8 displays the captured computer screen image of the patient motion and posture monitoring system. At the left of the screen, it displays the processed values from the sensors of the left knee and hip and the right knee and hip. At the right of the screen, the detected patient posture is shown with a green outline. A demonstration of the patient motion and posture monitoring system is shown in Figure 9. Therein, the patient performs the transitions: (1) standing → (2) sitting on the bed → (3) sitting knee extension on the bed → (4) supine on the bed → (5) sitting knee extension on the bed → (6) sitting on the bed→ (7) standing up → (8) walking to the wheelchair → (9) stopping in front of the wheelchair → (10) sitting on the wheelchair→ (11) moving on the wheelchair → (12) standing up. All patient motion and posture monitoring of Figure 9 are displayed in Video S1 as Supplementary Material.

**Figure 6.** Experimental setup for testing the patient motion and posture monitoring system.

**Figure 7.** Flow chart for motion and posture decisions.

**Figure 8.** Captured computer screen showing patient walking detection and signal processed values from the four sensor outputs.

**Figure 9.** Test of the patient motion and posture monitoring system. (**1**–**7**) The patient lies down on the bed, and then stands up. (**7**–**9**) He moves to a wheelchair. (**9**–**12**) He sits on the wheelchair, moves via the wheelchair, and stands up from the wheelchair.

Additionally, we conduct gait detection tests from the other 10 users wearing the patient clothing. Specifically, the users walk and stop 50 times for a total of 1000 tests, 100 tests per user. We give the users simple orders about walking and stopping without specifying speed. Table 2 shows the success rates from the user test of the patient motion and posture monitoring system. The success rates are over 93%, and the total average is 97.5%. Several failed cases missed the changes or falsely detected sitting. In addition, most errors happen at taller users than the subject with the treadmill test. These errors can be attributed to the slow walking speeds of users and the low sensor outputs of the loose patient clothing. The success rate can be improved by selecting the gait detection frequency through personal statistics. For example, when we use the monitoring system for weak patients, they can have low walking speed. In this case, we need to adjust the gait detection frequency as lower value.


**Table 2.** Success rate of the gait recognition in the patient motion and posture monitoring system.

#### **6. Conclusions**

In this paper, we introduced a gait recognition system with flexible piezoelectric sensors in loose clothing. The flexible sensors were attached to the knee and hip parts on the pants. We analyzed the sensor outputs of a user during walking. Specifically, we investigated raw signals from the sensors, signal processed signals from integration, and FFT magnitude spectra. From the analysis, we found that the signals of the hip sensors are more suitable for gait detection. We demonstrated the gait detection method in a real-time patient motion and posture monitoring system. The monitoring system was utilized to detect five postures including walking and eight transitions between the postures. We showed the successful operation of the monitoring system in a demo video, which consisted of motions and postures around a bed and wheelchair. Additionally, our gait detection system was tested on 10 subjects, and the success rates were over 93%. The successful performance of our system shows the feasibility of gait detection from loose clothes using flexible sensors. We anticipate that the same methodology can be utilized to explore gait recognition using alternative flexible sensors. Moreover, it can provide an insight into motion recognition using smart suits or textiles.

**Supplementary Materials:** The following are available online at www.mdpi.com/1424-8220/18/2/468/s1.

**Acknowledgments:** This work was supported by the Korea Institute of Science and Technology (KIST) flagship program "Connected Active Space for X" (Project No. 2E28250).

**Author Contributions:** Youngsu Cha led the experiment, performed data analysis, and drafted the manuscript. Hojoon Kim conducted the experiment. Doik Kim contributed to the setup of the experiment.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

*Review*

### **Type and Location of Wearable Sensors for Monitoring Falls during Static and Dynamic Tasks in Healthy Elderly: A Review**

**Rosaria Rucco 1,2,\*, Antonietta Sorriso 3, Marianna Liparoti 1,2, Giampaolo Ferraioli 4, Pierpaolo Sorrentino 2,3, Michele Ambrosanio <sup>3</sup> and Fabio Baselice <sup>3</sup>**


