(PDF) Apartments: Technology for Fall Prevention*

34th Annual International Conference of the IEEE EMBS San Diego, California USA, 28 August - 1 September, 2012 Radar Walking Speed Measurements of Seniors in their Apartments: Technology for Fall Prevention* Paul E. Cuddihy, Tarik Yardibi, Zachary J. Legenzoff, Liang Liu, Calvin E. Phillips, Carmen Abbott, Colleen Galambos, James Keller, Mihail Popescu, Jessica Back, Marjorie Skubic, and Marilyn J. Rantz Ahstract- Falls are a significant cause of injury and develop. Caregivers could follow up with clinical and accidental death among persons over the age of 65. Gait functional evaluations, correct any new underlying medical velocity is one of the parameters which have been correlated to causes, and put in place assistive or protective technologies if the risk of falling. We aim to build a system which monitors an increase in fall risk is indeed determined to exist. gait in seniors and reports any changes to caregivers, who can then perform a clinical assessment and perform corrective and To this end, we focus on the use of a pulse-Doppler range preventative actions to reduce the likelihood of falls. In this control radar (RCR). This device estimates the relative paper, we deploy a Doppler radar-based gait measurement velocities of targets within its detection range by transmitting system into the apartments of thirteen seniors. In scripted an electromagnetic wave signal and measuring frequency walks, we show the system measures gait velocity with a mean shifts in the reflected waves. Characterization of gait using error of 14.5% compared to the time recorded by a clinician. this device has been described previously [8]. With a calibration factor, the mean error is reduced to 10.5%. The radar is a promising sensing technology for gait velocity in In that work, the device was used in a laboratory setting a day-to-day senior living environment. with actors and clinicians simulating walks typically seen in seniors. The lab also contained a Vicon system, which uses I. INTRODUCTION infrared markers worn by the subjects and a sophisticated system of cameras to precisely measure limb and torso Each year in the United States, over one third of seniors movements during the walk. By comparing radar signals to over the age of 65 suffer a fall. Injuries sustained in these the output of this system, it was demonstrated that the RCR falls are one of the leading causes of accidental death in this was capable of estimating mean gait velocity, variability, population [1, 2], and the rate of deaths caused by falls in this stride duration, and stride duration variability. population has risen substantially in recent years [3]. Here we demonstrate the use of the RCR system in a Since several studies have shown that better outcomes are more natural senior living environment with walks performed correlated with rapid initiation of medical intervention by members of the target senior population. Research was immediately after a fall [4], the authors and others have performed at TigerPlace, an independent living environment explored the use of radar and other technologies for detecting specially designed and built through a partnership between falls and relaying the information to caregivers quickly [5]. the University of Missouri (MU) and Americare Corporation. In the work presented herein, we demonstrate technology This unique environment provides top quality long-term care that may be used to prevent falls altogether. Such technology while also supporting research and educational opportunities would take the form of an in-home monitoring system that for researchers at MU. captures gait characteristics on a daily basis and reports Walks are scripted, and monitored by a staff performing changes. Research has identified specific gait characteristics a fall risk assessment (FRA). Since the walk portion of this which are correlated with higher risk of a falls in older adult FRA measures only the time taken to walk a distance of 10 populations [6,7]. Nonetheless, many older adults fail to feet, and since additional sophisticated and intrusive systems have their gait assessed regularly. such as Vicon cannot be put into multiple senior living If gait analysis could occur daily in an automated and apartments, we focus here on walking speed. Although a unobtrusive fashion in the home, changes in gait could be more complete assessment of gait is desirable, gait speed detected and relayed to caregivers very soon after they alone has been associated with increased risk of falls even after adjusting for other confounders and clinical scores of balance and cognition [9]. * This project was supported by grant number ROl HSOl8477 from the Agency for Healthcare Research and Quality. The content is solely the Also deployed in the senior living apartments and used as responsibility of the authors and does not necessarily represent the official a gold standard in a portion of the walks is a Kinect-based views of the Agency for HeaIthcare Research and Quality. P.E. Cuddihy and T. Yardibi are with General Electric Global Research, gait analysis system described in [10, 11]. This system has Niskayuna, NY 12309 USA (phone: 518 387-5000; e-mail: {Cuddihy, different trade-offs in privacy, cost, and gait measurement Yardibi}@ ge.com). abilities. Although all sensing systems developed by the Z.J. Legenzoff, L. Liu, C.E. Phillips, J. Keller, M. Skubic are with the authors are implemented in a way to protect the identity and Electrical and Computer Engineering Department, C. Abbott is with the activities of residents, depth and vision cameras are widely School of Health Professions; C. Galambos is with the Department of Social Work; M. Popescu is with the Department of Health Management recognized as first collecting, then protecting this and Informatics; J. Back and M.J. Rantz are with the Sinclair School of information. Radar does not collect clearly identifiable Nursing; all of the University of Missouri, Columbia. 