Regenerative and Technological Section / Original Paper Received: August 19, 2013 Accepted: April 8, 2014 Published online: August 13, 2014
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
Sensitivity and False Alarm Rate of a Fall Sensor in Long-Term Fall Detection in the Elderly Maarit Kangas a, c Raija Korpelainen a–d Irene Vikman e Lars Nyberg e Timo Jämsä a, c
a
Department of Medical Technology, Institute of Biomedicine, b Institute of Health Sciences and c Medical Research Center Oulu, Oulu University Hospital, University of Oulu, and d Department of Sports and Exercise Medicine, Oulu Deaconess Institute, Oulu, Finland; e Department of Health Sciences, Luleå University of Technology, Luleå, Sweden
Abstract Background: About a third of home-dwelling older people fall each year, and institutionalized older people even report a two- or threefold higher rate for falling. Automatic fall detection systems have been developed to support the independent and secure living of the elderly. Even though good fall detection sensitivity and specificity in laboratory settings have been reported, knowledge about the sensitivity and specificity of these systems in real-life conditions is still lacking. Objective: The aim of this study was to evaluate the long-term fall detection sensitivity and false alarm rate of a fall detection prototype in real-life use. Methods: A total of 15,500 h of real-life data from 16 older people, including both fallers and nonfallers, were monitored using an accelerometry-based sensor system with an implemented fall detection algorithm. Results: The fall detection system detect-
© 2014 S. Karger AG, Basel 0304–324X/14/0611–0061$39.50/0 E-Mail [email protected] www.karger.com/ger
ed 12 out of 15 real-life falls, having a sensitivity of 80.0%, with a false alarm rate of 0.049 alarms per usage hour with the implemented real-time system. With minor modification of data analysis the false alarm rate was reduced to 0.025 false alarms per hour, equating to 1 false fall alarm per 40 usage hours. Conclusion: These data suggest that automatic accelerometric fall detection systems might offer a tool for improving safety among older people. © 2014 S. Karger AG, Basel
Introduction
The proportion of the older population aged over 65 years is growing rapidly in most countries. About a third of home-dwelling older people fall each year [1] and fall rate estimates range between 517 and 683 falls per 1,000 person years [2, 3]. Institutionalized older people present a two- or threefold higher rate of falling compared with the community-dwelling population [2, 4]. Between 10 and 20% of older people fall recurrently [2, 5, 6]. The proMaarit Kangas, PhD Department of Medical Technology, Institute of Biomedicine University of Oulu, PO Box 5000 FI–90014 Oulu (Finland) E-Mail maarit.kangas @ oulu.fi
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
Key Words Accelerometer · Fall detector · Specificity · Older people · Frequent faller
62
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
alarm rate of a fall detector prototype in real-life conditions with 16 older people. Minor modifications of the fall detection algorithm were also tested to further improve the specificity of the system. Materials and Methods Study Subjects and Data Collection People living in elderly care units were recruited to wear a prototype of an automatic fall detector. The study protocol was approved by the Ethical Review Boards in Oulu, Finland (39/2009) and in Luleå, Sweden (2443-2009). The participants and their close relatives received oral information about the study, and written informed consent was obtained. The inclusion criteria for the study included subjects being older than 65 years and possessing the ability to stand independently or with the help of one person. The field test was performed over a 10-month period with a sliding starting and ending point for each care unit, thus resulting in various lengths of individual test periods. Cognitive functions with the mini-mental state examination (MMSE) [27], Katz Index of Independence in Activities of Daily Living [28], 10-meter walking speed indoors, and mobility with Timed Up and Go (TUG) [29] were measured before the test period. The test population included 13 females and 3 males with an average age of 88.4 ± 5.2 years (n = 16), an MMSE of 13.13 ± 8.22 points (n = 13), a walking speed of 0.56 ± 0.31 m/s (n = 14) and a TUG score of 33.4 ± 33.8 s (n = 13). The status of falling 6 months before and during the test period was assessed from the reports of the care personnel. Fall Detector The test subjects wore the prototype of the accelerometrybased fall detector unit described earlier [22]. The device was designed to activate and collect acceleration data when the acceleration of all three axes fell below a threshold of 0.75 g. The block of around 14 s included buffered data before the activation and the collected data after that. After data storage activation, the implemented fall detection algorithm detected impact from an acceleration sum vector based on a threshold of 2 g and an end posture based on the vertical axis acceleration, with minor modifications (CareTech AB, Sweden) from our earlier studies [24, 26]. In the case of a detected fall, the collected acceleration data were labeled with a fall message and the data were transmitted further to the research database using IP-based technology. In a case of a fall message, an SMS message was sent to the care unit. In addition to the implemented fall detection, the sensor continuously collected activity data obtained as a number of vertical acceleration peaks above the threshold of 1.19 g in 1-min epochs (CareTech AB). The system transmitted the activity data using the same protocol used for transmitting acceleration data. Activity data were used to evaluate the usage of the sensor. The sensor (size 5 × 3 × 1.5 cm) was attached with a clip into a pocket on an elastic belt, except for 1 user where a textile pocket on the trouser waistband was used to fix the vertical axis of the device. The attachment site was at the waist, around the anterior superior iliac spine, but some variation was accepted due to the nature of the long-term field test.