Received: 29 January 2018; Accepted: 15 May 2018; Published: 18 May 2018

**Abstract:** In recent years, the meaning of successful living has moved from extending lifetime to improving the quality of aging, mainly in terms of high cognitive and physical functioning together with avoiding diseases. In healthy elderly, falls represent an alarming accident both in terms of number of events and the consequent decrease in the quality of life. Stability control is a key approach for studying the genesis of falls, for detecting the event and trying to develop methodologies to prevent it. Wearable sensors have proved to be very useful in monitoring and analyzing the stability of subjects. Within this manuscript, a review of the approaches proposed in the literature for fall risk assessment, fall prevention and fall detection in healthy elderly is provided. The review has been carried out by using the most adopted publication databases and by defining a search strategy based on keywords and boolean algebra constructs. The analysis aims at evaluating the state of the art of such kind of monitoring, both in terms of most adopted sensor technologies and of their location on the human body. The review has been extended to both dynamic and static analyses. In order to provide a useful tool for researchers involved in this field, the manuscript also focuses on the tests conducted in the analyzed studies, mainly in terms of characteristics of the population involved and of the tasks used. Finally, the main trends related to sensor typology, sensor location and tasks have been identified.

**Keywords:** falls in healthy elderly; fall risk assessment; fall prevention; fall detection; wearable sensors

#### **1. Introduction**

Falls and related accidents are a common and serious problem not only in a pathologic condition like Parkinson's Disease [1], stroke [2] or multiple sclerosis [3], in which they are due to the motor and cognitive characteristics of the specific disease, but also for healthy people aged 65 and over [4]. A fall is defined as coming to rest on the ground or floor or other lower level, suddenly and involuntarily [5,6].

Many epidemiological studies have reported a fall frequency of 28–35% in adults aged 65 and over [7–9], highlighting the need of developing effective and inexpensive ways to predict and prevent risk factors. [5,6,10].

Although some falls are probably unavoidable, most of them are due to the combination of intrinsic and extrinsic risk factors [11]. Intrinsic risk factors are related to the subject's characteristics, which include immutable biological features, sedentary lifestyle, concomitant presence of pathologies and use of medicines, age-related changes such as cognitive impairment, gait pattern alterations, and inability to maintain postural stability [12]. Musculoskeletal weakness [13] and decline of cognitive functions are known to be correlated with falls risk [14]. More specifically, a close relationship between motor and cognitive functions in both healthy elders and cognition-compromised subjects has been observed [15,16].

Nevertheless, falls may also depend upon extrinsic factors, including environmental, ergonomic and organizational aspects (inadequate housing, insufficient illumination, hazards in domestic and public places, lack of social and health care services and restricted social interaction) [12].

As reported in [17], in most cases, falls occur as the result of the individual's inability to adapt to environmental conditions. For example, increasing the slipperiness of a floor surface (i.e., from dry to wet) creates a high risk of slipping. In fact, the forces of the foot that are normally generated during gait require friction to counteract the shear forces and maintain balance. When the friction available at the level of the shoe can not meet the biomechanical requirements, a slip becomes more likely, and the consequent risk of injury with it.

In the case of highly predisposing risk factors, the frequency of falls increases, leading to a growth in the mortality rate and in the probability of causing/worsening disabilities [18]. Fractures of the femur, head injuries, damages to the lower and upper limbs and post-fall syndrome are fall-related events, which could cause a loss of confidence, hesitation and, consequently, a reduction of the quality of life [4,6,19]. The incidence of fall-related mortality is greater in men [6] than in women, although the latter show a higher prevalence of falls. The incidence of mortality is higher in men probably due to more comorbidities or to the tendency to not seek medical care until the condition becomes severe, or to delay to the access to prevention and management of diseases [6]. Furthermore, men are usually involved in intense and dangerous physical activity and risky behaviours that predispose to major risk of fall. Many epidemiological studies predict twice as many injuries up to the year 2030 [6,20].

Fall-related costs can be categorized in direct costs (which include medical and diagnostic examinations, rehabilitation treatments) and indirect costs (which have an impact on family economy) [4]. Some injuries related to falls, in addition to physical and psychological consequences on the subject, require a rehabilitative or surgical treatment, resulting in an increased hospitalization time, which has a significant economic impact on the healthcare system [6,21]. Indeed, this results in the highest direct costs, which represent about 50% of the total share in the case of falls, as shown by epidemiological data from the World Health Organization (WHO) [6].

Given the high incidence of falls in healthy elderly, in order to prevent them, it is necessary to identify predisposing risk factors, to analyze subject specific needs and to use a targeted preventive strategy, like behavioral changes [6] or promoting a healthy lifestyle [22], despite some evidence showing the opposite [23,24]. In addition, in order to make elderly people less predisposed to falls, it would be useful to plan age-friendly environments through the installation of staircases, adequate lighting and non-slip rugs [25].

