- Email: [email protected]
A proposal for the classification and evaluation of fall detectors
IRBM 29 (2008) 340–349
General review
A proposal for the classification and evaluation of fall detectors Une proposition pour la classification et l’évaluation des détecteurs de chutes N. Noury a,b,∗ , P. Rumeau a,c , A.K. Bourke d,e,f,g , G. ÓLaighin d,e , J.E. Lundy h,i a
Team AFIRM, UMR UJF–CNRS 5525, laboratory TIMC-IMAG, faculté de médecine de Grenoble, centre Jean-Roget, 38706 La Tronche cedex, France b Team MMB, laboratory INL, insa Lyon, 69627 Villeurbanne cedex, France c Hospital La Grave-Casselardit, Toulouse, France d Department of Electronic Engineering, National University of Ireland, Galway, Ireland e National Centre for Biomedical Engineering Science, National University of Ireland, Galway, Ireland f Biomedical Electronics Laboratory, Department of Electronic and Computer Engineering, Faculty of Science and Engineering, University of Limerick, Ireland g Wireless Access Research Group, Department of Electronic and Computer Engineering, Faculty of Science and Engineering, University of Limerick, Galway, Ireland h Hospital Cochin, Emergency Unit, Paris, France i Vigilio s.a., Evry, France Received 10 June 2008; accepted 17 August 2008
Abstract Falls affect, each year, tens of million of elderly people throughout the world. It can have immediate lethal consequences but also causes many disabling fractures and dramatic psychological consequences which reduce the independence of the person. Falls in the elderly is thus a major public health problem. The “early” detection of fall consequently raises the interest of searchers, as most of elderly fallers cannot return to a standing position on their own following a fall. It is also an interesting scientific problem because it is an ill-defined process. The goals of this study were to classify various approaches used to detect the fall and to point out the difficulty to compare the results of these studies, as there is currently no common evaluation benchmark. © 2008 Elsevier Masson SAS. All rights reserved. Résumé La chute est un problème majeur de santé publique qui touche chaque année plusieurs dizaines de millions de personnes âgées de par le monde, avec des conséquences immédiates, mortelles, mais aussi des complications handicapantes, physiques ou psychologiques. Le plus souvent la personne âgée ne peut se relever seule après la chute, aussi faut-il intervenir très rapidement, donc pouvoir détecter cet évènement dans les plus brefs délais. C’est donc un problème qui intéresse les chercheurs et, qui plus est, attise leur curiosité car la chute est un processus mal défini qui s’exprime en une grande variété de situations. L’objet de la présente étude est de clarifier les idées sur ce phénomène en exposant diverses approches contemporaines pour parvenir à sa détection. On montre aussi le défaut de cadre commun d’évaluation. © 2008 Elsevier Masson SAS. All rights reserved.
1. Introduction Falls affect each year one in two people over 80 years of age. It thus concerns tens of million of elderly people throughout the ∗ Corresponding author. UMR UJF-CNRS 5525, TIMC-IMAG Laboratory, Research team AFIRM, Faculty of Medicine, University Joseph-Fourier bâtiment Jean-Roget, 38706 La Tronche, France. E-mail address: [email protected] (N. Noury).
1959-0318/$ – see front matter © 2008 Elsevier Masson SAS. All rights reserved. doi:10.1016/j.irbm.2008.08.002
world; approximately 100 million people in developed countries. In this segment of the population, a fall may have lethal consequences. It also causes many disabling fractures [1] and dramatic psychological consequences, which reduces the independence of the person [2–4]. Falls in the elderly is thus a major public health problem. The fall consequently raises the interest of searchers, and particularly the “early” detection of the fall. Indeed, as most the elderly fallers cannot return to a standing position on their own
N. Noury et al. / IRBM 29 (2008) 340–349
following a fall [5], a relationship was established between the delay before intervention and the morbidity–mortality rate [6,7]. It is therefore important to respond very quickly to a fall. The detection of the fall is also an interesting scientific problem because it is an ill-defined process and can thus be approached using various methods. The goals of this study were to classify the various approaches to detect the fall and to point out the difficulty of comparing the results of these studies, thus the need for a common evaluation benchmark is evident. This article starts with a definition of the fall of the elderly, a discussion on the physics of a fall and its detection, and reviews the main academic and industrial research in this developing field. The authors then suggest a classification of the various approaches. In the last part, a set of evaluation parameters and a proposal of a common evaluation framework for fall detection systems are presented. 2. Definition of a fall detection Everybody has experienced an unwanted fall, whether in childhood while training to walk, or occasionally in adulthood. The fall mechanism is thus well-known to everybody. To face the fall, corrective and protective mechanisms were developed; athletes can even control “high energy” falls. Nevertheless, it is difficult to describe precisely the phenomenon, and even harder to imagine the means for its detection. Obviously, the fall of a person can be described as the rapid change from the upright/sitting position to the reclining or almost lying position. It is not a controlled movement, like lying down, for example. In 1987 the Kellogg international working group on the prevention of falls in the elderly defined a fall as “unintentionally coming to ground, or some lower level not as a consequence of sustaining a violent blow, loss of consciousness, sudden onset of paralysis as in stroke or an epileptic seizure” [8]. This definition was used in many research studies, and later extended to include falls resulting from dizziness and syncope, and other consequences of an epileptic fit or cardiovascular collapse such as postural hypotension and transient ischemic attacks, all of which may have resulted in a fall and its potential consequences. According to the above-mentioned definitions, a fall detector should be able to detect falling to the ground, in an unexpected and uncontrolled way, whatever the cause of this accidental fall. Additionally, it should have the capacity to discriminate the fall from any intentional movement. 3. Material and methods for fall detection The fall may be broken down into four phases (Fig. 1), that is, the prefall phase, the critical phase, the postfall phase and the recovery phase. During the “prefall” phase the person performs usual activities of daily living (ADL), with occasional sudden movements, like sitting or lying down rapidly, which must be distinguished from a fall. The “critical phase” consists in the sudden movement of the body toward the ground, ending with a vertical shock
341
Fig. 1. The four phases of a fall event.
on the ground. The duration of this phase is extremely short (T1 − T0 = 300–500 ms). During the “postfall” phase, the person remains inactive, frequently lying on the ground. The postfall phase shouldn’t last too long (T2 − T1 < 1 h) to reduce the consequences of the fall. Eventually the “recovery” phase is either intentional – the person stands up on his own – or with help from another person. We may classify the different studies on fall detection according to whether they focus on the “direct” detection of the critical phase or the impact shock, and the “indirect” detection of the postfall phase. Some attempted to combine both approaches. 3.1. Early detection of the critical phase of the fall During the critical phase of a fall, there is a temporary period of “free fall” during which the vertical speed increases linearly with time due to gravitational acceleration. Wu ([9] University of Vermont, USA) showed, with a video analysis of markers placed on the subject, that vertical and horizontal speeds are three times higher during a fall than for any other controlled movement. She also showed that both speeds will increase near simultaneously during a fall whereas they are strongly dissimilar during “controlled” movements. Thus, if the vertical and horizontal speed of controlled movements of the person (to rise, to bend down, to sit down) is measured, it is possible to discriminate these speeds from those occurring during a fall, which would exceed a pre-determined threshold. This inspired NaitCharif and Mckenna [10] and Rougier and Meunier [11] who attempted to detect falls by tracking the head movements with computer vision techniques and particles filters algorithms with fixed thresholds on vertical and horizontal velocities. The difficulty was, in this case, to determine this threshold. If it is too low, the system then also detects negative events (“false positive”). If the threshold is too high it will not detect positive events (“false negative”). This threshold is also dependent on the subject-to-subject variability. To overcome this last difficulty, a “learning period” may be used. This learning period may be “supervised” by asking the wearer to carry out a series of voluntary movements in order to “mark” the normal speed of execution. If the learning period is “unsupervised” the movements of the person are simply recorded during a few hours or several days, and then a statistical analysis on the measured speeds is carried out. Depeursinge ([12], CSEM in Switzerland) described a device which, using three orthogonally arranged accelerometers and three successive integrations, could locate in real-time the spatial position of the device and, by training a neural network, could detect unusual events such as the
342
N. Noury et al. / IRBM 29 (2008) 340–349
fall of the wearer. Noury et al. ([13], University of Grenoble in France) designed an autonomous sensor module, attached under the armpit, including accelerometers, an inclinometer and a vibration sensor. The sensor module detects, when the velocity of the movement exceeds a specific threshold, the sequence from a vertical posture to the lying posture, as well as the absence of movements after the fall. Prado et al. [14] developed an intelligent four-axis accelerometer unit (IAU) worn like a patch, fixed to the low back at the sacrum level. Some other studies were reported on the use of triaxial accelerometers placed either at the waist, Mathie et al. [15], or on the chest, for Hwang et al. [16] who also used a gyroscope. Recently, Bourke and Lyons proposed a fall algorithm based on thresholds applied to biaxial gyroscope signals [17]. Eventually, the detection of the very early phase of the fall could allow launching an airbag as proposed by Fukaya [18].