978-1-4577-1787-1/12/$26.00 ©2012 IEEE 260 information on identities or activities, and it is likely to be housed unobtrusively as shown in Figure 2. The radars have deployed at a lower cost. It is the goal of this work to begin been modified so that their baseband signal outputs can be demonstrating that the radar can collect the same clinically recorded by an external data acquisition system. The radar relevant gait information as the vision and depth systems, works by periodically transmitting a 5.8 GHz pulse. The such that the these advantages can be leveraged. transmitted and returned signals received within a certain time period, which determines the range of the device, are In this paper, we will compare the walking speeds then mixed and low pass filtered. The transmitted signals are measured by each of the three systems: RCR, Kinect, and reflected from stationary objects at the same frequency, manual stopwatch measurement in the FRA. A key strength whereas a frequency difference is introduced when a non­ of this research lies in the fact that these scripted walks were stationary object is in the range. A number of different performed in actual senior living apartments by the residents frequency shifts can be observed in the radar measurements of these apartments. Further, the RCR is unobtrusive, as rising from the motion of the various body parts. The evidenced by the fact that it has operated day and night in the dominant (in terms of signal energy) return signal is due to apartments. the torso and can be used to estimate walking speed. The We leave for future research the measurement of baseband radar output is sampled at 960 Hz with a additional gait parameters such as stride length and commercially available AID converter, specifically the variability, as well as the evaluation of daily unscripted DATAQ DI-710 data logger [13]. The measurements are walks. transmitted wirelessly to a computer, where they are recorded into a database for post-processing. This radar has been II. METHODS previously demonstrated to be effective in estimating gait Thirteen residents of TigerPlace participated in the velocity, stride rates and the variability associated with these overall study. Subjects consist of five males and eight variables in [8]. females, aged 75 to 97. Eleven residents walk independently According to the Doppler principle, velocity is related to without a walker, cane, or wheelchair during their walking the frequency shift in the measurements as follows: v = speed assessment, and two use a walker. Seven participants cb.f f(2fJ, where c is the speed of light, b.f is the frequency live alone, and the remaining six were made up of three shift, and fe is the radar carrier frequency. In order to couples sharing their apartment. estimate the frequency shifts from the radar measurements, a Walks were collected during fall risk assessments Spectrogram (DrRantzSeniorSlow3) conducted by study personnel over a three month period from 1� r---�----�--� Nov 2011 through Jan 2012. Each subject provided a ·10 maximum of one walk per month. Walks were simultaneously observed through the RCR and Kinect gait ·20 systems, all described below. Gait velocity measurements ,,100 obtained through these three different sources were then ;;. � -30 compared. c .. :J <T � 40 A. Fall Risk Assessment u. � Subjects enrolled in the research agreed to participate in -50 monthly FRAs. These assessments were performed in the subjects' apartments by study staff including a trained -60 clinical staff observer who watched protocol and scored each instrument using its standard rubric. The FRA protocol consists of a series of sequences of standing, reaching, Figure 1 The spectrogram of the radar signal. The unique signatures of walking, and sitting motions chosen to quantitatively measure the torso and leg motion can be clearly observed. Levels are in dB [8]. both functional performance and fall risk. They included Functional Reach, Timed Up and Go, Berg Balance Scale Fourier transform based algorithm is employed. The and others. algorithm is outlined below and further details can be found in [8]. In the context of the greater FRA, each subject was asked to walk 10 ft starting and ending in a standing position, then For estimating velocity, the raw data is first passed turn and repeat in the opposite direction. These were each through a band-pass filter with a pass-band region of 5-100 recorded as separate walks. Subjects were oriented such that Hz. This removes very low frequency contamination from the these walks were oriented directly towards and away from data as well as high frequency noise since normal gait the radar units. Distance between the RCR unit and the velocities correspond to frequencies much lower than 100 Hz closest end of each walk varied by apartment from just over (8.5 ftls). The algorithm then divides the filtered signal into three feet to almost fourteen feet. The time of each walk was overlapping time segments. The fast Fourier transform (FFT) measured with a stopwatch. This measurement was of