The interest in the studying of mechanisms of balance and body-orientation control has increased in the last years since it has proved to be an effective tool for fall monitoring. Different techniques of measurement and assessment have been used to quantify the postural stability in elderly individuals both in static and dynamic conditions [26,27].

Human posture can be defined as "the position of one or many body segments in relation to one another and their orientation in space" [28]. Head, trunk, pelvis, lower limbs and feet are known as body "segments", whereas spinal, hips, knees, ankle and shoulder joints are denoted as body "linkages" [29]. Human posture depends on a great number of factors—among them, we cite the muscle tone [30], the orientation and the equilibrium [31], which are involved in the control of the body position with respect to the walking surface (the relationship between the Center of Mass (CoM) and

Base of Support (BoS) [32]), and in the control of alignment of Center of Pressure (CoP) with respect to the Center of Gravity (CoG) [26].

According to a biomechanical approach, the falls could be due to a change in position of the CoM projection beyond the BoS without any correction [10]. Even when standing, the body sways constantly and for this reason such posture is called static standing or stand still. The most common way to study the balance in static standing is through the trajectories of the CoP with respect to the CoG [33–35], although nonlinear measures have been proposed in some studies [36–38]. The assessment of the degree of postural sway during static standing could be useful to evaluate the ability to preserve balance control. For this reason, participants are commonly asked to perform a static upright task, on a force platform, in open- and closed-eyes condition, with both arms along the body [39,40]. During this test, it is also possible to analyze the relationship between the CoM and CoP, it being useful for fall prediction [39] and risk assessment [40].

One of the major health issues for an older adult is the loss of stability or a slip and the subsequent fall during walking. For this reason, the evaluation of dynamic stability, the assessment of activity and mobility in daily living conditions (for example, through ADLs) or through tasks such as simple walk, "time up and go" (TUG) test [41] and the 6-minute walk test [42], could provide useful motor outcomes to detect the early signs of balance alteration and then provide key information about fall risk [43]. The latter results could, however, be capitalized in a predictive fashion so as to accomplish fall detection/prediction. Many articles in the literature are focused on defining gait characteristics in the elderly with the aim of identifying the causes of fall. In the recent review article by [27], the authors tried to evaluate the sensitivity of biomechanical measures that quantify gait stability in older populations. In detail, they analyzed different approaches to study gait stability based on two different classes: (i) linear variability of temporal measures (such as swing and stance time, step width, stride velocity), and (ii) (nonlinear) orbital or local stability measures (such as Floquet multipliers or Lyapunov exponents). They concluded that, although some biomechanical approaches for determining specific parameters of mobility can assess the function and stability in the elderly, they hardly have been taken up in clinical settings because of their unclear sensitivity and specificity, together with the time and effort required for their use. Within this framework, a key role is played by the adopted sensing technologies, both in terms of hardware and software approaches.

Recently, the interest in falls is increased not only in healthy subjects, but also in pathological conditions such as Parkinson's disease [1,44,45], multiple sclerosis [3,46], stroke [2,47] and Alzheimer's disease [48,49]. However, this review investigates the studies published in the last two decades in the field of stability control for fall risk assessment, prevention and detection based on wearable sensors in healthy elderly people only. This is because, in a pathologic condition, the stability, and therefore the fall risk, strongly depends on the motor and cognitive characteristics of the disease.

An interesting review can be found in [50], where the authors focus their attention on wearable inertial sensors for fall risk assessment in geriatric populations. Within this manuscript, we extend the analysis to other sensor typologies, such as pressure sensors, electromyographic sensors and cameras, and to fall detection and prevention. Therefore, the aim of our work is to provide an overview of the most adopted sensing technologies in these fields, with a focus on the type of sensors (rather than algorithms), their position on the body and the kind of tasks they are used in. The existing literature has been found by defining a search criteria and considering some of the most important publication databases. In more detail, the search criterion is reported in Section 2, and the search results are reported in Section 3. Starting from these outcomes, some trends are identified and reported in Section 4. Finally, conclusions are drawn in Section 5.

#### **2. Methods**

This section is focused on the definition of the research criteria and strategy. After the introduction of the keywords, which are required in order to correctly understand the flow chart of the selection procedure, the framework of the research strategy is proposed and motivated. Finally, an overview of the results is briefly summarized in the last subsection.