3.2. Direct detection of the end of the critical phase of the fall At the end of the critical phase, the body frequently hits the ground or an obstacle. The mechanical consequence of this is a sudden polarity inversion of the acceleration vector in the direction of the trajectory; this is commonly referred to as the “impact shock”. This event can be detected with an accelerometer or a shock detector; which is actually an accelerometer with a previously determined fixed threshold. Williams et al. ([19] Ireland) first described in 1998 an autonomous device carried on the belt, which featured a piezoelectric sensor to detect the impact shock when the person hit the ground as well as a mercury tilt switch to detect when the person was in a horizontal position. This research was continued by Doughty et al. [20], and a commercial version of this system was developed and marketed by the company Tunstall (Whitley lodge, Yorkshire, England). Lindemann [21] placed a 3D accelerometer external into the housing of a hearing aid. Three trigger thresholds were identified so that a fall could be identified: a sum-vector of acceleration in the xy-plane exceeding 2 g; a sum-vector of velocity of all spatial components right before the impact superior to 0.7 m/s; and a sum-vector of acceleration of all spatial components superior to 6 g. Bourke et al. [22] proposed fall algorithms based on thresholds on triaxial accelerometer signals upon impact. The difficulty of this approach is to determine the direction of the trajectory, which is obviously variable from one fall to another. Actually, most of the falls occur in the “sagittal” plane (forward or backward) as the fall often follows a voluntary movement, which is mainly performed in the saggital plane (to sit down, to rise, to walk, to bend down). A second difficulty is the location of the sensor on the body relatively to the point of impact. Depending on whether or not the sensor is near the point of impact, the “signature” of the signal recorded at the time of the shock can be significantly different and it thus becomes more difficult to recognize a fall when it occurs, thus leading to a significant number of “false positives”.
3.3. Indirect detection of the lying during the postfall As many falls end with the person lying on the ground, the simplest approach is to detect the horizontal position with a tilt sensor (a mercury contact or a ball trapped in a guide). Another approach is to detect when the feet are no longer in contact with the ground. Tamura et al. [23] proposed a photo-switch, which outputs a trigger signal to record the falling time. In a preliminary study, the system was tested for normal adults and hemiplegic patients, without any problem. The above-mentioned methods are more appropriate to monitoring an “isolated worker”, not supposed to work in the lying position. But it is obviously not suitable for the detection of falls of an older person in their home environment because, at home, lying can quite naturally occur even outside normal sleeping hours. Therefore this method is prone to many “false positives”, that is detection as falls of situations, which are not falls. A complementary solution is to detect more precisely if the body is lying on the floor, with sensitive floor tiles installed in all the rooms [24]. But the coverage is rather limited as falls do not always result in the person lying on the ground, or may occur in locations, which are not equipped with the special tiles. Furthermore, this requires a very complex wiring with potential deterioration of the home decoration. 3.4. Indirect detection of the absence of motion during the postfall After a “serious” fall (risk of severe injury), the person frequently remains immobile in a posture and/or a place. The lack of movement can thus be a consequence of a fall event. This can be detected with a basic movement or vibration sensor, placed on one of the extremities of the body, which are more mobile (e.g., wrist or ankle). Zhang et al. [25] produced a fall detection device by placing a triaxial accelerometer in a mobile phone; he proposed that, for a severe fall, the sequence of events is: a daily activity, a fall and then the person remaining motionless. Mathie et al. [26] used a similar method to identify a fall with a triaxial accelerometer. He used a range of parameters including the tilt angle, length of time during which a subject maintained a posture, metabolic energy expenditure, and previous and next activity. He also investigated the effects of three parameters: the length of a smoothing median filter, the width of the averaging window, and the value of the acceleration magnitude threshold. The main difficulty with these approaches is in the choice of the latency time (the delay T2 − T1 before taking a decision), which must be long enough to decrease the number of “false positives”. However, if the latency time is too long, it will result in a longer delay before an intervention is initiated. Lack of movements in some parts of the home (bathroom, toilets) can also be simply detected using presence infrared sensors [27] or video. Mihailidis ([28], University of Toronto, Canada) placed a video camera on the ceiling and developed scene algorithms to detect a fall. Image processing of video signals can also be used to detect a fall by detecting abrupt movements using vector analysis. This last method typically consists of subtracting successive images to keep only the variations, which are then sorted according to their direction and/or
N. Noury et al. / IRBM 29 (2008) 340–349
their amplitude. While these techniques work well in controlled environments (laboratory, scene), they must be adapted in noncontrolled environments in which neither the lighting nor the framing is controlled (it is obviously necessary that the subject be in the field of vision). Moreover, as the subject moves in a threedimensional space, it is also necessary to call upon more complex techniques, namely the use of two cameras (“stereovision”). Nevertheless, these techniques are becoming accessible, both technically and financially, thanks to the emergence of low cost cameras (web cams), the wireless transmission of images over short distances, and the possibility of embedding the required algorithms. But the video technology poses a major problem of acceptance as it requires the placement of video cameras in private living quarters, and especially in the bedroom and the bathroom, with consequent concerns about privacy. Most of the approaches attempted to produce an analytical model of the fall from knowledge or by reasoning from rules. Furthermore, a more “intuitive” approach can also be used with the development of fall detection systems based on machine learning with an observation phase (a training period) and then a classification phase. It starts with the definition of a set of classification criteria, significant and independent (discriminating). The training period may be supervised. A neural network can be trained, and then used to automatically classify future situations. Only the situations met during training can be classified, the others being mixed in a class labelled “others” (to stumble, to slip, etc.). If the training is “unsupervised”, a class “fall” is likely to be detected if the training period is long enough, even infinite if the event of fall is rare. Moreover, the first fall event is likely to be missed since its class is unknown before its first occurrence.
343
Table 1 Terminology for the classification of fall detectors. Direct detection
Detection of a sudden postural variation
Accelerometric
Detection of sudden movements and shocks to the body Detection of body height above ground level Detection of body spatial position Video monitoring of the body
Statimetric Topometric Eidolimetric Peripheral detection
Detection of behavioral/contextual variations
Behavioral Geotaximetric
Monitoring of behavioral sequences Monitoring of spatial localization of the person Monitoring of the body orientation (verticality) Monitoring of the body movements Monitoring of vital signs
Orthostatimetric Kinesimetric Physiometric Combinatorial Detection
Combination of direct and peripheral detection
Bidimensional Tridimensional Multidimensional
One direct and one peripheral parameter One direct and two peripheral parameters One direct and three or more peripheral parameters
reason is the rate of “false alarms”, resulting in inappropriate alerts of the remote monitoring service, leading to the “rejection” of the equipment by the remote supervisors. It should thus be stressed that there is currently no satisfactory solution on the market and that fall detection should be better defined.
4. Industrial developments and commercial products 5. Evaluation of the fall sensors Some of the previous results were successfully implemented in functional prototypes, and an extensive range of patents has been filed in the fall sensor area. In a few cases, final products were built and are available on the market place. For the sake of clarity, we first propose a classification of the patents with a specific terminology (Table 1). Using our terminology, we attempted to classify most of the existing patents (Table 2). We observed a regular increase in the submission of patents, beginning in 1996, with a peak (11 patents) in 2002, and now stabilized (4 patents in 2006). Most of the products available on the market are worn on the wrist or on a belt (Fig. 2). Other devices, also on the market, offer alternative solutions by “peripheral” detection of a fall (Fig. 3). In spite of the need for technological interventions to face the fall problem, and several devices being available in the market place, there is no significant industrial development of fall sensors and little use of these devices in daily geriatric practice. There are probably multiple reasons. Some devices exhibit low performance on the field. Some are not accepted by the users due to inadequate ergonomics or due to the stigmatization of the fragility of the old person. But the major
Most of the academic studies started with the design of a test instrument in order to record signals, or images, during simulated situations of fall/non fall. Eventually, they suggested algorithms, which were tested off-line on recorded signals/images. But few studies reported result data and conditions of assessment. Lord and Colvin [29] reported in 1991 some early studies on fall detection using accelerometers with very little details. Nait-Charif [10] and Rougier and Meunier [11] announced the possibility to detect a fall. The IAU of Prado et al. [14] was further evaluated by Diaz et al. [30] in a laboratory study carried out on eight volunteers to show that the device was able to distinguish true falling events from normal activities like fast walking or going up/downstairs. The prototypes made by Williams et al. [19] and Doughty et al. [20] were evaluated on 20 people to observe “false positive alarms”. The final design allowed for a reliable detection in 180 different falling scenarios. The algorithm developed by Lindemann [21] allowed discriminating ADL from intentional falls. The system made by Mihailidis was tested on 21 volunteers who carried out simulated falls and was capable of detecting 77% of falls. Mathie et al. [15] collected data from 26 normal subjects performing
344
N. Noury et al. / IRBM 29 (2008) 340–349
Table 2 Classification of the patents for fall detectors. Direct detection Accelerometric
Monoaxial Multiaxial
Statimetric Topometric Eidolimetric
Peripheral detection Behavioral
Geotaximetric
Orthostatimetric Kinesimetric
Physiometric Combinatorial detection Bidimensional analysis
Accelerometry + time without movement
Accelerometry + shock Accelerometry + image Accelerometry + physiological parameter Accelerometry + geolocalisation
Accelerometry + Δ accelerometrics
Tridimensional analysis
Multidimensional analysis
Accelerometry + time without movement + geolocalisation Accelerometry + cardiac ballistics + time without movement Accelerometry + cardiac ballistics + time without movement + data fusion
P. WO 2004/086325 A2 (07/10/04) MurataM. US 2004/0046675 A1 (11/03/04) Franco T.S. Jr US 2005/0093709 A1 (05/05/05) Cederstrom E. WO 2004/114245 A1 (29/12/04) Karner F.C. US 6,965,311 B1 (15/11/05) Fourniquet A. 2 870 378 (17/05/04) Thacker N. WO 01/63576 A2 (30/08/01) Fredriksson A. WO 2004/047039 A1 (03/06/04)
Aphycare WellCare Systems Inc IST EDF University Manchester Wespot A.B.