Kangas/Korpelainen/Vikman/Nyberg/ Jämsä
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
portion of recurrent fallers increases with age, resulting in almost half of the home-dwelling elderly over 85 years being recurrent fallers [7, 8]. Fear of not being able to recover after falling is common among older people [9]. Approximately half of fallers have been incapable of recovery to a standing position unaided [10]. To prevent long lies, commercially available personal emergency response systems provide applications to call for help. However, around 80% of older people wearing these systems who are unable to recover after a fall do not use their alarm system to call for help [11, 12]. Some commercially available automatic fall detection systems exist, as summarized by Noury et al. [13]. Even though good fall detection sensitivity and specificity in laboratory settings have been reported [14–17], knowledge on the sensitivity and specificity, as well as the acceptability and usability, of these systems in real-life conditions is still lacking. Tamura [18] reported 19 real-life falls detected with an algorithm monitoring acceleration characteristics from a Parkinson’s patient, but no false alarm rate evaluations were presented. In some studies, false fall alarms have been evaluated in real-life conditions, but no falls have occurred during the test period. A vest-worn fall detection system generated 42 false alarms during 833 real-life monitoring hours in older people [19]. In another study, a waist-worn fall detection system was tested for 52 h among older people [20]. A difference between laboratory and real-life samples was detected, since the algorithms resulting in 100% specificity with instructed activities of daily living resulted in a false alarm rate between 0.04 and 0.21 alarms per hour with data collected from real-life occurrences [20]. However, self-initiated, intentional falls may differ from sudden, unexpected falls [21]. Recent reports have indicated that intentional falls in a laboratory environment have both similar and different features compared with reallife falls among older people [22, 23], thus indicating the importance of evaluating fall detection systems in reallife conditions. We have previously validated a fall detection system based on a waist-worn accelerometer with an algorithm for impact and end posture detection [24, 25]. This system has achieved a fall detection sensitivity of 97% and a specificity of 100% in a laboratory setting with intentional falls and instructed activities of daily living from middle-aged test persons [26]. Since knowledge of the sensitivity and specificity of fall detection systems in real-life conditions is still lacking, the aim of this study was to evaluate the long-term fall detection sensitivity and false
Table 1. Use of the fall detector and false alarm rate in a study population of 16 elderly subjects: original fall detection algorithm (raw
data) and modified algorithms (Mod 1–Mod 3) Person
13 21 31 41 51 62 73 81 92 103 112 122 132 143 153 163 Total
Days, n
Alarms, n
worn
not worn
148 127 33 34 5 69 155 61 14 135 17 72 11 118 40 66 1,105
13 15 18 48 0 22 24 13 14 30 0 1 19 25 59 6 306
21 45 65 13 2 43 124 44 0 132 0 105 3 118 22 11 748
Usage hours
False alarm rate (alarms/usage hour)
total, h
average ± SD, h/day
raw data
Mod 1
Mod 2
Mod 3
1,667 1,232 359 265 25 956 3,179 909 165 2,407 197 873 120 1,688 361 1,097 15,500
11.5±2.6 9.8±2.8 10.9±6.1 7.8±2.9 5.0±2.6 8.6±7.1 20.6±4.0 15.0±6.1 11.4±6.2 18.4±6.2 10.9±4.2 12.1±1.3 10.9±3.8 14.1±4.9 9.0±2.0 15.5±5.7 14.2±6.3
0.013 0.037 0.181 0.050 0.080 0.045 0.040 0.048 0.000 0.055 0.000 0.121 0.025 0.070 0.061 0.010 0.049
0.004 0.011 0.136 0.015 0.000 0.029 0.034 0.021 0.000 0.027 0.000 0.086 0.025 0.041 0.011 0.006 0.030
0.010 0.033 0.150 0.038 0.080 0.035 0.029 0.042 0.000 0.035 0.000 0.096 0.025 0.056 0.053 0.007 0.037
0.005 0.011 0.114 0.011 0.000 0.023 0.025 0.021 0.000 0.173 0.000 0.079 0.025 0.036 0.008 0.005 0.025
Falls, n
2 2 2 1 1 1 9 04 04 04 0 0 0 0 0 0 18
Days: technical failure days excluded. Alarms: nonuse days excluded. Average usage hours/day: nonuse days excluded. Falls: number of falls reported during the test period; for detected falls see table 2. Data modifications Mod 1–3 explained in Material and Methods. Superscript numbers 1–3 indicate status at the end of the test period. 1 Dead or withdrawn due to decline in health status. 2 Withdrawal due to other reasons. 3 Completed the test period. 4 Did not fall during the test period but had a fall history before the test period.