#### *2.1. Keywords Definition*

In this section, a brief description of the main keywords is provided in order to clarify the search criteria.

*"Aged"*: with this term, we considered all those works whose studies either take into account the stability analysis for subjects aged over 64 years or consider a younger population for a preliminary assessment but with an intended application to old people.

*"Postural control"*: in biomechanics, it refers to the ability of maintaining the line of gravity (vertical line from centre of mass) of a body within the base of support with minimal postural sway. In order to classify the different articles, we adopted the terms reported in Figure 1 both for the *dynamic analysis* (i.e., "gait" and "walk" words) as well as the *statical analysis* (i.e., "static", "stationary", "standing" and "stance" words). The former refers to the study of a walking subject, while the latter investigates the function of the balance system during quiet stance.

*"Fall detection"*: it refers to the drop from a standing position or during an activity. Falls are a major public health problem among old people and the use of a fall detection system as an assistive device is important to alert when an event of fall occurs.

*"Fall prevention"*: it refers to all those articles that analyze strategies aimed at preventing the fall of older people, including doing exercises to improve muscle strength and balance, and making simple changes to home or wearing sensible shoes.

*"Fall risk assessment"*: it refers to those manuscripts that focus on the cause of falls, such as environmental issues, age, mental state and mobility.

*"Wearable sensor systems"*: national and international guidelines on fall prevention all reinforce the need to screen and assess older people, also with an attempt to identify the ones with increased risk of falling. In order to achieve this goal, one of the adopted solutions is to use wearable or portable sensors systems.

**Figure 1.** Block scheme of the adopted search strategy for the papers selection.

#### *2.2. Literature Research Strategy*

The aim of the manuscript is to provide an overview of the type and location of wearable sensors for the monitoring and assessment of falls during static and dynamic tasks (i.e., walking and standing) in healthy elderly people [51–53]. For this purpose, a bibliographical research on the most important scientific publication databases was performed. In particular, an approach similar to the one proposed

in [54] in case of the newborns has been followed. We focused on IEEE Xplore, SpringerLink, Science Direct and PubMed databases. Articles related to both dynamic and static stability analysis have been taken into account, since, as previously reported, there is a close relationship between falls and static (i.e., [26,55–58]) or dynamic (i.e., [27,59–62]) tasks. In order to perform this research, the set of keywords previously introduced has been applied. Figure 1 illustrates the flow chart of the selection process. The key point is represented by the AND block, which merges the other peripheral branches, in which each word is related to its synonyms by means of an OR condition. By doing so, the research is carried out via the intersection of at least one word per each peripheral block, otherwise the result of the query is null.

#### *2.3. Study Selection and Screening Process*

After a preliminary search, 675 results have been found for the dynamic task and 4064 articles for the static task, as shown in Figure 2. In the first step, based on paper titles and abstracts, the following exclusion criteria have been identified:


**Figure 2.** Adopted research methodology. The flow chart illustrates the two steps of the selection procedure (title and abstract filtering and full-text reading).

This selection drove into the exclusion of 4680 articles. Regarding the remaining 59 manuscripts, a second refinement step has been performed, which consists of a full-text reading of the articles. This filtering process was based on the following inclusion criteria:


Finally, 42 articles remained and were considered in this work. In more detail, we started the selection procedure by considering all the previous databases for collecting the articles, in order to provide more information regarding their distribution among the databases, although most of the papers were present on Pubmed. Therefore, we decided to indicate with the "Pubmed" label all those articles not present on the other databases.

Figure 2 provides an overview of the whole selection process and the results of the literature research are reported in Table 1. In addition in Table 2 are reported the acronyms for the tasks, in Table 3 the acronyms for the sensor positions while in Table 4 the acronyms for the sensors types, used in Table 1. Starting from these results, the quantity of works published in last years, both in conference proceedings and journals, the number and age distribution of participants, the number and typology of sensors and the types of conducted tasks have been reported as graphs. In Figure 3, the distribution of the manuscripts of Table 1 is shown. In order to make the trend more clear, a five point moving average has been applied, resulting in the red dotted line of Figure 3. It is evident that the topics of fall risk assessment, falls monitoring and falls preventing in elderly people have gained increasing interest in last years, with an almost linear growth of the published manuscripts.