each segment is computed after applying a Hanning considered the gold standard gait velocity. spectral conditioning window and using an appropriate amount of zero-padding. (Specifically, in this study, we have B. Radar used segment lengths of 400 time samples, 75% overlap The radars used in this study are low-cost commercially between segments and 4096 point FFTs.) The resulting short­ available pulse-Doppler range control radars (see, e.g., [12]), time Fourier transform (STFT) image is smoothed out in the 261 time dimensions with a 3 sample moving average (at every velocity, the cosine rule was used to translate the centroid's frequency bin). The frequency with the maximum spectral location, velocity, and direction of travel into a measure of energy at each time instant is then used to estimate the gait subject's velocity relative to the radar. It is this velocity that velocity. The above process can be repeated iteratively, while a Doppler radar actually measures. adjusting the band-pass filter frequencies adaptively as a As part of algorithms designed to save storage space, the function of the estimated gait frequencies in the previous Kinect system only saved data during brief periods where a iteration, for improved performance (see [8] for details). walk was believed to be present. These algorithms are still in Figure 1 shows a sample STFT plot obtained from the radar. development, and some of the walks--or portions thereof-­ As mentioned previously, the dominant returns from the torso were not captured for validation by the Kinect system. can be readily observed. The secondary returns are from the leg and arm motions. III. RESULTS AND DISCUSSION A. Experiment Scope From the fall risk assessments conducted from November, 2011 to January, 2012, 16 data sets were available with time-synchronized gait measurements from the FRA, Kinect and radar. In addition 12 data sets were available with time-synchronized measurements from the FRA and radar. There exists at least one walk from each of 11 different subjects in these measurements. These subjects include 4 men and 7 women, age 75 to 97. Nine of the 11 walked independently during their FRA and 2 used walkers. Seven participants lived alone and the remaining 4 were couples. Figure 2 Radar unit in unobtrusive setting (left) Lid removed, showing data acquisition and radar (right) B. Comparative Analysis The FRA velocities are computed by using the time it C. Kinect took for the subjects to walk the pre-designated 10 ft paths, as As part of the broader research project, subjects agree to measured by a stopwatch. Kinect is used to compute the have the Kinect-based depth camera fall risk assessment walking velocities as well as the angles during these walks. system deployed and operational in their apartments round The eventual goal in this study is to evaluate the gait velocity the clock as shown in Figure 3. This included coverage of estimation performance of the radar and compare it with the times during which FRAs were performed. FRA velocity estimates. Accordingly, in this study, the ,-----,------,------,-----,------,-----, 30 25 '" <I> 20 !!! go :E. '" 15� g> « I --Kinect 10 0.5 --...;. ,---_·_--_·_--_·_--_·_-- t -.. Radar i -6- Radar Corrected i -. Kinect Angles o 5 10 15 20 25 Run # Figure 4 The velocity estimates from FRA, Kinect, Radar and walking Figure 3 Kinect sensor installed in apartment with computer angles estimated by Kinect(angles are with respect to the radar). over the refrigerator Kinect sensor will be used to estimate the walking angles As described in previous research [10, 11], this system with respect to the radar. The radar velocity estimates can starts with an image from a Kinect depth camera, performs then be corrected by using these angle estimates and foreground extraction, and estimates walking speed by compared with the FRA (the radar measured velocity is the calculating the centroid of the foreground 3D point cloud. true velocity of the gait multiplied by the cosine of the angle That centroid is then projected onto the ground plane, and between the walk orientation and the radar). speed is computed from the change in position measured over Figure 4 shows the velocity estimates obtained from the each frame. FRA, Kinect and radar measurements for the first 16 runs, For this research, the Kinect system was used primarily and from the FRA and Kinect measurements for the for secondary validation. Both the centroids and images were remaining 12 runs. The average walking angles during each used to verify the walk data. In addition to absolute gait walk with respect to the radar and as estimated by Kinect are 262 also superimposed on these plots. It is observed that the radar the radar is known, and corrected for, the accuracy is estimates (both before and after accounting for walking improved to an average of 11.9%. With a calibration factor, angle) are in very good agreement with the FRA the average error is 10.5%. This is a promising first step in measurements. The radar velocity errors can be computed showing the ability of the radar to measure gait velocity in a with respect to the FRA as follows: I vRadar - VFRA I / VFRA' typical senior living environment. The error levels obtained using this formula are shown in Measurements were taken using a continuous monitoring Figure 5. The blue dots in this figure represent the 16 samples system housed in a small unobtrusive piece of furniture, with for which Kinect walk angles are available. The remaining 12 the system remaining in operation