• Mod 3 was calculated by combining modifications Mod 1 and Mod 2. To calculate the fall detection sensitivity, falls that occurred when the sensor was worn were included. Fall detection sensitivity was calculated as the percentage of falls correctly identified as falls by the system.
Analyses of Data Real-life falls were identified based on the routine reporting by the care personnel. Based on the time of the fall event, the acceleration database was analyzed to find the corresponding acceleration data. Care personnel were asked to report alarms and categorize them as fall alarms or false fall alarms. The number of fall alarms was collected from the database based on the fall label generated by the implemented fall detection system. For usage hour analyses, inactivity periods exceeding 3 h were excluded during the daytime. At nighttime, if an indication of any activity was seen, all hours were included. The false alarm rate was calculated by dividing the number of false alarms by the number of usage hours during the test period. Usage time (in hours) for 1 false alarm to occur is also reported. Average values for the test population were weighted by the usage hours. The false alarm rate was first calculated based on the raw data. The effect of minor data analysis procedures was studied offline using three different data modifications (Mod 1–Mod 3): • Mod 1 was calculated by excluding the false fall alarms related to the attachment or taking off the sensor system. This modification was performed by deleting the 15-min time period from the beginning and from the end of the daily usage period if available. • Mod 2 was calculated by excluding multiple false fall alarms in a short time period. This modification was performed by combining alarms clustered in 1-min sliding epochs.
The total length of the test period was 1,540 days, consisting of 1,105 days when the fall detection system was used by the test persons and 306 days when the sensor was not used. A total of 129 days were discarded from analyses due to technical reasons (table 1). The sensor system was mostly used during waking hours but some test persons used the system also at nighttime. Overall, 6 test persons completed the test period; 10 withdrew from the study, 5 due to death or a decline in health status and 5 for other reasons (table 1). During the test period, 18 falls were reported to have occurred among 7 of the test persons (table 1); 3 falls oc-
Fall Detection Sensitivity and False Alarm Rate
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
63
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
Results
Table 2. Description and detection of the 15 falls of the study subjects during the monitoring period
Person No./fall No.
Time
Description
Fall detection
1/1
02:00
The person was trying to get up from her bed but the blanket was entangled around her legs; she slid on the floor; no injuries reported
Detected
1/2
19:20
The person was found lying on the floor of her room after a fall; she had gotten entangled in a blanket when moving from the bed; resulted in bruises in the left side rib cage
Detected
2/3
12:35
The person was found on the floor near the toilet seat; resulted in bruises and pain in her neck
Detected
2/4
14:45
The person was found lying on the floor; resulted in a hip fracture
Detected
3/5
02:00
The person was found lying on the floor next to her bed; no injuries reported
Detected
7/6
04:00
The person was found sitting on the floor; she reported that she had dragged herself from the toilet to the bedroom; no injuries reported
Detected
7/7
04:30
The person was found lying on the floor in her room; she had been on her way to the toilet when 1 leg gave way; no injuries reported
Detected
7/8
09:30
The person was walking and fell down in the joined living room; she had been helped back up; resulted in bruises on her right knee
Detected
7/9
15:00
The person fell down to her knees next to the toilet seat because 1 leg gave up on her; no injuries reported
Detected
7/10
17:39
The person fell in her room; she had crawled back next to the bed and was able to get up by herself; no injuries reported
Detected
7/11
18:15
The person was found lying on the floor in her room; she had been changing clothes when she lost her balance and fell; no injuries reported
Detected
7/12
23:30
The care personnel heard a noise from the room and found the person sitting on the floor next to the toilet room door; no injuries reported
Detected
7/13
19:50
The person was walking to her room and fell in the doorway; she was helped back up; no injuries reported
Data collected, but no fall alarm
4/14
16:30
The person was found sitting on the floor next to her bed; not known if the person has fallen down, but leg radiographed indicated no fracture
Data not collected
5/15
14:30
The person was found lying on the floor; he was alone on his way to the toilet; no injuries were reported
Data not collected
64
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
person having fallen in the doorway. Based on the acceleration data (fig. 1b) the person would have fallen into a seated end posture; however, this was not mentioned on the fall report. Most of the detected falls occurred during the night or afternoon (table 2). Only a minority of the false fall alarms were documented by the care personnel. Based on the personal communication and available alarm reports, most of the false fall alarms were reported to be related to clothing and to taking Kangas/Korpelainen/Vikman/Nyberg/ Jämsä
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
curred at times when the person did not have the sensor on. The implemented fall detection system detected 12 falls from 6 persons (table 2) out of 15 falls where the system was used during the accident, thus resulting in a fall detection sensitivity of 80.0%. Examples of acceleration data are presented in figure 1a–d. In 2 of the nondetected falls the person was found sitting or lying on the floor with no more details available. The third nondetected fall (fall 7/13 in table 2) was described as the
5
4
4
3
3
2
2
1
1
0
0
–1
1
2
3
b
–1
5
5
4
4
3
3
2
2
1
1
0
0
–1
c
0
0
1
2
–1
3
Time (s)
0
1
0
1
d
2
3
2
3
Time (s)
Fig. 1. Examples of fall and nonfall acceleration signals. Acceleration as a sum vector (solid line), and low-pass-filtered posture calculated from vertical acceleration (dashed line) [24, 26] from fall 7/7, resulting in an online fall alarm (a), fall 7/13, resulting in data collection but no fall alarm due to a nonlying end posture (b), false
fall alarm reported to have occurred when the test subject removed the sensor belt and placed it on the table (c) and nonfall-related activity, in which data collection was activated but the algorithm did not generate a fall alarm (d).