**Figure 3.** Number of found manuscripts that have been published in the last 15 years (blue continuous line) and its average (red dotted line).

The manuscripts resulting from the selection have been critically analyzed in order to provide an overview of the general trends and of the most adopted technologies in the field of falls detection, falls risk assessment and falls prevention in healthy elderly. The analysis has been carried out according to three aspects: the characteristics of the participants involved in each study, the approaches for sensing and the typology of tasks that the cohort has been asked to perform.




**Table 2.** Legend of acronyms for the tasks in Table 1.

**Table 3.** Legend of acronyms for the sensor positions in Table 1.


**Table 4.** Legend of acronyms for the sensors types in Table 1.


#### **3. Results of the Bibliographic Research**

The papers resulting from the selection are summarized in Table 1 in a compact way. We chose to categorize the papers according to three main aspects: the participants involved in the study, the adopted sensors and the body portion under focus. The assessment of the participants gives an idea of the target of the study in terms of subjects, and defines if one is dealing mainly with a technical experiment of with clinical research. The second aspect allows to identify the technical features of the research and its costs. The analysis of the body parts mainly involved in the study allows to define the main biomechanical aspects considered in the work. Relevant information about such parameters can be found in the columns of the table. In more details, papers have been organized according to the following aspects:

• *Author (year)*: the family name of the first author is reported, and the publication year between brackets;


#### **4. Discussion**

#### *4.1. Participants*

In Figure 4, we report the number of participants for each study (grouped in four categories) for the considered articles, excluding the ones that do not specify the population age and/or participants' number, and that discuss a specific sensor/methodology only qualitatively, without any population involved. From the inspection of the figure, the most frequent bins are very different from a numerical viewpoint. Additionally, only a few articles have more than 100 participants in the study, reflecting the difficulty in recruiting a wide audience for these studies.

**Figure 4.** Participant age distribution in the case of different population sizes.

From the conducted review of the literature, it emerges that there are mainly two lines of research: (a) studies concerning the comparison between old versus young subjects [42,65,71,84] and (b) works aiming at underlining some peculiar differences between groups of elderly fallers and non-fallers [43,79–81]. In both cases, there is a prevalent presence of healthy elderly groups. This observation is confirmed by Figure 4, where it is shown that most of the participants in all studies are 64 years old or more, which is not surprising provided that this was used as an inclusion criterion. On the other hand, the prevalence of the young and middle (30 to 64) groups in studies with less than 10 subjects reflects the need of a fast validation in the case of a novel instrumental technology.

#### *4.2. Typology of Sensors*

By considering the sensors adopted in the studies, a first important consideration can be drawn. Specifically, most of the proposed sensing methodologies are simple from a technological point of view, as a significant share of the works adopt a few sensor technologies for monitoring the subjects:

16 papers use only one sensor and 20 papers use two sensors. On the other hand, only a small percentage of the papers proposes the joint use of three or more sensors types for improving the monitoring performance. Given the above numerical categorization (e.g. single sensor, two sensors and ≥3 sensors, respectively), we proceed in what follows with category-specific observations.

#### 4.2.1. Single Sensor

From Figure 5, it can be seen that the accelerometer is by far the most adopted sensor in the group of single-sensor type paper with a percentage of adoption of about 70%. This is probably due to the low-cost and to the plethora of devices available on the market, most of them with integrated wireless communication system, small size, weight and long-lasting batteries. For example, Cola et al. [68] propose a method requiring the use of a single waist-mounted accelerometer to achieve a continuous monitoring of deviation in the gait of elderly people. Among the single sensor approaches, about one fourth of the papers adopts a wide range of sensor technologies, both commercial products (e.g. Kinect, Wii, cameras) and prototypes. For instance, Diraco et al. [73] propose a non-invasive technique for posture classification suitable to be used in several in-home scenarios. This procedure exploits 3D point cloud sequences acquired by using a single time-of-flight sensor to classify the posture hierarchy by using a support vector machine (SVM) approach. Conversely, from the cameras proposed in [73], Najavi et al. [86] propose a commercial prototype. A more recent technology can also be employed for gait analysis and fall detection. Stone et al. [91] monitored a population of aged people (67–97 years old) for a few months using a Microsoft Kinect sensor (Redmond, WA, USA) in order to evaluate fall risk by means of the standard methodologies timed-up-and-go (TUG) time and habitual gait speed (HGS) tests.