Coulthard J.J. US 6,972,677 B2 (06/12/05) Esteve D. EP 1 482 464 A1 (01/12/04) Power M.W. US 2002/0060630 A1 (23/05/02) Esteve D. 2 828 317 (01/08/01) Markhovsky R. US 2006/0012476 A1 (19/1/2006) Burbaud T. 2 857 140 (23/06/03) Tallman E. US 6,175,308 B1 (16/01/01) Kelly P.B. Jr WO 01/50957 A1 (19/01/01) Najafi Bijan EP 1 195 139 A1 (10/04/02) Crisco III J.J. US 2002/0060633 A1 (23/05/02) Brilman A.J. US 6,975,230 B1 (13/12/05) Bader G. US 6,337,629 B1 (08/01/02)
EDF
Roussy C.A. WO 2004/100092 A2 (18/11/04) Holdsworth D. EP 1 128 49 A1 (29/08/01) Cadet P.H. WO 98/29852 (31/12/96) Lepaul G. 2 760 116 A1 (26/02/97) Myllymaki M. WO 99/56262 A1 (04/11/99) Burbaud T. EP 1 632 920 A1 (08/03/06) Fulton J.G. US 6,377,179 B1 (23/04/02) Cosquer P. WO 01/85025 A1 (19/11/01)
DHS Tunstall
Depeursinge Y. EP 0 877 346 A1 (11/11/98) Depeursinge Y. US 6,201,476 B1 (13/03/01) Jacobsen S.C. US 6,160,478 (12/12/00) Jacobsen S.C. 2 785 073 A1 (26/10/99) Petelenz T.J. US 6,433,690 B2 (13/08/02) Lehrman M.L. WO 01/20571 A1 (15/09/00) Lehrman M.L. US 6 ?864,796 B2 (08/03/05) Lehrman M.L. WO 02/45043 A1 (06/06/02) Lehrman M.L. US 2002/0118121 A1 (29/08/02) Lehrman M.L. WO 02/061704 A1 (08/08/02) Lehrman M.L. WO 02/073564 A1 (19/09/02) Lehrman M.L. US 2002/0008630 A1 (24/01/02) Azcoita Arr. J.M WO 2006/000605 (05/01/06)
CSEM CSEM
Noury N. 01/12046 (18/09/01)
UJF Grenoble
Noury N. 06/08584 (29/09/06)
Vigi’Fall (Vigilio, UJF Grenoble)
sit-to-stand and stand-to-sit transitions and walking. His system successfully distinguished between activity and rest, giving sensitivity greater than 0.98 and specificity between 0.88 and 0.94. Noury et al. [31], in an early test, asked five volunteers to perform 15 fall scenarios five times, and obtained a sensitivity and specificity reaching 85%. Bourke et al. [17,22] announced a performance of 100% from a test involving 10 healthy young subjects performing simulated falls on large crash mats and elderly healthy subjects performing normal ADL.
Actall Corp.
Noc’Watch/fall Saver EPFL Lausanne Barlow, Joseph & Homes Ltd. Biosys AB
Inst. Polytechnique Sevenans IST Vital base Univ. Rennes
Sarcos Inc. ILife Systems Inc. ILife Systems Inc. ILife Systems Inc. ILife Systems Inc. ILife Systems Inc. ILife Systems Inc. ILife Systems Inc.
We can say that the proposed solutions and associated devices for fall detection in the elderly are numerous but their performances are difficult to compare. For the sake of clarity, it is therefore necessary to define a common set of criteria to evaluate the performance of each system. Obviously the first intrinsic “performance” of fall sensors is their ability to detect the fall, and only the fall. Unfortunately, this information is frequently unavailable in the literature. Furthermore, it is interesting to compare other objective criteria, such as the detection
N. Noury et al. / IRBM 29 (2008) 340–349
345
Fig. 2. Some commercially available fall detectors.
method, the “intrusivity” and the lifespan of the device; the last criteria are related to the acceptance so their selection may be subjective because selection depends on the aims of the researcher/engineer/marketing agent. 5.1. Quality criteria for the fall detector Fall detection is either positive if the detector properly recognizes a fall event, or negative if it does not. As the output is binary, the quality of the detector cannot be evaluated simply from a single test, instead, it is necessary to carry out a statistical analysis on a series of tests. There are four possible situations: • true positive (TP): a fall occurs, the device detects it; • false positive (FP): the device announces a fall, but it did not occur; • true negative (TN): a normal (no fall) movement is performed, the device does not declare a fall;
• false negative (FN): a fall occurs but the device does not detect it. Two criteria are proposed to evaluate the response to these four situations: • sensitivity is the capacity to detect a fall Eq. (1): sensitivity =
TP ; TP + FN
(1)
• specificity is the capacity to detect only a fall Eq. (2): specificity =
TN . TN + FP
(2)
5.2. Criteria for the evaluation of a fall sensor In addition to the intrinsic performance of the fall detector, other criteria must be taken in consideration when comparing
346
N. Noury et al. / IRBM 29 (2008) 340–349
Fig. 3. Commercial devices for “behavioral” detection of the fall.
the different approaches. Such criteria relate to the form factor, the acceptance, the usability, and the level of “accomplishment”. The form factor (mass and volume of the sensor) is particularly important if the sensor is carried by the wearer. The acceptance of the sensor is usually lower if the sensor is carried on the person (endosensor) rather than installed in his environment (exosensors). It is also influenced by the “visibility”: a visible sensor (on the wrist, the belt, the chest. . .) may stigmatize the handicap or dependence. This can also be the case with external sensors visibly placed in the domestic environment. The user friendliness depends on the autonomy, the range, and the packaging. The autonomy is the estimated time between two successive maintenance interventions, that is, to change/reload the batteries. The range is the maximum distance between the person and the sensor so that the latter can efficiently perform measuring. In the case of a carried sensor, communicating wirelessly, it is the distance between the base station and the sensor. The type of packaging concerns the biological innocuity, for instance biocompatibility, or the risk of injury for sensors directly in contact with the skin. The level of “accomplishment” was added because some promising systems are still being tested while others are already available on the market place. We tried to evaluate (Table 3) some of the most widely used devices according to these criteria. For the sake of clarity, we
gave additional information on the principle of operation, the location of the sensor and the evaluation of the performance (duration, number of test subjects). 5.3. Protocol for the evaluation of fall sensor performance It is currently practically impossible to compare the performances of various fall sensors based on literature data, as common criteria for their evaluation were not used and no common procedures were used to perform the tests. Therefore, it is very important to propose a common framework for the evaluation of the fall sensors. There are many scenarios of falling and it is not possible to consider all of them in an exhaustive way. The device must be tested in a limited number of situations, which are sufficiently representative of falls (positive situations), as well as of “pseudo” falls (negative situations). The first situation to be considered is a fall in the anteroposterior plane. As already mentioned, most falls occur during intentional movements, initiated by the person in the anteroposterior plane, forward or backward; a large number of falls occur after stumbling on an obstacle while walking, when slipping backwards on wet ground, or during a “stand-to-sit” transfer. If a person looses her/his balance in the forward direction, s/he will initially try to recover by taking some steps forward, thus amplifying the movement of the fall, and s/he will perhaps finally fall while projecting her/his arms forwards for protection. S/he can also drop to her/his knees. If loss of balance occurs backwards,
N. Noury et al. / IRBM 29 (2008) 340–349
347
Table 3 Classification of existing devices. Name
Principle
Location
Form factor
Functions
Level of evaluation
Sereoz (Aphycare, France)
Accelerometric (shock)
Wrist
Accelerometric + geolocalisation
Chest
Ynolis (“Ecureuil”, DHS, France) Noc Watch “Fall Saver” (USA)
Accelerometric + motionless time Orthostatimetric (Orientation of corporal segments)
Wrist Thigh
Watch 18 g Credit Card Tegaderm Patch
Call Button Heart Beat Surface body temperature Cancellation button Analysis of activity Heart beat Detections of standing position
Two years in institution
FDSS (csem, Switzerland)
Watch 30 g Hypoallergenic polymer 10 × 5 × 2 cm
ACTIM3D (Labo TIMC, France)
Accelerometric + motionless time
Chest (under the arm pit)
15 g 5 × 3 × 1 cm Hypoallergenic polymer Six months autonomy (on a 3 V coin cell battery)
the person will try to sit down to possibly limit the intensity of the shock impact. In some cases, the fall occurs sideways, either during a badly controlled “sit-to-stand” transfer, or if the person, when loosing balance, tries to grip the wall. Eventually, some daily life movements may present a similar amplitude or intensity to the one encountered in accidental situations, for example the action of lying down, or of sitting down, if carried out “quickly”. It is also possible to observe situations of fall initiation with recovery (stumbling). We suggest (Table 4) a set of 20 scenarios for the evaluation of fall sensors, with 50% “negative” and 50% “positive”, based on the above discussion, our experience in evaluating the Noury et al. fall sensor [31], completed by some scenarios used by Bourke et al. to evaluate his fall algorithms [17,22]. It is statistically insufficient to have only one test for each scenario for each subject. However, to avoid making excessive demands on the subjects, a maximum of three tests could be undergone by each subject in each condition. The subject should be allowed to rest as much as he wishes, when he wishes, during the tests. It should be noted that it is our experience that repetition of the same scenario results in “adaptation” to the gesture, which thus becomes less natural, it would be beneficial to vary the order of the tests, or to let the subject choose the order, only recording the order he selected. During the test, the subject will be free to adapt her/his speed in following any of the predetermined scenarios. As a consequence, s/he will have to keep her/his eyes opened. This will also enable her/him to perform self-protection strategies. As the subject carries out three times each scenario, we would collect 60 tests per subject. With a reduced sample of 10 subjects (the sample should also respect gender), that would amount to 600 data points, which is a statistically acceptable sample, significant enough to compute the specificity and the sensitivity of the device.