off or attaching the sensor at the waist (data not shown). An example figure of a false fall alarm is presented in figure 1c. During the field test, the fall detection system generated 812 fall alarms of which 52 occurred on days that were discarded due to technical reasons and 12 were fall related, thus resulting in 748 false fall alarms (table 1). The maximum number of daily false fall alarms per person was 10 and the corresponding median and average values were 0 and 0.61, respectively A total of 15,500 h of data were analyzed for false fall alarm rates. Based on the raw data, the false alarm rate was 0.049 false alarms per usage hour, equating to 20.4 usage hours per false fall alarm (table 1). The false alarm rate was reduced by around 40% when 15-min time periods from the beginning and ending of the daily usage period were excluded from the analyses (table 1). When fall alarms which were clustered within a 1-min time period were combined, the false alarm rate was reduced by 25% to 0.037 false alarms per usage hour, equating to 27.0 usage
hours per false fall alarm (table 1). When combining modifications Mod 2 and Mod 3, the false alarm rate was reduced to 0.025 false alarms per usage hour (40.0 usage hours per false alarm). Test persons who either had a fall history during the previous 6 months or had fallen during the test period had the trend of lower average false fall alarm rate based on the raw data (22.7 usage hours per false alarm) compared with the nonfaller group (16.7 usage hours per false alarm). After data modifications the trend remained, with the result for data modification Mod 3 being 46.1 and 38.9 usage hours per false alarm for the faller and nonfaller groups, respectively.
Fall Detection Sensitivity and False Alarm Rate
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
Discussion
In this study, we analyzed the sensitivity and false alarm rate of an accelerometer-based fall detector in reallife conditions among 16 older persons living in elderly care units. A total of 15,500 h of real-life data were moni65
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
Acceleration (g) Acceleration (g)
a
5
66
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
According to the Technology Acceptance Model [32], the actual usage of a device or service depends on the experience of its ease of use and usefulness. The ease of use is a design issue. The usefulness, however, is very much dependent on the alarm accuracy. It could be expected that for a frequent faller, the high sensitivity rate would motivate the use of the detector for the person wearing it, care staff and relatives, reducing fear and anxiety as well as the risk of long lies after a fall. For persons with a low risk of falls, on the other hand, even very low false alarm rates could heavily outnumber possible true alarms, causing nuisance and disruptions in daily activities and reducing the detector’s usefulness. The rate of 1 false alarm per 40 usage hours in this study may still be considered high and further development and a further reduction of false alarm rates is therefore important. This could be achieved simply by providing the user with the possibility to deactivate the alarm when needed and with more sophisticated data analyses for online algorithm improvements. In this study, 1 fall was recorded which was not recognized as a fall by the system. According to the signal, this was due to the requirement for a horizontal end posture of the fall in the algorithm. Based on the description of the event, the person had been walking and fell in the doorway. The acceleration data suggest that the fall ended in a sitting posture. Changing the threshold for posture detection might improve the fall detection sensitivity, although it might also have an effect on specificity. The exact definition of a fall event is vague and varies between studies. Two commonly used definitions, one used by the World Health Organization and another by the Kellogg International Work Group [33], do not require a horizontal end posture. This was also the case in a proposal for a fall model where a fall was defined as an unexpected event in which the person comes to rest on the ground, floor or lower level [34]. However, it has been shown that combining posture detection with impact and preimpact velocity improves the fall detection specificity and reduces the false alarm rate compared with impact and velocity only. Bourke et al. [20] monitored the posture change before and after a fall, without the criteria of a lying end posture. In some fall detection algorithms, the horizontal end posture for the lying posture after the fall-associated impact has been used as one criterion for a fall [26, 35]. This study has some limitations. Falls were monitored by the care personnel through their routine protocol at the care units. In most cases, the faller was alone at the time of the event and she/he was found sitting or lying Kangas/Korpelainen/Vikman/Nyberg/ Jämsä
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