around the clock for samples are shown with the circles. The solid line shows the several weeks. This further demonstrates the feasibility of average radar error over all of the 28 runs. The dashed line, collecting daily gait monitoring data in a home environment. on the other hand, is the average of the radar error samples The next important step will be expanding the system to measure unscripted walks throughout the day. This will 40 ,-----�----�--_, o involve either controlling or measuring the angle of the walk 35 ················ 0 ·····································............. . in order to maximize accuracy of the radar. It will also require removing outliers caused by VISitors, pets, 30 ...................................................................... meandering walks, etc. From the collection of daily walks, a � gait characterization score will be obtained. It is this score � 25 ...................................................................... that can be monitored with trend or anomaly detection « • 0::: : algorithms in order to trigger alerts to caregivers. Such a i II.. 20 . . . . . . . . . . . . . . . . � ..................................: • ................ system holds great promise to reduce falls and fall-related injuries in seniors. � 15 ...·...· ..... ·��!�: ...·...· ........................ · ............ . . . I W REFERENCES 10 : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .--r • ... . . .- . . . . . . . . . . . . . [1] Centers for Disease Control and Prevention, National Center for 5 ................ S...................................•................ Injury Prevention and Control. Web -based Injul)' Statistics QueI)' and Reporting System (WISQARS) [online]. Accessed November o 30, 2010 • • [2] Murhy SL, "Deaths: Final Data for 1998," National Vital Statistics O L-----�----�--� Radar Radar Corrected Reports, vol. 48, no. 11. Hyattsville, Maryland: National Center for Health Statistics, 2000. Figure 5 The velocity estimation error of the radar computed with respect [3] Stevens JA. Fatalities and injuries from falls among older adults­ to the FRA. The blue dots represent the 16 samples where corrected radar United States, 1993-2003 and 2001-2005. MMWR 2006a;55(45). estimates are available and the remaining 12 samples are shown with the [4] C. G. Moran, R.T. Wenn, M. Sikand, A.M. Taylor, Early mortality black circles. The solid line represents the overall mean of the radar after hip fracture: is delay before surgery important ", J. of Bone and velocity errors and the dashed lines represent the mean of the radar Joint Surgery, pp. 483-9, 2005. velocity errors for the first 16 samples (the blue dots). [5] Liu L, Popescu M, Skubic M, Rantz M, Yardibi T & Cuddihy P, "Automatic Fall Detection Based on Doppler Radar Motion, " Proceedings, 5th International Conference on Pervasive Computing for which the Kinect angles are available. In addition, the Technologies for Healthcare, Dublin, Ireland, May 23-26, 2011, pp errors after angle corrections for these sample points are 222-225. [6] BE. Maki, Gait changes in older adults: predictors of falls or shown on the right hand side. The average absolute error is indicators of fear, Journal of the American Geriatrics Society, vol. 14.5%. The average error decreases to 11.9% when the radar 45(3), pp. 313-20, 1997. velocities are corrected for angle in the smaller subset for [7] JM Hausdorff, DA Rios, HK Edelberg, Gait variability and fall risk in which that angle was available from the Kinect system. Note community-living older adults: a I-year prospective study, Arch Phys that the difference is not dramatic because the walk angles Med Rehabil, 2001;82:1050-6. [8] Yardibi T, Cuddihy P, Genc S, Bufi C, Skubic M, Rantz M, Liu L & are relatively small in most of the runs as observed in Figure Phillips C, "Gait characterization via pulse-Doppler radar, " 4. During FRA' s the walking paths are arranged to be as Proceedings, 9th Annual IEEE International Conference on Pervasive much towards and away from the radar as possible, hence Computing and Communications: Workshop on Smart Environments resulting in small angles. to Enhance Health Care, Seattle, WA, March 21-25, 2011, pp 662- 667. It is important to note that in the broader dataset, the radar [9] J. Verghese, R. Holzer, RB. Lipton, and C. Wang, "Quantitative Gait is underestimating gait velocity in the vast majority of cases. Markers and Incident Fall Risk in Older Adults," J Gerontol A BioI By adding a calibration factor of 0.19 ftls to the radar gait Sci Med Sci (2009). [10] Stone E & Skubic M, "Passive In-Home Measurement of Stride-to­ velocity, the average error is reduced to 10.5%. Stride Gait Variability Comparing Vision and Kinect Sensing, " Proceedings, 33rd Annual International Conference of the IEEE IV. CONCLUSIONS AND FUTURE DIRECTIONS Engineering in Medicine and Biology Society (EMBC' 11), Boston, MA, August 30-September 3, 2011, pp 6491-6494. We have presented a Doppler radar-based gait monitoring [11] Stone E & Skubic M, "Evaluation of an Inexpensive Depth Camera system and deployed it to eleven senior apartments. In 28 for In-Home Gait Assessment, " Journal of Ambient Intelligence and scripted walks, the system measured gait velocity to within Smart Environments, 3(4):349-361, 2011. [12] Information available at: 14.5% of measurements taken by a clinician with a www.clrwtr.com/PDF/GE-Security/Sentrol-PrecisionLine-Series.pdf stopwatch. When the precise angle of the walk relative to [13] See www.dataq.com. 263