Analysis of activity
96% true positives
47 subjects (evaluation during 1 week) 21 subjects in lab Specificity = 94% Sensitivity = 95%
Table 4 Scenarios for the evaluation of fall detectors. Category
Name
Kind
Backward falla
Ending sitting Ending lying Ending in lateral position With recovery On the knees With forward arm protection Ending lying flat With rotation, ending in the right lateral position With rotation, ending in the left lateral position With recovery Ending lying flat With recovery Ending lying flat With recovery Vertical slipping against a wall finishing in sitting position Sitting down on a chairb then standing up Lying down on the bed then rising up Walking a few meters Bending down, picking something up on the floor, then rising up Coughing or sneezing
Positive Positive Positive Negative Positive Positive Positive Positive
Forward falla
Lateral fall to the righta Lateral fall to the lefta Syncope Neutral
Positive Negative Positive Negative Positive Negative Negative Negative Negative Negative Negative
Negative
a
The test may be performed starting with both legs straight or with knee flexion. b The height and kind of chair could also lead to separate test conditions.
Although the goal of a fall sensor is to detect the fall of elderly people, it is actually difficult to test the fall situations with them. Thus fall situations may be simulated by younger persons, or even athletes, and normal activities may be tested on elderly in the age group at risk for falls.
348
N. Noury et al. / IRBM 29 (2008) 340–349
6. Conclusion
References
The ideal fall detection system should exhibit both sensitivity and specificity reaching 100%. This was sometimes reached in experimental set ups [17,22], but when developed as an autonomous integrated fall sensor, there is a dramatic loss in performances [28,31]. No individual method seems satisfactory, it thus seems appropriate that a combination of several methods would allow for a greater selectivity. Also, a multidimensional approach, combining both kinematics and physiological parameters, could probably come closer to the 100% sensitivity and specificity target. New promising techniques could also be investigated in considering the fall as a chaotic event with stable states and “bifurcations” of “complex systems”. On the technological side, the main improvements will come with the miniaturization of the devices (integration) and reduction in the level of maintenance required. The smaller the size of the sensor, the easier it will be to fit on the person, in garments or in accessories. Also, the maintenance frequency should ideally reach one or two years to reduce maintenance and associated costs. The functionality of these devices should also be improved. First of all, the activation must be fully automatic with no need of carrier intervention. Secondly, the device with communication features could bring enhanced services to the person, such as the provision of a “social link” which would improve the conventional alert system. Daily use of an intelligent device also introduces ethical issues concerning the respect of intimacy and privacy (each movement of the person can be recorded) and also the risk of subject dependency on technology. But for the time being, the most urgent task facing the scientific community is to accept a common definition of a fall and of fall detection, and to agree on a common protocol for the evaluation of fall detection systems such as the one suggested in this study. This article focuses on the fall detection methods and devices because getting help quickly after a fall reduces risk of hospitalization by 26% and death by over 80%. Nevertheless, there are obviously several advantages offered by the analysis of the environmental/behavioral/physiological sources of the fall, in order to develop prevention systems or actions aiming at reducing the intrinsic/extrinsic causes of the fall [32–37]. Some falls are preventable when caused by environmental factors (like fall hazards at home) and by side effects of medication because they can be foreseen and avoided. Preventing falls and the resulting injuries can promote independence, by reducing or delaying the need to move out of home.
[1] Sadigh S, Reimers A, Andersson R, Laflamme L. Falls and fall-related injuries among the elderly: a survey of residential-care facilities in a Swedish municipality. J Community Health 2004;29:129–40. [2] Gostynski M. Prevalence, circumstances and consequences of falls in institutionalized elderly; a pilot study. Soz Praventivmed 1991;36:341–5. [3] Ryynanen OP, Kivela SL, Honkanen R, Laippala P. Falls and lying helpless in the elderly. Z Gerontol 1992;25:278–82. [4] Kiel DP. Falls. RI Med 1991;J74:75–9. [5] Alexander NB, Ulbrich J, Raheja A, Channer D. Rising from the floor in older adults. J Am Geriatr Soc 1997;45:564–9. [6] Gurley RJ, Lum N, Sande M, Lo B, Katz MH. Persons found in their homes helpless or dead. N Engl J Med 1996;334:1710–6. [7] Wild, Nayak DUS, Isaacs B. How dangerous are falls in old people at home? Br Med J (Clin Res Ed) 1981;282:266–8. [8] Gibson MJ, et al. The prevention of falls in later life. A report of the Kellogg International Work Group on the Prevention of Falls by the Elderly. Danish Medical Bulletin 1987;34(Suppl. 4):1–24. [9] Wu G. Distinguishing fall activities from normal activities by velocity characteristics. Jour of Biomechanics 2000;33:1497–500. [10] Nait-Charif H, Mckenna S. Activity Summarisation and Fall Detection in a Supportive Home Environment. In: 17th International Conference on Pattern Recognition, 2004. [11] Rougier C, Meunier J. Demo: Fall Detection Using 3D Head Trajectory Extracted From a Single Camera Video Sequence. In: First International Workshop on Video Processing for Security (VP4S-06), June 7–9. Quebec City, Canada: 2006. [12] Depeursinge Y. Device for monitoring the activity of a person and/or detecting a fall. 2001. [US patent 6,201,476 B1]. [13] Noury N, et al. Monitoring behavior in home using a smart fall sensor and position sensors. In: IEEE-EMBS Microtechnologies in Medicine & Biology, Lyon France: 2000. p. 607–610. [14] Prado MJ, Reina-Tosina, Roa L S.Distributed intelligent architecture for falling detection and physical activity analysis in the elderly. 2002. [15] Mathie MJ, et al. Classification of basic daily movements using a tri-axial accelerometer. Med Biol Eng Comput 2004;42:679–87. [16] Hwang JY, et al. Development of Novel Algorithm and Real-time monitoring Ambulatory system using Bluetooth module for fall detection in the elderly. In: Proceedings of the 26th Annual International Conference of the IEEE-EMBS. San Francisco, CA, USA: September 1–5, 2004. [17] Bourke AK, Lyons GM. A threshold-based detection algorithm using a bi-axial gyroscope sensor. Med Eng Phys 2008;30(1):89–90. [18] Fukaya K. Fall detection sensor for fall protection airbag. In: SICE Annual conference, Osaka, August 5–7, 2002. [19] Williams G, et al. A smart fall and activity monitor for telecare applications. In: 20th international Conference IEEE-EMBS 1998, Hong Kong: 29 October–1 November 1998, p. 1151–1154. [20] Doughty, Lewis KR, McIntosh A. The design of a practical and reliable fall detector for community and institutional telecare. J Telemed Telecare 2000;6(Suppl. 1):S150–4. [21] Lindemann U. Evaluation of a fall detector based on accelerometers: a pilot study. Med Biol Eng Comput 2005;43:548–51. [22] Bourke AKVJ, O’Brien J, Lyons GM. Evaluation of a threshold-based triaxial accelerometer fall detection algorithm. Gait Posture 2007;2(2):194–9. [23] Tamura T, et al. An ambulatory fall monitor for the elderly. In: Annual International Conference IEEE Eng Med Biol Soc. 2000. Chicago, IL: 2000, p. 2608–2610. [24] Srinivasan P, Birchfield D, Gang Qian, Kidane A. Design of a pressure sensitive floor for multimodal sensing. In: Proceedings 9th International Conference on Information Visualisation. 6–8 July 2005, p. 41–46. [25] Zhang T, et al. Fall detection by embedding an accelerometer in cell phone and using KFD algorithm. Int J Comp Sci Netw Secur 2006;6: 277–84. [26] Mathie MJ, Basilakis and Celler BG. A system for monitoring posture and physical activity using accelerometers. In: 23rd Annual International Conference IEEE Engineering in Medicine and Biology Society, Istanbul 25–28 October 2001.
Acknowledgements The authors wish to thank Dr. Pierre-Emanuel Colle, senior lecturer and head of the Grenoble Medical school language department, for his extensive corrections of the article.
N. Noury et al. / IRBM 29 (2008) 340–349 [27] Virone G, Noury N, Demongeot J. A system for automatic measurement of circadian activity deviations in telemedicine. IEEE Trans Biomed Eng 2002;49:1463–9. [28] Mihailidis A. An intelligent emergency response system: preliminary development and testing of automated fall detection. J Telemed Telecare 2005;11:194–8. [29] Lord CJ, Colvin DP, Falls in the elderly: Detection and Assessment. In: Annual International Conference. IEEE Eng. Med. Biol. Society. 1991. [30] Diaz A, et al. Preliminary evaluation of a full-time falling monitor for the elderly. In: Proceedings of the 26th Annual International Conference of the IEEE-EMBS, San Francisco, CA, USA. September 1–5, 2004. [31] Noury N, Barralon P, Virone G, Boissy P, Hamel M, Rumeau P. A Smart Sensor Based on Rules and its Evaluation in Daily Routines. In: IEEEEMBC 2003. Cancun-Mexico: 2003. p. 3286–3289. [32] Aharonoff GB, Dennis MG, Elshinawy A, Zuckerman JD, Koval KJ. Circumstances of falls causing hip fractures in the elderly. Clin Orthop Relat Res 1998:10–4.
349
[33] DeVito CA, Lambert DA, Sattin RW, Bacchelli S, Ros A, Rodriguez JG. Fall injuries among the elderly. Community-based surveillance. J Am Geriatr Soc 1988;36:1029–35. [34] Milisen K, Detroch E, Bellens K, Braes T, Dierickx K, Smeulders W, et al. Falls among community-dwelling elderly: a pilot study of prevalence,circumstances and consequences in Flanders. Gerontol Geriatr 2004;35:15–20. [35] Norton R, Campbell AJ, Lee-Joe T, Robinson E, Butler M. Circumstances of falls resulting in hip fractures among older people. J Am Geriatr Soc 1997;45:1108–12. [36] Ryynanen OP, Kivela SL, Honkanen R. Times, places, and mechanisms of falls among the elderly. Gerontol 1991;24:154–61. [37] Thapa PB, Brockman KG, Gideon P, Fought RL, Ray WA. Injurious falls in nonambulatory nursing home residents: a comparative study of circumstances, incidence, and risk factors. J Am Geriatr Soc 1996;44: 273–8.