tored. The results showed that the fall detection system detected 12 out of 15 real-life falls, having a sensitivity of 80%, with a false alarm rate as low as 0.025 alarms per usage hour. These data suggest that automatic accelerometric fall detection systems might offer a tool for improving safety among older people. A limited number of reports on fall detection sensitivity and specificity in real-life conditions are published. In the present study, the false alarm rate of 0.025 alarms per usage hour was in agreement with the range of 0.04–0.21 false alarms per hour presented previously [20]. Another study reported a group of 8 older people using fall detectors for a total of 168 days, resulting in a fall detection sensitivity of 62.5% and 25 false fall alarms [30]. In a recent study, 13 published fall detection algorithms were applied and tested on acceleration data from 29 real-life falls collected from 15 older persons [31]. Our previously published algorithms [24, 26] applied here were among those tested. The results from Bagalá et al. [31] show a high fall detection specificity close to 100% for our fall detection algorithms, as has been shown in our earlier studies with offline data analyses with experimental falls [26]. The fall detection sensitivity of 80% with real-life falls presented in the present study is lower than that in our previous study with experimental falls (92.5%) [26] but higher than the fall detection sensitivity of 48% from Bagalá et al. [31]. Their falls database included more subjects whose end posture was not lying, which may explain the difference. Furthermore, the false alarm rate in this study is lower than that generated in 24-hour recordings from 3 fallers showing values around 6 false alarms per day for our algorithms [31]. The data analyses between studies – offline data analysis [31] versus implemented real-time software in our study – are different. Also, the study populations are different: 15 test persons suffering from progressive supranuclear palsy, thus prone to recurrent falls, versus a population of 16 people including both fallers and nonfallers in our study. Based on the discussions with the care personnel, most of the false fall alarms were associated with the time periods when the sensor was being attached or taken off. Data processing involving deleting these 15-min time periods at the beginning and end of the daily usage period had a positive effect on the false alarm rate. The data processing with the clustering of alarms within a 1-min time period resulted in a lower false alarm rate than the nonmodified data. Combining these two modifications decreased the false alarm rate by approximately 50% compared with nonmodified data.
on the floor. In 1 case, the fall was discovered based on the fall alarm, even though the person was able to recover by herself. In practice, these kinds of falls with recovery may have occurred even more. Thus, the exact number of falls per person might be inaccurate. Generally, it has been estimated that at least 30% of falls remain unreported [36]. This study includes data from institutionalized older persons only. However, the monitored time period was longer than in previous studies [19, 20], and both fall detection sensitivity and false alarm rates were analyzed. Unfortunately, the compliance of care personnel to use the phone for SMS fall alarm messages remained low, also resulting in low compliancy in reporting false fall alarms. However, the acceleration signal for false alarms is available for further studies to improve fall detection accuracy. In conclusion, we have evaluated a real-time fall detection algorithm among 16 older persons during 15,500 h of recording. The results showed a fall detection sensitivity of 80%, with 1 false alarm per 40 usage hours. In the
future, more advanced algorithms to further reduce the false alarm rate would increase the applicability of the system.
Acknowledgments The authors would like to thank the subjects who participated in the study and the staff at the care units. Peter Wallström and Anders Lundbäck, CareTech AB, are acknowledged for technical contributions. This study was supported by the European Regional Development Fund of the European Union under the Interreg IV A North program, the Regional Council of Lapland, the County Administrative Board of Norrbotten, the Norrbotten County Council, the Norrbottens Forskningsråd, the Finnish Funding Agency for Technology and Innovations, the Academy of Finland, National Semiconductor Finland, CareTech AB, and Kalix Electropolis AB.
Disclosure Statement The authors have no conflicts of interest.
References
Fall Detection Sensitivity and False Alarm Rate