**Figure 5.** Sensor typologies in the case of papers exploiting one, two, three or more than three sensors.

#### 4.2.2. Two Sensors

By focusing on the papers that use two sensors, two families can be identified: approaches that combine the accelerometer with a pressure sensor (generally placed within the shoes), and methodologies that jointly use the accelerometer and the gyroscope sensors, usually placed on the same electronic board. For instance, the study proposed in [75] introduces a new robust classifier for sit-to-stand (SiSt) and stand-to-sit (StSi) detection in daily activity based on both gyroscope and accelerometer. The monitoring system consists of a single inertial sensor placed on the trunk. By using dynamic time warping, the trunk acceleration patterns of SiSt and StSi are classified based on their similarity to specific templates. Gopalai et al. [76] also use accelerometers and gyroscopes to improve postural control and shorten rehabilitation periods among the young and the elderly. Beyond such "standard" sensors, the use of some peculiar devices can be explored for the stability control of elderly people. Greene et al. [77] investigate the gait variable Minimum Ground Clearance (MGC) using shank-mounted inertial sensors containing both a tri-axial accelerometer and an add-on tri-axial

gyroscope. The aim is to estimate clinically meaningful parameters, which may be used in the screening for falls risk.

#### 4.2.3. Three or More Sensors

In the case of three or more devices, other sensing technologies are adopted, including magnetometer, camera or electromyography (EMG), as described in [64], where Bertolotti et al. have designed and built an autonomous wearable 9-degrees-of-freedom system embedding three axial accelerometers, three-axial gyroscopes, and three-axial magnetometers. In [88], a support system for detecting falls of an elder person is presented. This system is an AAL (Ambient Assisted Living) system that allows to infer a potential fall by the combination of a wearable wireless sensor node, based on an accelerometer, and a static wireless non-intrusive sensory device, based on heterogeneous sensor nodes. The sensors network would spread throughout the environment, in any room of the house, routing and linking nodes to the base station. The wearable node is not intended for determining a falling situation, but to advise the reasoner layer about specific acceleration patterns that could eventually imply a fall.

#### *4.3. Position of Sensors*

Regarding the position of the sensors, 25 studies monitored two or three points of the body (Figure 6). Most importantly, the placement of sensors on the human body is mostly on the trunk.

**Figure 6.** Sensor locations in case of papers exploiting one, two, three or more than three sensors.

Among all, the work of Curone et al. [70] needs to be mentioned, who developed a processing methodology to monitor the posture and detect the activity level of the subject based on data from an accelerometer. The sensor is placed on the upper trunk, inside a garment similar to a jacket. Moreover, the algorithm is able to associate a reliability value in order to launch alarms only in case of effectively dangerous conditions. The methodology exhibits a very high accuracy in task classification (about 96%), although the evaluation has been done on a very small cohort (6 participants). In [86], authors present a method for evaluating postural transitions. The algorithm analyzes the parameters in the wavelet domain, and related them to the falling risk of the subject. In addition, Ref. [68] monitors the trunk for its analysis, reaching a very interesting accuracy value (84%) in the case of a 30 people cohort. Unfortunately, no evaluation is done in the case of older subjects.

Besides the trunk, sensors are usually placed on the foot (almost 30% of all selected studies, corresponding to pressure sensors within the shoes) and the leg, when one to three sensors are considered. In addition, the inertial device presented by Bertolotti [64] provides objective measurements of limb movements for the assessment of motor and balance control. It can be adopted for the fall risk assessment, quantifying sports exercise, studying people habits, and monitoring the elderly. Several acquisition configurations are presented as the sensors can be positioned on the trunk, on the thighs, on the arms and also on the head. In [75], the authors adopt inertial sensors positioned on

the trunk and on the leg to monitor subjects, reaching high accuracies in classifying postural transitions (up to 95%).

Very few works adopt not-body worn sensors for monitoring subjects, although such kind of systems have the advantage of being totally non invasive. In particular, in [69], a video system is proposed for monitoring Alzheimer's patients. A software module has been developed in order to recognize both physical tasks and instrumental activities of daily living. The proposed video processing algorithm is quite effective in the physical task recognition, achieving an F- score of 93. The approach proposed in [73] uses time of flight sensors for collecting 3D point cloud sequences and classifying the subject's activities. Such methodology is interesting and capable of good performances, at least without any partial occlusion of the subject.