General review
A proposal for the classification and evaluation of fall detectors Une proposition pour la classification et l’évaluation des détecteurs de chutes N. Noury a,b,∗ , P. Rumeau a,c , A.K. Bourke d,e,f,g , G. ÓLaighin d,e , J.E. Lundy h,i a
Team AFIRM, UMR UJF–CNRS 5525, laboratory TIMC-IMAG, faculté de médecine de Grenoble, centre Jean-Roget, 38706 La Tronche cedex, France b Team MMB, laboratory INL, insa Lyon, 69627 Villeurbanne cedex, France c Hospital La Grave-Casselardit, Toulouse, France d Department of Electronic Engineering, National University of Ireland, Galway, Ireland e National Centre for Biomedical Engineering Science, National University of Ireland, Galway, Ireland f Biomedical Electronics Laboratory, Department of Electronic and Computer Engineering, Faculty of Science and Engineering, University of Limerick, Ireland g Wireless Access Research Group, Department of Electronic and Computer Engineering, Faculty of Science and Engineering, University of Limerick, Galway, Ireland h Hospital Cochin, Emergency Unit, Paris, France i Vigilio s.a., Evry, France Received 10 June 2008; accepted 17 August 2008
Abstract Falls affect, each year, tens of million of elderly people throughout the world. It can have immediate lethal consequences but also causes many disabling fractures and dramatic psychological consequences which reduce the independence of the person. Falls in the elderly is thus a major public health problem. The “early” detection of fall consequently raises the interest of searchers, as most of elderly fallers cannot return to a standing position on their own following a fall. It is also an interesting scientific problem because it is an ill-defined process. The goals of this study were to classify various approaches used to detect the fall and to point out the difficulty to compare the results of these studies, as there is currently no common evaluation benchmark. © 2008 Elsevier Masson SAS. All rights reserved. Résumé La chute est un problème majeur de santé publique qui touche chaque année plusieurs dizaines de millions de personnes âgées de par le monde, avec des conséquences immédiates, mortelles, mais aussi des complications handicapantes, physiques ou psychologiques. Le plus souvent la personne âgée ne peut se relever seule après la chute, aussi faut-il intervenir très rapidement, donc pouvoir détecter cet évènement dans les plus brefs délais. C’est donc un problème qui intéresse les chercheurs et, qui plus est, attise leur curiosité car la chute est un processus mal défini qui s’exprime en une grande variété de situations. L’objet de la présente étude est de clarifier les idées sur ce phénomène en exposant diverses approches contemporaines pour parvenir à sa détection. On montre aussi le défaut de cadre commun d’évaluation. © 2008 Elsevier Masson SAS. All rights reserved.
1. Introduction Falls affect each year one in two people over 80 years of age. It thus concerns tens of million of elderly people throughout the ∗ Corresponding author. UMR UJF-CNRS 5525, TIMC-IMAG Laboratory, Research team AFIRM, Faculty of Medicine, University Joseph-Fourier bâtiment Jean-Roget, 38706 La Tronche, France. E-mail address: [email protected] (N. Noury).
1959-0318/$ – see front matter © 2008 Elsevier Masson SAS. All rights reserved. doi:10.1016/j.irbm.2008.08.002
world; approximately 100 million people in developed countries. In this segment of the population, a fall may have lethal consequences. It also causes many disabling fractures [1] and dramatic psychological consequences, which reduces the independence of the person [2–4]. Falls in the elderly is thus a major public health problem. The fall consequently raises the interest of searchers, and particularly the “early” detection of the fall. Indeed, as most the elderly fallers cannot return to a standing position on their own
N. Noury et al. / IRBM 29 (2008) 340–349
following a fall [5], a relationship was established between the delay before intervention and the morbidity–mortality rate [6,7]. It is therefore important to respond very quickly to a fall. The detection of the fall is also an interesting scientific problem because it is an ill-defined process and can thus be approached using various methods. The goals of this study were to classify the various approaches to detect the fall and to point out the difficulty of comparing the results of these studies, thus the need for a common evaluation benchmark is evident. This article starts with a definition of the fall of the elderly, a discussion on the physics of a fall and its detection, and reviews the main academic and industrial research in this developing field. The authors then suggest a classification of the various approaches. In the last part, a set of evaluation parameters and a proposal of a common evaluation framework for fall detection systems are presented. 2. Definition of a fall detection Everybody has experienced an unwanted fall, whether in childhood while training to walk, or occasionally in adulthood. The fall mechanism is thus well-known to everybody. To face the fall, corrective and protective mechanisms were developed; athletes can even control “high energy” falls. Nevertheless, it is difficult to describe precisely the phenomenon, and even harder to imagine the means for its detection. Obviously, the fall of a person can be described as the rapid change from the upright/sitting position to the reclining or almost lying position. It is not a controlled movement, like lying down, for example. In 1987 the Kellogg international working group on the prevention of falls in the elderly defined a fall as “unintentionally coming to ground, or some lower level not as a consequence of sustaining a violent blow, loss of consciousness, sudden onset of paralysis as in stroke or an epileptic seizure” [8]. This definition was used in many research studies, and later extended to include falls resulting from dizziness and syncope, and other consequences of an epileptic fit or cardiovascular collapse such as postural hypotension and transient ischemic attacks, all of which may have resulted in a fall and its potential consequences. According to the above-mentioned definitions, a fall detector should be able to detect falling to the ground, in an unexpected and uncontrolled way, whatever the cause of this accidental fall. Additionally, it should have the capacity to discriminate the fall from any intentional movement. 3. Material and methods for fall detection The fall may be broken down into four phases (Fig. 1), that is, the prefall phase, the critical phase, the postfall phase and the recovery phase. During the “prefall” phase the person performs usual activities of daily living (ADL), with occasional sudden movements, like sitting or lying down rapidly, which must be distinguished from a fall. The “critical phase” consists in the sudden movement of the body toward the ground, ending with a vertical shock
341
Fig. 1. The four phases of a fall event.
on the ground. The duration of this phase is extremely short (T1 − T0 = 300–500 ms). During the “postfall” phase, the person remains inactive, frequently lying on the ground. The postfall phase shouldn’t last too long (T2 − T1 < 1 h) to reduce the consequences of the fall. Eventually the “recovery” phase is either intentional – the person stands up on his own – or with help from another person. We may classify the different studies on fall detection according to whether they focus on the “direct” detection of the critical phase or the impact shock, and the “indirect” detection of the postfall phase. Some attempted to combine both approaches. 3.1. Early detection of the critical phase of the fall During the critical phase of a fall, there is a temporary period of “free fall” during which the vertical speed increases linearly with time due to gravitational acceleration. Wu ([9] University of Vermont, USA) showed, with a video analysis of markers placed on the subject, that vertical and horizontal speeds are three times higher during a fall than for any other controlled movement. She also showed that both speeds will increase near simultaneously during a fall whereas they are strongly dissimilar during “controlled” movements. Thus, if the vertical and horizontal speed of controlled movements of the person (to rise, to bend down, to sit down) is measured, it is possible to discriminate these speeds from those occurring during a fall, which would exceed a pre-determined threshold. This inspired NaitCharif and Mckenna [10] and Rougier and Meunier [11] who attempted to detect falls by tracking the head movements with computer vision techniques and particles filters algorithms with fixed thresholds on vertical and horizontal velocities. The difficulty was, in this case, to determine this threshold. If it is too low, the system then also detects negative events (“false positive”). If the threshold is too high it will not detect positive events (“false negative”). This threshold is also dependent on the subject-to-subject variability. To overcome this last difficulty, a “learning period” may be used. This learning period may be “supervised” by asking the wearer to carry out a series of voluntary movements in order to “mark” the normal speed of execution. If the learning period is “unsupervised” the movements of the person are simply recorded during a few hours or several days, and then a statistical analysis on the measured speeds is carried out. Depeursinge ([12], CSEM in Switzerland) described a device which, using three orthogonally arranged accelerometers and three successive integrations, could locate in real-time the spatial position of the device and, by training a neural network, could detect unusual events such as the
342
N. Noury et al. / IRBM 29 (2008) 340–349
fall of the wearer. Noury et al. ([13], University of Grenoble in France) designed an autonomous sensor module, attached under the armpit, including accelerometers, an inclinometer and a vibration sensor. The sensor module detects, when the velocity of the movement exceeds a specific threshold, the sequence from a vertical posture to the lying posture, as well as the absence of movements after the fall. Prado et al. [14] developed an intelligent four-axis accelerometer unit (IAU) worn like a patch, fixed to the low back at the sacrum level. Some other studies were reported on the use of triaxial accelerometers placed either at the waist, Mathie et al. [15], or on the chest, for Hwang et al. [16] who also used a gyroscope. Recently, Bourke and Lyons proposed a fall algorithm based on thresholds applied to biaxial gyroscope signals [17]. Eventually, the detection of the very early phase of the fall could allow launching an airbag as proposed by Fukaya [18].