8 Lehtola S, Koistinen P, Luukinen H: Falls and injurious falls late in home-dwelling life. Arch Gerontol Geriatr 2006;42:217–224. 9 Melander-Wikman A, Jansson M, Hallberg J, Mortberg C, Gard G: The lighthouse alarm and locator trial – a pilot study. Technol Health Care 2007;15:203–212. 10 Bueno-Cavanillas A, Padilla-Ruiz F, JimenezMoleon JJ, Peinado-Alonso CA, Galvez-Vargas R: Risk factors in falls among the elderly according to extrinsic and intrinsic precipitating causes. Eur J Epidemiol 2000;16:849–859. 11 Fleming J, Brayne C; Cambridge City over-75s Cohort (CC75C) Study Collaboration: Inability to get up after falling, subsequent time on floor, and summoning help: prospective cohort study in people over 90. BMJ 2008;337:a2227. 12 Heinbuchner B, Hautzinger M, Becker C, Pfeiffer K: Satisfaction and use of personal emergency response systems. Z Gerontol Geriatr 2010;43:219–223. 13 Noury N, Rumeau P, Bourke AK, ÓLaighin G, Lundy JE: A proposal for the classification and evaluation of fall detectors. IRBM 2008; 29:340–349. 14 Bourke AK, O’Brien JV, Lyons GM: Evaluation of a threshold-based tri-axial accelerometer fall detection algorithm. Gait Posture 2007;26:194–199. 15 Karantonis DM, Narayanan MR, Mathie M, Lovell NH, Celler BG: Implementation of a real-time human movement classifier using a triaxial accelerometer for ambulatory moni-
16
17
18 19
20
21
22
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
toring. IEEE Trans Inf Technol Biomed 2006; 10:156–167. Lindemann U, Hock A, Stuber M, Keck W, Becker C: Evaluation of a fall detector based on accelerometers: a pilot study. Med Biol Eng Comput 2005;43:548–551. Benocci M, Tacconi C, Farella E, Benini L, Chiari L, Vanzago L: Accelerometer-based fall detection using optimized ZigBee data streaming. Microelectronics J 2010;41:703–710. Tamura T: Wearable accelerometer in clinical use. Conf Proc IEEE Eng Med Biol Soc 2005; 7:7165–7166. Bourke AK, van de Ven PW, Chaya AE, OLaighin GM, Nelson J: Testing of a longterm fall detection system incorporated into a custom vest for the elderly. Conf Proc IEEE Eng Med Biol Soc 2008;2008:2844–2847. Bourke AK, van de Ven P, Gamble M, O’Connor R, Murphy K, Bogan E, et al: Evaluation of waist-mounted tri-axial accelerometer based fall-detection algorithms during scripted and continuous unscripted activities. J Biomech 2010;43:3051–3057. Robinovitch SN, Brumer R, Maurer J: Effect of the ‘squat protective response’ on impact velocity during backward falls. J Biomech 2004;37:1329–1337. Kangas M, Vikman I, Nyberg L, Korpelainen R, Lindblom J, Jämsä T: Comparison of reallife accidental falls in older people with experimental falls in middle-aged test subjects. Gait Posture 2012;35:500–505.
67
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
1 Tinetti ME, Speechley M, Ginter SF: Risk factors for falls among elderly persons living in the community. N Engl J Med 1988; 319: 1701–1707. 2 Luukinen H, Koski K, Hiltunen L, Kivelä SL: Incidence rate of falls in an aged population in northern Finland. J Clin Epidemiol 1994; 47:843–850. 3 Campbell AJ, Borrie MJ, Spears GF, Jackson SL, Brown JS, Fitzgerald J: Circumstances and consequences of falls experienced by a community population 70 years and over during a prospective study. Age Ageing 1990;19:136–141. 4 Jensen J, Lundin-Olsson L, Nyberg L, Gustafson Y: Falls among frail older people in residential care. Scand J Public Health 2002; 30: 54–61. 5 Fletcher PC, Hirdes JP: Risk factors for falling among community-based seniors using home care services. J Gerontol A Biol Sci Med Sci 2002;57:M504–M510. 6 Pluijm SM, Smit JH, Tromp EA, Stel VS, Deeg DJ, Bouter LM, et al: A risk profile for identifying community-dwelling elderly with a high risk of recurrent falling: results of a 3-year prospective study. Osteoporos Int 2006; 17: 417–425. 7 Fleming J, Matthews FE, Brayne C; Cambridge City over-75s Cohort (CC75C) Study Collaboration: Falls in advanced old age: recalled falls and prospective follow-up of over90-year-olds in the Cambridge City over-75s Cohort Study. BMC Geriatr 2008;8:6.
68
28 Katz S, Ford AB, Moskowitz RW, Jaffe MW: Studies of illness in the aged. The index of ADL: a standardized measure of biological and psychosocial function. J Am Med Assoc 1963;185:914–919. 29 Podsiadlo D, Richardson S: The timed ‘Up & Go’: a test of basic functional mobility for frail elderly persons. J Am Geriatr Soc 1991; 39: 142–148. 30 Bloch F, Gautier V, Noury N, Lundy JE, Poujaud J, Claessens YE, et al: Evaluation under real-life conditions of a stand-alone fall detector for the elderly subjects. Ann Phys Rehabil Med 2011;54:391–398. 31 Bagalá F, Becker C, Cappello A, Chiari L, Aminian K, Hausdorff JM, et al: Evaluation of accelerometer-based fall detection algorithms on real-world falls. PLoS One 2012; 7:e37062. 32 Holden RJ, Karsh B: The Technology Acceptance Model: its past and its future in health care. J Biomed Informatics 2010;43:159–172.
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
33 Gibson MJ, Andres RO, Isaacs B, Radebaugh T, Worm-Petersen J: The prevention of falls in later life. A report of the Kellogg International Work Group on the prevention of falls by the elderly. Danish Med Bullet 1987; 34(suppl 4):1–24. 34 Becker C, Schwickert L, Mellone S, Bagala F, Chiari L, Helbostad JL, et al: Proposal for a multiphase fall model based on real-world fall recordings with body-fixed sensors. Z Gerontol Geriatr 2012;45:707–715. 35 Chao PK, Chan HL, Tang FT, Chen YC, Wong MK: A comparison of automatic fall detection by the cross-product and magnitude of tri-axial acceleration. Physiol Meas 2009;30:1027–1037. 36 Lord SR, Sherrington C, Menz H, Close J: Falls in older people. Risk factors and strategies for prevention. Cambridge, Cambridge University Press, 2007.