The reason why the trunk is the most used segment for the sensor location is tightly coupled to the upright gait, being a human characteristic that requires the ability to preserve the upper body balance during walking. Interestingly, in 1992, Perry [100] defined gait analysis as centered on the lower limbs, in particular on hip, knee and ankle joint kinematics, while defining the upper body as a mere "static passenger unit" of the locomotor apparatus. Later, empirical evidence greatly changed this point of view. Indeed, in recent years, many studies have moved their focus and confirmed that the trunk plays a fundamental role both in static and in dynamic stability, see for example the study of Bertolotti et al. [64]. Furthermore, in [68], the authors used a single waist-mounted accelerometer in order to assess frailty and risk of falls in the elderly. Curone [70] designed an algorithm that analyzes, in real time, the signals produced by one three-axial accelerometer placed on the trunk, in order to classify human activities and posture transitions. In addition, Karel et al. [84] used a a triaxial accelerometer worn at the level of the sacrum to evaluate the incidence of stumbling in the elderly in daily life.

A loss of gait stability in older people is documented in many studies and the corresponding literature shows that different aging factors, such as the loss of muscle strength [101], the decline in vision and the peripheral and vestibular sensations [102], directly hamper the ability to keep the upper-body stable during walking. According to these results, the scientific interest in the trunk increased in recent years, and the development of wearable sensor technologies, which enable reliable measurements of trunk movements, has contributed to such increased interest.

As observed earlier, there is a significant adoption of pressure sensors within the shoes. The reason is that monitoring the CoP both in static and dynamic conditions, together with measuring spatio-temporal gait variability, can identify the fall risk, and help reducing it. In particular, Aminina [42] proposed a 3D method for gait analysis using the foot orientation and trajectory during each cycle based on two inertial modules worn on each foot. Using data transmitted by the modules during long distance walking periods, a dedicated algorithm provides relevant gait parameters for the evaluation of the outcome of the rehabilitation program for fall prevention in elderly subjects. In addition, Di Rosa [72] used a pair of electronic insoles, for the "Wireless Insole for Independent and Safe Elderly Living" project, in order to assess the risk of falls through a novel Fall Risk Index based on multiple gait parameters and gait pattern recognition.

#### *4.4. Tasks*

The assessment of the ability to maintain the balance in static and dynamic conditions and the estimation of the risk factors of falls in older people has been done by asking the participants to perform one or more tasks, whether simple or complex. In Figure 7, the distribution of the different types of task in case of one, two or more required activities is reported. We have divided the "Task" paragraph accordingly.

**Figure 7.** The type of tasks the participants have been asked to perform is case of one, two or more activities for fall risk assessment (FRA), fall prevention (FP) and fall detection (FD). Notice that (a) three out of all the articles listed in Table 1 were not reported in the figure (because they did not perform any test) and (b) to avoid an excessive number of categories, less investigated activities were clustered within "other" label.

#### 4.4.1. One Task

Among the papers selected in this study, roughly 23 works include in the experimental protocol the execution of one task only. As it can be see in Table 1, 15 papers focus on the analysis of dynamics using one task, such as a walk at self-selected velocity, the six minute walk test, falling test or the "time up and go". These tests are simple to manage, facilitate patients collaboration by being tolerable and also reflect most of the activities that are frequently performed by elderly people. Through the use of sensors, such as accelerometers or pressure sensors, the simple walk allows not only to detect the functional mobility of elderly people, but also to quantify their spatio-temporal parameters (speed, walking time, length and width stride etc.), the joint range of motion, and to identify a small set of features for the classification of falls. Indeed, Howcroft et al. [79] proposed the use of accelerometers and plantar pressure sensors during the 7.62 m (25 feet) test to detect, from a broader set of features, a smaller one to be used for falls classification, which could further improve falls classification. Although a greater walking distance might be a more suitable indicator of everyday walking activities, the assessment of brief gait distance is commonly used in clinical practice and could allow to find a small set of features in order to improve the clinical fall risk prediction. In addition, Similä et al. [41] suggested that gait characteristics can be used to estimate the outcome of clinical assessment of dynamic stability and predict balance decline. Furthermore, walking monitored by wearable sensors could be a useful screening tool to identify people with early signs of balance deficit [103]. As shown in Figure 7, 28% of the papers includes the analysis of static condition by employing a single task, such as standing in an anatomical reference position, with open or closed eyes. These tasks allow for detecting age-related postural changes in antero-posterior or mid-lateral directions. In Turcato et al. [39], an inertial sensor has been useful to detect the angular velocity and the linear accelerations of trunk in the static condition. The results of the study highlight that it is possible to reliably predict, with a resolution by up to 400 ms, the higher frequency CoM displacements. Since the loss of static balance occurs when the CoM goes beyond the BoS, the approach described can be useful to develop wearable devices able to warn subject of the risk of falling, so that they can implement strategies to keep the balance. Furthermore, as suggested by Gopalai et al. [76], with vibrotactile feedback tools, the static evaluation of body allows to monitor and correct the postural control. In fact, they found that the use of an intelligent biofeedback system integrated by a vibrating platform is useful to improve postural control.