3.2. Direct detection of the end of the critical phase of the fall At the end of the critical phase, the body frequently hits the ground or an obstacle. The mechanical consequence of this is a sudden polarity inversion of the acceleration vector in the direction of the trajectory; this is commonly referred to as the “impact shock”. This event can be detected with an accelerometer or a shock detector; which is actually an accelerometer with a previously determined fixed threshold. Williams et al. ([19] Ireland) first described in 1998 an autonomous device carried on the belt, which featured a piezoelectric sensor to detect the impact shock when the person hit the ground as well as a mercury tilt switch to detect when the person was in a horizontal position. This research was continued by Doughty et al. [20], and a commercial version of this system was developed and marketed by the company Tunstall (Whitley lodge, Yorkshire, England). Lindemann [21] placed a 3D accelerometer external into the housing of a hearing aid. Three trigger thresholds were identified so that a fall could be identified: a sum-vector of acceleration in the xy-plane exceeding 2 g; a sum-vector of velocity of all spatial components right before the impact superior to 0.7 m/s; and a sum-vector of acceleration of all spatial components superior to 6 g. Bourke et al. [22] proposed fall algorithms based on thresholds on triaxial accelerometer signals upon impact. The difficulty of this approach is to determine the direction of the trajectory, which is obviously variable from one fall to another. Actually, most of the falls occur in the “sagittal” plane (forward or backward) as the fall often follows a voluntary movement, which is mainly performed in the saggital plane (to sit down, to rise, to walk, to bend down). A second difficulty is the location of the sensor on the body relatively to the point of impact. Depending on whether or not the sensor is near the point of impact, the “signature” of the signal recorded at the time of the shock can be significantly different and it thus becomes more difficult to recognize a fall when it occurs, thus leading to a significant number of “false positives”.
3.3. Indirect detection of the lying during the postfall As many falls end with the person lying on the ground, the simplest approach is to detect the horizontal position with a tilt sensor (a mercury contact or a ball trapped in a guide). Another approach is to detect when the feet are no longer in contact with the ground. Tamura et al. [23] proposed a photo-switch, which outputs a trigger signal to record the falling time. In a preliminary study, the system was tested for normal adults and hemiplegic patients, without any problem. The above-mentioned methods are more appropriate to monitoring an “isolated worker”, not supposed to work in the lying position. But it is obviously not suitable for the detection of falls of an older person in their home environment because, at home, lying can quite naturally occur even outside normal sleeping hours. Therefore this method is prone to many “false positives”, that is detection as falls of situations, which are not falls. A complementary solution is to detect more precisely if the body is lying on the floor, with sensitive floor tiles installed in all the rooms [24]. But the coverage is rather limited as falls do not always result in the person lying on the ground, or may occur in locations, which are not equipped with the special tiles. Furthermore, this requires a very complex wiring with potential deterioration of the home decoration. 3.4. Indirect detection of the absence of motion during the postfall After a “serious” fall (risk of severe injury), the person frequently remains immobile in a posture and/or a place. The lack of movement can thus be a consequence of a fall event. This can be detected with a basic movement or vibration sensor, placed on one of the extremities of the body, which are more mobile (e.g., wrist or ankle). Zhang et al. [25] produced a fall detection device by placing a triaxial accelerometer in a mobile phone; he proposed that, for a severe fall, the sequence of events is: a daily activity, a fall and then the person remaining motionless. Mathie et al. [26] used a similar method to identify a fall with a triaxial accelerometer. He used a range of parameters including the tilt angle, length of time during which a subject maintained a posture, metabolic energy expenditure, and previous and next activity. He also investigated the effects of three parameters: the length of a smoothing median filter, the width of the averaging window, and the value of the acceleration magnitude threshold. The main difficulty with these approaches is in the choice of the latency time (the delay T2 − T1 before taking a decision), which must be long enough to decrease the number of “false positives”. However, if the latency time is too long, it will result in a longer delay before an intervention is initiated. Lack of movements in some parts of the home (bathroom, toilets) can also be simply detected using presence infrared sensors [27] or video. Mihailidis ([28], University of Toronto, Canada) placed a video camera on the ceiling and developed scene algorithms to detect a fall. Image processing of video signals can also be used to detect a fall by detecting abrupt movements using vector analysis. This last method typically consists of subtracting successive images to keep only the variations, which are then sorted according to their direction and/or
N. Noury et al. / IRBM 29 (2008) 340–349
their amplitude. While these techniques work well in controlled environments (laboratory, scene), they must be adapted in noncontrolled environments in which neither the lighting nor the framing is controlled (it is obviously necessary that the subject be in the field of vision). Moreover, as the subject moves in a threedimensional space, it is also necessary to call upon more complex techniques, namely the use of two cameras (“stereovision”). Nevertheless, these techniques are becoming accessible, both technically and financially, thanks to the emergence of low cost cameras (web cams), the wireless transmission of images over short distances, and the possibility of embedding the required algorithms. But the video technology poses a major problem of acceptance as it requires the placement of video cameras in private living quarters, and especially in the bedroom and the bathroom, with consequent concerns about privacy. Most of the approaches attempted to produce an analytical model of the fall from knowledge or by reasoning from rules. Furthermore, a more “intuitive” approach can also be used with the development of fall detection systems based on machine learning with an observation phase (a training period) and then a classification phase. It starts with the definition of a set of classification criteria, significant and independent (discriminating). The training period may be supervised. A neural network can be trained, and then used to automatically classify future situations. Only the situations met during training can be classified, the others being mixed in a class labelled “others” (to stumble, to slip, etc.). If the training is “unsupervised”, a class “fall” is likely to be detected if the training period is long enough, even infinite if the event of fall is rare. Moreover, the first fall event is likely to be missed since its class is unknown before its first occurrence.
343
Table 1 Terminology for the classification of fall detectors. Direct detection
Detection of a sudden postural variation
Accelerometric
Detection of sudden movements and shocks to the body Detection of body height above ground level Detection of body spatial position Video monitoring of the body
Statimetric Topometric Eidolimetric Peripheral detection
Detection of behavioral/contextual variations
Behavioral Geotaximetric
Monitoring of behavioral sequences Monitoring of spatial localization of the person Monitoring of the body orientation (verticality) Monitoring of the body movements Monitoring of vital signs
Orthostatimetric Kinesimetric Physiometric Combinatorial Detection
Combination of direct and peripheral detection
Bidimensional Tridimensional Multidimensional
One direct and one peripheral parameter One direct and two peripheral parameters One direct and three or more peripheral parameters
reason is the rate of “false alarms”, resulting in inappropriate alerts of the remote monitoring service, leading to the “rejection” of the equipment by the remote supervisors. It should thus be stressed that there is currently no satisfactory solution on the market and that fall detection should be better defined.
4. Industrial developments and commercial products 5. Evaluation of the fall sensors Some of the previous results were successfully implemented in functional prototypes, and an extensive range of patents has been filed in the fall sensor area. In a few cases, final products were built and are available on the market place. For the sake of clarity, we first propose a classification of the patents with a specific terminology (Table 1). Using our terminology, we attempted to classify most of the existing patents (Table 2). We observed a regular increase in the submission of patents, beginning in 1996, with a peak (11 patents) in 2002, and now stabilized (4 patents in 2006). Most of the products available on the market are worn on the wrist or on a belt (Fig. 2). Other devices, also on the market, offer alternative solutions by “peripheral” detection of a fall (Fig. 3). In spite of the need for technological interventions to face the fall problem, and several devices being available in the market place, there is no significant industrial development of fall sensors and little use of these devices in daily geriatric practice. There are probably multiple reasons. Some devices exhibit low performance on the field. Some are not accepted by the users due to inadequate ergonomics or due to the stigmatization of the fragility of the old person. But the major
Most of the academic studies started with the design of a test instrument in order to record signals, or images, during simulated situations of fall/non fall. Eventually, they suggested algorithms, which were tested off-line on recorded signals/images. But few studies reported result data and conditions of assessment. Lord and Colvin [29] reported in 1991 some early studies on fall detection using accelerometers with very little details. Nait-Charif [10] and Rougier and Meunier [11] announced the possibility to detect a fall. The IAU of Prado et al. [14] was further evaluated by Diaz et al. [30] in a laboratory study carried out on eight volunteers to show that the device was able to distinguish true falling events from normal activities like fast walking or going up/downstairs. The prototypes made by Williams et al. [19] and Doughty et al. [20] were evaluated on 20 people to observe “false positive alarms”. The final design allowed for a reliable detection in 180 different falling scenarios. The algorithm developed by Lindemann [21] allowed discriminating ADL from intentional falls. The system made by Mihailidis was tested on 21 volunteers who carried out simulated falls and was capable of detecting 77% of falls. Mathie et al. [15] collected data from 26 normal subjects performing
344
N. Noury et al. / IRBM 29 (2008) 340–349
Table 2 Classification of the patents for fall detectors. Direct detection Accelerometric
Monoaxial Multiaxial
Statimetric Topometric Eidolimetric
Peripheral detection Behavioral
Geotaximetric
Orthostatimetric Kinesimetric
Physiometric Combinatorial detection Bidimensional analysis
Accelerometry + time without movement
Accelerometry + shock Accelerometry + image Accelerometry + physiological parameter Accelerometry + geolocalisation
Accelerometry + Δ accelerometrics
Tridimensional analysis
Multidimensional analysis
Accelerometry + time without movement + geolocalisation Accelerometry + cardiac ballistics + time without movement Accelerometry + cardiac ballistics + time without movement + data fusion
P. WO 2004/086325 A2 (07/10/04) MurataM. US 2004/0046675 A1 (11/03/04) Franco T.S. Jr US 2005/0093709 A1 (05/05/05) Cederstrom E. WO 2004/114245 A1 (29/12/04) Karner F.C. US 6,965,311 B1 (15/11/05) Fourniquet A. 2 870 378 (17/05/04) Thacker N. WO 01/63576 A2 (30/08/01) Fredriksson A. WO 2004/047039 A1 (03/06/04)
Aphycare WellCare Systems Inc IST EDF University Manchester Wespot A.B.