Kangas/Korpelainen/Vikman/Nyberg/ Jämsä
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
23 Klenk J, Becker C, Lieken F, Nicolai S, Maetzler W, Alt W, et al: Comparison of acceleration signals of simulated and real-world backward falls. Med Eng Phys 2011;33:368–373. 24 Kangas M, Konttila A, Lindgren P, Winblad I, Jämsä T: Comparison of low-complexity fall detection algorithms for body attached accelerometers. Gait Posture 2008;28:285–291. 25 Kangas M, Konttila A, Winblad I, Jämsä T: Determination of simple threshold for accelerometry-based parameters for fall detection. Conf Proc IEEE Eng Med Biol Soc 2007; 2007:1367–1370. 26 Kangas M, Vikman I, Wiklander J, Lindgren P, Nyberg L, Jämsä T: Sensitivity and specificity of fall detection in people aged 40 years and over. Gait Posture 2009;29:571–574. 27 Folstein MF, Folstein SE, McHugh PR: ‘Minimental state’. A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res 1975;12:1891–1898.
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
Sensitivity and False Alarm Rate of a Fall Sensor in Long-Term Fall Detection in the Elderly Maarit Kangas a, c Raija Korpelainen a–d Irene Vikman e Lars Nyberg e Timo Jämsä a, c
a
Department of Medical Technology, Institute of Biomedicine, b Institute of Health Sciences and c Medical Research Center Oulu, Oulu University Hospital, University of Oulu, and d Department of Sports and Exercise Medicine, Oulu Deaconess Institute, Oulu, Finland; e Department of Health Sciences, Luleå University of Technology, Luleå, Sweden
Abstract Background: About a third of home-dwelling older people fall each year, and institutionalized older people even report a two- or threefold higher rate for falling. Automatic fall detection systems have been developed to support the independent and secure living of the elderly. Even though good fall detection sensitivity and specificity in laboratory settings have been reported, knowledge about the sensitivity and specificity of these systems in real-life conditions is still lacking. Objective: The aim of this study was to evaluate the long-term fall detection sensitivity and false alarm rate of a fall detection prototype in real-life use. Methods: A total of 15,500 h of real-life data from 16 older people, including both fallers and nonfallers, were monitored using an accelerometry-based sensor system with an implemented fall detection algorithm. Results: The fall detection system detect-
© 2014 S. Karger AG, Basel 0304–324X/14/0611–0061$39.50/0 E-Mail [email protected] www.karger.com/ger
ed 12 out of 15 real-life falls, having a sensitivity of 80.0%, with a false alarm rate of 0.049 alarms per usage hour with the implemented real-time system. With minor modification of data analysis the false alarm rate was reduced to 0.025 false alarms per hour, equating to 1 false fall alarm per 40 usage hours. Conclusion: These data suggest that automatic accelerometric fall detection systems might offer a tool for improving safety among older people. © 2014 S. Karger AG, Basel
Introduction
The proportion of the older population aged over 65 years is growing rapidly in most countries. About a third of home-dwelling older people fall each year [1] and fall rate estimates range between 517 and 683 falls per 1,000 person years [2, 3]. Institutionalized older people present a two- or threefold higher rate of falling compared with the community-dwelling population [2, 4]. Between 10 and 20% of older people fall recurrently [2, 5, 6]. The proMaarit Kangas, PhD Department of Medical Technology, Institute of Biomedicine University of Oulu, PO Box 5000 FI–90014 Oulu (Finland) E-Mail maarit.kangas @ oulu.fi
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
Key Words Accelerometer · Fall detector · Specificity · Older people · Frequent faller
62
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
alarm rate of a fall detector prototype in real-life conditions with 16 older people. Minor modifications of the fall detection algorithm were also tested to further improve the specificity of the system. Materials and Methods Study Subjects and Data Collection People living in elderly care units were recruited to wear a prototype of an automatic fall detector. The study protocol was approved by the Ethical Review Boards in Oulu, Finland (39/2009) and in Luleå, Sweden (2443-2009). The participants and their close relatives received oral information about the study, and written informed consent was obtained. The inclusion criteria for the study included subjects being older than 65 years and possessing the ability to stand independently or with the help of one person. The field test was performed over a 10-month period with a sliding starting and ending point for each care unit, thus resulting in various lengths of individual test periods. Cognitive functions with the mini-mental state examination (MMSE) [27], Katz Index of Independence in Activities of Daily Living [28], 10-meter walking speed indoors, and mobility with Timed Up and Go (TUG) [29] were measured before the test period. The test population included 13 females and 3 males with an average age of 88.4 ± 5.2 years (n = 16), an MMSE of 13.13 ± 8.22 points (n = 13), a walking speed of 0.56 ± 0.31 m/s (n = 14) and a TUG score of 33.4 ± 33.8 s (n = 13). The status of falling 6 months before and during the test period was assessed from the reports of the care personnel. Fall Detector The test subjects wore the prototype of the accelerometrybased fall detector unit described earlier [22]. The device was designed to activate and collect acceleration data when the acceleration of all three axes fell below a threshold of 0.75 g. The block of around 14 s included buffered data before the activation and the collected data after that. After data storage activation, the implemented fall detection algorithm detected impact from an acceleration sum vector based on a threshold of 2 g and an end posture based on the vertical axis acceleration, with minor modifications (CareTech AB, Sweden) from our earlier studies [24, 26]. In the case of a detected fall, the collected acceleration data were labeled with a fall message and the data were transmitted further to the research database using IP-based technology. In a case of a fall message, an SMS message was sent to the care unit. In addition to the implemented fall detection, the sensor continuously collected activity data obtained as a number of vertical acceleration peaks above the threshold of 1.19 g in 1-min epochs (CareTech AB). The system transmitted the activity data using the same protocol used for transmitting acceleration data. Activity data were used to evaluate the usage of the sensor. The sensor (size 5 × 3 × 1.5 cm) was attached with a clip into a pocket on an elastic belt, except for 1 user where a textile pocket on the trouser waistband was used to fix the vertical axis of the device. The attachment site was at the waist, around the anterior superior iliac spine, but some variation was accepted due to the nature of the long-term field test.