Additionally, as it can be seen in Table 1, two studies implement daily activities and two other studies analyze falling task. Specifically, Di Rosa et al. [72] evaluated the performance of free-activities of daily life, in order to develop a self-learning and wearable system to identify the walking decline and to detect the risks of falls at home. Furthermore, a recent paper [82] proposed a portable device able to detect falls with accuracy and to provide monitoring and timely help for the the elderly.

#### 4.4.2. Two Tasks

Of all papers included in this review, nine of them consider two tasks, as reported in Figure 7. In this case, the most adopted tests are (single and dual task) walking, sitting down and standing up. Some studies [80,81] reported in our review have explored the use of wearable sensors, the motor control and the risk of falls during the performance of a single task (walking for 7.62 m) and dual task (walking for 7.62 m with a cognitive load). These studies found that wearable sensors, such as accelerometers and pressure-sensitive insoles, can be useful, during the performance of two tasks, to detect alterations of the motor functions such as gait variability, which is indicative of greater instability and is predictive of the risk of falls. Activities like SiSt or StSi carried out daily by old people allow for detecting postural transitions, analyzing the adopted postural strategy and preventing the fall events, as reported in [75]. In [86], the authors ask to perform two tasks in order to develop a device that can assess quantitatively the degree of mobility in old people and monitor risk factors for falls.

#### 4.4.3. More than Two Tasks

In this review, seven selected papers include more than two activities. The most adopted tasks are walking, sitting down and standing up as can be seen in Figure 7. In order to improve the quality of life of elderly people, for the authors in [67], both biomechanical evaluation and detection by sensors of multiple consecutive activities are necessary, in order to reflect the static and dynamic transition period, which could predispose to a greater risk of falling if not adequately controlled. Indeed, three types of tasks (walking, standing up and falling down) were analyzed in order to monitor and to detect fall events.

Finally, five works have not been included in Figure 7 as they only provide an overview of the importance of using sensors to assess the risk of falls in the elderly, without specifying the task used for the analysis.

#### **5. Conclusions**

Within this manuscript, a literature review of monitoring technologies for fall risk assessment, fall prevention and fall detection in case of healthy elderly people is provided. The analysis showed that several methodologies have been proposed in the literature, and it is very difficult to find a general rule. From the perspective of sensors, most of the considered methodologies uses two sensors at maximum. Among them, accelerometers and gyroscopes are the most widespread technologies, probably because they can combine low cost and informative signals. Nevertheless, some works adopt unconventional sensors, such as radar or cameras, with the aim of investigating the effectiveness of other sensing technologies for monitoring issues. Most of them only show some preliminary results, and still require a more complete statistical validation.

Regarding the position of the sensors, the trunk is the most used segment because it plays a fundamental role both in static and dynamic stability.

By looking at the task standpoint, in case of static stability assessment, the majority of the works adopt a quiet standing test with open or closed eyes, allowing for detecting age-related postural changes in antero-posterior or mid-lateral directions. Instead, for dynamic evaluations, a walking task or a sit-to-stand test is usually adopted to detect postural variations.

Regarding the validation of sensors, the information arising from the scientific literature is reported in Table 1. However, the information about sensibility, specificity and accuracy of the considered methodologies are too diverse and do not allow for evaluating the impact of parameters such as sampling rate or sensor precision. From the analysis, there is a need for the definition of one (or more) gold standards in terms of sensors (both types and location) and tasks to be performed in order to face the extremely high variety of proposed approaches.

**Funding:** This research received no external funding.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **Abbreviations**

The following abbreviations are used in this manuscript:

BoS: Base of Support CoG: Center of Gravity CoM: Center of Mass CoP: Center of Pressure

#### **References**


c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