Coulthard J.J. US 6,972,677 B2 (06/12/05) Esteve D. EP 1 482 464 A1 (01/12/04) Power M.W. US 2002/0060630 A1 (23/05/02) Esteve D. 2 828 317 (01/08/01) Markhovsky R. US 2006/0012476 A1 (19/1/2006) Burbaud T. 2 857 140 (23/06/03) Tallman E. US 6,175,308 B1 (16/01/01) Kelly P.B. Jr WO 01/50957 A1 (19/01/01) Najafi Bijan EP 1 195 139 A1 (10/04/02) Crisco III J.J. US 2002/0060633 A1 (23/05/02) Brilman A.J. US 6,975,230 B1 (13/12/05) Bader G. US 6,337,629 B1 (08/01/02)
EDF
Roussy C.A. WO 2004/100092 A2 (18/11/04) Holdsworth D. EP 1 128 49 A1 (29/08/01) Cadet P.H. WO 98/29852 (31/12/96) Lepaul G. 2 760 116 A1 (26/02/97) Myllymaki M. WO 99/56262 A1 (04/11/99) Burbaud T. EP 1 632 920 A1 (08/03/06) Fulton J.G. US 6,377,179 B1 (23/04/02) Cosquer P. WO 01/85025 A1 (19/11/01)
DHS Tunstall
Depeursinge Y. EP 0 877 346 A1 (11/11/98) Depeursinge Y. US 6,201,476 B1 (13/03/01) Jacobsen S.C. US 6,160,478 (12/12/00) Jacobsen S.C. 2 785 073 A1 (26/10/99) Petelenz T.J. US 6,433,690 B2 (13/08/02) Lehrman M.L. WO 01/20571 A1 (15/09/00) Lehrman M.L. US 6 ?864,796 B2 (08/03/05) Lehrman M.L. WO 02/45043 A1 (06/06/02) Lehrman M.L. US 2002/0118121 A1 (29/08/02) Lehrman M.L. WO 02/061704 A1 (08/08/02) Lehrman M.L. WO 02/073564 A1 (19/09/02) Lehrman M.L. US 2002/0008630 A1 (24/01/02) Azcoita Arr. J.M WO 2006/000605 (05/01/06)
CSEM CSEM
Noury N. 01/12046 (18/09/01)
UJF Grenoble
Noury N. 06/08584 (29/09/06)
Vigi’Fall (Vigilio, UJF Grenoble)
sit-to-stand and stand-to-sit transitions and walking. His system successfully distinguished between activity and rest, giving sensitivity greater than 0.98 and specificity between 0.88 and 0.94. Noury et al. [31], in an early test, asked five volunteers to perform 15 fall scenarios five times, and obtained a sensitivity and specificity reaching 85%. Bourke et al. [17,22] announced a performance of 100% from a test involving 10 healthy young subjects performing simulated falls on large crash mats and elderly healthy subjects performing normal ADL.
Actall Corp.
Noc’Watch/fall Saver EPFL Lausanne Barlow, Joseph & Homes Ltd. Biosys AB
Inst. Polytechnique Sevenans IST Vital base Univ. Rennes
Sarcos Inc. ILife Systems Inc. ILife Systems Inc. ILife Systems Inc. ILife Systems Inc. ILife Systems Inc. ILife Systems Inc. ILife Systems Inc.
We can say that the proposed solutions and associated devices for fall detection in the elderly are numerous but their performances are difficult to compare. For the sake of clarity, it is therefore necessary to define a common set of criteria to evaluate the performance of each system. Obviously the first intrinsic “performance” of fall sensors is their ability to detect the fall, and only the fall. Unfortunately, this information is frequently unavailable in the literature. Furthermore, it is interesting to compare other objective criteria, such as the detection
N. Noury et al. / IRBM 29 (2008) 340–349
345
Fig. 2. Some commercially available fall detectors.
method, the “intrusivity” and the lifespan of the device; the last criteria are related to the acceptance so their selection may be subjective because selection depends on the aims of the researcher/engineer/marketing agent. 5.1. Quality criteria for the fall detector Fall detection is either positive if the detector properly recognizes a fall event, or negative if it does not. As the output is binary, the quality of the detector cannot be evaluated simply from a single test, instead, it is necessary to carry out a statistical analysis on a series of tests. There are four possible situations: • true positive (TP): a fall occurs, the device detects it; • false positive (FP): the device announces a fall, but it did not occur; • true negative (TN): a normal (no fall) movement is performed, the device does not declare a fall;
• false negative (FN): a fall occurs but the device does not detect it. Two criteria are proposed to evaluate the response to these four situations: • sensitivity is the capacity to detect a fall Eq. (1): sensitivity =
TP ; TP + FN
(1)
• specificity is the capacity to detect only a fall Eq. (2): specificity =
TN . TN + FP
(2)
5.2. Criteria for the evaluation of a fall sensor In addition to the intrinsic performance of the fall detector, other criteria must be taken in consideration when comparing
346
N. Noury et al. / IRBM 29 (2008) 340–349
Fig. 3. Commercial devices for “behavioral” detection of the fall.
the different approaches. Such criteria relate to the form factor, the acceptance, the usability, and the level of “accomplishment”. The form factor (mass and volume of the sensor) is particularly important if the sensor is carried by the wearer. The acceptance of the sensor is usually lower if the sensor is carried on the person (endosensor) rather than installed in his environment (exosensors). It is also influenced by the “visibility”: a visible sensor (on the wrist, the belt, the chest. . .) may stigmatize the handicap or dependence. This can also be the case with external sensors visibly placed in the domestic environment. The user friendliness depends on the autonomy, the range, and the packaging. The autonomy is the estimated time between two successive maintenance interventions, that is, to change/reload the batteries. The range is the maximum distance between the person and the sensor so that the latter can efficiently perform measuring. In the case of a carried sensor, communicating wirelessly, it is the distance between the base station and the sensor. The type of packaging concerns the biological innocuity, for instance biocompatibility, or the risk of injury for sensors directly in contact with the skin. The level of “accomplishment” was added because some promising systems are still being tested while others are already available on the market place. We tried to evaluate (Table 3) some of the most widely used devices according to these criteria. For the sake of clarity, we
gave additional information on the principle of operation, the location of the sensor and the evaluation of the performance (duration, number of test subjects). 5.3. Protocol for the evaluation of fall sensor performance It is currently practically impossible to compare the performances of various fall sensors based on literature data, as common criteria for their evaluation were not used and no common procedures were used to perform the tests. Therefore, it is very important to propose a common framework for the evaluation of the fall sensors. There are many scenarios of falling and it is not possible to consider all of them in an exhaustive way. The device must be tested in a limited number of situations, which are sufficiently representative of falls (positive situations), as well as of “pseudo” falls (negative situations). The first situation to be considered is a fall in the anteroposterior plane. As already mentioned, most falls occur during intentional movements, initiated by the person in the anteroposterior plane, forward or backward; a large number of falls occur after stumbling on an obstacle while walking, when slipping backwards on wet ground, or during a “stand-to-sit” transfer. If a person looses her/his balance in the forward direction, s/he will initially try to recover by taking some steps forward, thus amplifying the movement of the fall, and s/he will perhaps finally fall while projecting her/his arms forwards for protection. S/he can also drop to her/his knees. If loss of balance occurs backwards,
N. Noury et al. / IRBM 29 (2008) 340–349
347
Table 3 Classification of existing devices. Name
Principle
Location
Form factor
Functions
Level of evaluation
Sereoz (Aphycare, France)
Accelerometric (shock)
Wrist
Accelerometric + geolocalisation
Chest
Ynolis (“Ecureuil”, DHS, France) Noc Watch “Fall Saver” (USA)
Accelerometric + motionless time Orthostatimetric (Orientation of corporal segments)
Wrist Thigh
Watch 18 g Credit Card Tegaderm Patch
Call Button Heart Beat Surface body temperature Cancellation button Analysis of activity Heart beat Detections of standing position
Two years in institution
FDSS (csem, Switzerland)
Watch 30 g Hypoallergenic polymer 10 × 5 × 2 cm
ACTIM3D (Labo TIMC, France)
Accelerometric + motionless time
Chest (under the arm pit)
15 g 5 × 3 × 1 cm Hypoallergenic polymer Six months autonomy (on a 3 V coin cell battery)
the person will try to sit down to possibly limit the intensity of the shock impact. In some cases, the fall occurs sideways, either during a badly controlled “sit-to-stand” transfer, or if the person, when loosing balance, tries to grip the wall. Eventually, some daily life movements may present a similar amplitude or intensity to the one encountered in accidental situations, for example the action of lying down, or of sitting down, if carried out “quickly”. It is also possible to observe situations of fall initiation with recovery (stumbling). We suggest (Table 4) a set of 20 scenarios for the evaluation of fall sensors, with 50% “negative” and 50% “positive”, based on the above discussion, our experience in evaluating the Noury et al. fall sensor [31], completed by some scenarios used by Bourke et al. to evaluate his fall algorithms [17,22]. It is statistically insufficient to have only one test for each scenario for each subject. However, to avoid making excessive demands on the subjects, a maximum of three tests could be undergone by each subject in each condition. The subject should be allowed to rest as much as he wishes, when he wishes, during the tests. It should be noted that it is our experience that repetition of the same scenario results in “adaptation” to the gesture, which thus becomes less natural, it would be beneficial to vary the order of the tests, or to let the subject choose the order, only recording the order he selected. During the test, the subject will be free to adapt her/his speed in following any of the predetermined scenarios. As a consequence, s/he will have to keep her/his eyes opened. This will also enable her/him to perform self-protection strategies. As the subject carries out three times each scenario, we would collect 60 tests per subject. With a reduced sample of 10 subjects (the sample should also respect gender), that would amount to 600 data points, which is a statistically acceptable sample, significant enough to compute the specificity and the sensitivity of the device.
Analysis of activity
96% true positives
47 subjects (evaluation during 1 week) 21 subjects in lab Specificity = 94% Sensitivity = 95%
Table 4 Scenarios for the evaluation of fall detectors. Category
Name
Kind
Backward falla
Ending sitting Ending lying Ending in lateral position With recovery On the knees With forward arm protection Ending lying flat With rotation, ending in the right lateral position With rotation, ending in the left lateral position With recovery Ending lying flat With recovery Ending lying flat With recovery Vertical slipping against a wall finishing in sitting position Sitting down on a chairb then standing up Lying down on the bed then rising up Walking a few meters Bending down, picking something up on the floor, then rising up Coughing or sneezing
Positive Positive Positive Negative Positive Positive Positive Positive
Forward falla
Lateral fall to the righta Lateral fall to the lefta Syncope Neutral
Positive Negative Positive Negative Positive Negative Negative Negative Negative Negative Negative
Negative
a
The test may be performed starting with both legs straight or with knee flexion. b The height and kind of chair could also lead to separate test conditions.
Although the goal of a fall sensor is to detect the fall of elderly people, it is actually difficult to test the fall situations with them. Thus fall situations may be simulated by younger persons, or even athletes, and normal activities may be tested on elderly in the age group at risk for falls.
348
N. Noury et al. / IRBM 29 (2008) 340–349
6. Conclusion
References
The ideal fall detection system should exhibit both sensitivity and specificity reaching 100%. This was sometimes reached in experimental set ups [17,22], but when developed as an autonomous integrated fall sensor, there is a dramatic loss in performances [28,31]. No individual method seems satisfactory, it thus seems appropriate that a combination of several methods would allow for a greater selectivity. Also, a multidimensional approach, combining both kinematics and physiological parameters, could probably come closer to the 100% sensitivity and specificity target. New promising techniques could also be investigated in considering the fall as a chaotic event with stable states and “bifurcations” of “complex systems”. On the technological side, the main improvements will come with the miniaturization of the devices (integration) and reduction in the level of maintenance required. The smaller the size of the sensor, the easier it will be to fit on the person, in garments or in accessories. Also, the maintenance frequency should ideally reach one or two years to reduce maintenance and associated costs. The functionality of these devices should also be improved. First of all, the activation must be fully automatic with no need of carrier intervention. Secondly, the device with communication features could bring enhanced services to the person, such as the provision of a “social link” which would improve the conventional alert system. Daily use of an intelligent device also introduces ethical issues concerning the respect of intimacy and privacy (each movement of the person can be recorded) and also the risk of subject dependency on technology. But for the time being, the most urgent task facing the scientific community is to accept a common definition of a fall and of fall detection, and to agree on a common protocol for the evaluation of fall detection systems such as the one suggested in this study. This article focuses on the fall detection methods and devices because getting help quickly after a fall reduces risk of hospitalization by 26% and death by over 80%. Nevertheless, there are obviously several advantages offered by the analysis of the environmental/behavioral/physiological sources of the fall, in order to develop prevention systems or actions aiming at reducing the intrinsic/extrinsic causes of the fall [32–37]. Some falls are preventable when caused by environmental factors (like fall hazards at home) and by side effects of medication because they can be foreseen and avoided. Preventing falls and the resulting injuries can promote independence, by reducing or delaying the need to move out of home.
[1] Sadigh S, Reimers A, Andersson R, Laflamme L. Falls and fall-related injuries among the elderly: a survey of residential-care facilities in a Swedish municipality. J Community Health 2004;29:129–40. [2] Gostynski M. Prevalence, circumstances and consequences of falls in institutionalized elderly; a pilot study. Soz Praventivmed 1991;36:341–5. [3] Ryynanen OP, Kivela SL, Honkanen R, Laippala P. Falls and lying helpless in the elderly. Z Gerontol 1992;25:278–82. [4] Kiel DP. Falls. RI Med 1991;J74:75–9. [5] Alexander NB, Ulbrich J, Raheja A, Channer D. Rising from the floor in older adults. J Am Geriatr Soc 1997;45:564–9. [6] Gurley RJ, Lum N, Sande M, Lo B, Katz MH. Persons found in their homes helpless or dead. N Engl J Med 1996;334:1710–6. [7] Wild, Nayak DUS, Isaacs B. How dangerous are falls in old people at home? Br Med J (Clin Res Ed) 1981;282:266–8. [8] Gibson MJ, et al. The prevention of falls in later life. A report of the Kellogg International Work Group on the Prevention of Falls by the Elderly. Danish Medical Bulletin 1987;34(Suppl. 4):1–24. [9] Wu G. Distinguishing fall activities from normal activities by velocity characteristics. Jour of Biomechanics 2000;33:1497–500. [10] Nait-Charif H, Mckenna S. Activity Summarisation and Fall Detection in a Supportive Home Environment. In: 17th International Conference on Pattern Recognition, 2004. [11] Rougier C, Meunier J. Demo: Fall Detection Using 3D Head Trajectory Extracted From a Single Camera Video Sequence. In: First International Workshop on Video Processing for Security (VP4S-06), June 7–9. Quebec City, Canada: 2006. [12] Depeursinge Y. Device for monitoring the activity of a person and/or detecting a fall. 2001. [US patent 6,201,476 B1]. [13] Noury N, et al. Monitoring behavior in home using a smart fall sensor and position sensors. In: IEEE-EMBS Microtechnologies in Medicine & Biology, Lyon France: 2000. p. 607–610. [14] Prado MJ, Reina-Tosina, Roa L S.Distributed intelligent architecture for falling detection and physical activity analysis in the elderly. 2002. [15] Mathie MJ, et al. Classification of basic daily movements using a tri-axial accelerometer. Med Biol Eng Comput 2004;42:679–87. [16] Hwang JY, et al. Development of Novel Algorithm and Real-time monitoring Ambulatory system using Bluetooth module for fall detection in the elderly. In: Proceedings of the 26th Annual International Conference of the IEEE-EMBS. San Francisco, CA, USA: September 1–5, 2004. [17] Bourke AK, Lyons GM. A threshold-based detection algorithm using a bi-axial gyroscope sensor. Med Eng Phys 2008;30(1):89–90. [18] Fukaya K. Fall detection sensor for fall protection airbag. In: SICE Annual conference, Osaka, August 5–7, 2002. [19] Williams G, et al. A smart fall and activity monitor for telecare applications. In: 20th international Conference IEEE-EMBS 1998, Hong Kong: 29 October–1 November 1998, p. 1151–1154. [20] Doughty, Lewis KR, McIntosh A. The design of a practical and reliable fall detector for community and institutional telecare. J Telemed Telecare 2000;6(Suppl. 1):S150–4. [21] Lindemann U. Evaluation of a fall detector based on accelerometers: a pilot study. Med Biol Eng Comput 2005;43:548–51. [22] Bourke AKVJ, O’Brien J, Lyons GM. Evaluation of a threshold-based triaxial accelerometer fall detection algorithm. Gait Posture 2007;2(2):194–9. [23] Tamura T, et al. An ambulatory fall monitor for the elderly. In: Annual International Conference IEEE Eng Med Biol Soc. 2000. Chicago, IL: 2000, p. 2608–2610. [24] Srinivasan P, Birchfield D, Gang Qian, Kidane A. Design of a pressure sensitive floor for multimodal sensing. In: Proceedings 9th International Conference on Information Visualisation. 6–8 July 2005, p. 41–46. [25] Zhang T, et al. Fall detection by embedding an accelerometer in cell phone and using KFD algorithm. Int J Comp Sci Netw Secur 2006;6: 277–84. [26] Mathie MJ, Basilakis and Celler BG. A system for monitoring posture and physical activity using accelerometers. In: 23rd Annual International Conference IEEE Engineering in Medicine and Biology Society, Istanbul 25–28 October 2001.
Acknowledgements The authors wish to thank Dr. Pierre-Emanuel Colle, senior lecturer and head of the Grenoble Medical school language department, for his extensive corrections of the article.
N. Noury et al. / IRBM 29 (2008) 340–349 [27] Virone G, Noury N, Demongeot J. A system for automatic measurement of circadian activity deviations in telemedicine. IEEE Trans Biomed Eng 2002;49:1463–9. [28] Mihailidis A. An intelligent emergency response system: preliminary development and testing of automated fall detection. J Telemed Telecare 2005;11:194–8. [29] Lord CJ, Colvin DP, Falls in the elderly: Detection and Assessment. In: Annual International Conference. IEEE Eng. Med. Biol. Society. 1991. [30] Diaz A, et al. Preliminary evaluation of a full-time falling monitor for the elderly. In: Proceedings of the 26th Annual International Conference of the IEEE-EMBS, San Francisco, CA, USA. September 1–5, 2004. [31] Noury N, Barralon P, Virone G, Boissy P, Hamel M, Rumeau P. A Smart Sensor Based on Rules and its Evaluation in Daily Routines. In: IEEEEMBC 2003. Cancun-Mexico: 2003. p. 3286–3289. [32] Aharonoff GB, Dennis MG, Elshinawy A, Zuckerman JD, Koval KJ. Circumstances of falls causing hip fractures in the elderly. Clin Orthop Relat Res 1998:10–4.
349
[33] DeVito CA, Lambert DA, Sattin RW, Bacchelli S, Ros A, Rodriguez JG. Fall injuries among the elderly. Community-based surveillance. J Am Geriatr Soc 1988;36:1029–35. [34] Milisen K, Detroch E, Bellens K, Braes T, Dierickx K, Smeulders W, et al. Falls among community-dwelling elderly: a pilot study of prevalence,circumstances and consequences in Flanders. Gerontol Geriatr 2004;35:15–20. [35] Norton R, Campbell AJ, Lee-Joe T, Robinson E, Butler M. Circumstances of falls resulting in hip fractures among older people. J Am Geriatr Soc 1997;45:1108–12. [36] Ryynanen OP, Kivela SL, Honkanen R. Times, places, and mechanisms of falls among the elderly. Gerontol 1991;24:154–61. [37] Thapa PB, Brockman KG, Gideon P, Fought RL, Ray WA. Injurious falls in nonambulatory nursing home residents: a comparative study of circumstances, incidence, and risk factors. J Am Geriatr Soc 1996;44: 273–8.