Kangas/Korpelainen/Vikman/Nyberg/ Jämsä
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
portion of recurrent fallers increases with age, resulting in almost half of the home-dwelling elderly over 85 years being recurrent fallers [7, 8]. Fear of not being able to recover after falling is common among older people [9]. Approximately half of fallers have been incapable of recovery to a standing position unaided [10]. To prevent long lies, commercially available personal emergency response systems provide applications to call for help. However, around 80% of older people wearing these systems who are unable to recover after a fall do not use their alarm system to call for help [11, 12]. Some commercially available automatic fall detection systems exist, as summarized by Noury et al. [13]. Even though good fall detection sensitivity and specificity in laboratory settings have been reported [14–17], knowledge on the sensitivity and specificity, as well as the acceptability and usability, of these systems in real-life conditions is still lacking. Tamura [18] reported 19 real-life falls detected with an algorithm monitoring acceleration characteristics from a Parkinson’s patient, but no false alarm rate evaluations were presented. In some studies, false fall alarms have been evaluated in real-life conditions, but no falls have occurred during the test period. A vest-worn fall detection system generated 42 false alarms during 833 real-life monitoring hours in older people [19]. In another study, a waist-worn fall detection system was tested for 52 h among older people [20]. A difference between laboratory and real-life samples was detected, since the algorithms resulting in 100% specificity with instructed activities of daily living resulted in a false alarm rate between 0.04 and 0.21 alarms per hour with data collected from real-life occurrences [20]. However, self-initiated, intentional falls may differ from sudden, unexpected falls [21]. Recent reports have indicated that intentional falls in a laboratory environment have both similar and different features compared with reallife falls among older people [22, 23], thus indicating the importance of evaluating fall detection systems in reallife conditions. We have previously validated a fall detection system based on a waist-worn accelerometer with an algorithm for impact and end posture detection [24, 25]. This system has achieved a fall detection sensitivity of 97% and a specificity of 100% in a laboratory setting with intentional falls and instructed activities of daily living from middle-aged test persons [26]. Since knowledge of the sensitivity and specificity of fall detection systems in real-life conditions is still lacking, the aim of this study was to evaluate the long-term fall detection sensitivity and false
Table 1. Use of the fall detector and false alarm rate in a study population of 16 elderly subjects: original fall detection algorithm (raw
data) and modified algorithms (Mod 1–Mod 3) Person
13 21 31 41 51 62 73 81 92 103 112 122 132 143 153 163 Total
Days, n
Alarms, n
worn
not worn
148 127 33 34 5 69 155 61 14 135 17 72 11 118 40 66 1,105
13 15 18 48 0 22 24 13 14 30 0 1 19 25 59 6 306
21 45 65 13 2 43 124 44 0 132 0 105 3 118 22 11 748
Usage hours
False alarm rate (alarms/usage hour)
total, h
average ± SD, h/day
raw data
Mod 1
Mod 2
Mod 3
1,667 1,232 359 265 25 956 3,179 909 165 2,407 197 873 120 1,688 361 1,097 15,500
11.5±2.6 9.8±2.8 10.9±6.1 7.8±2.9 5.0±2.6 8.6±7.1 20.6±4.0 15.0±6.1 11.4±6.2 18.4±6.2 10.9±4.2 12.1±1.3 10.9±3.8 14.1±4.9 9.0±2.0 15.5±5.7 14.2±6.3
0.013 0.037 0.181 0.050 0.080 0.045 0.040 0.048 0.000 0.055 0.000 0.121 0.025 0.070 0.061 0.010 0.049
0.004 0.011 0.136 0.015 0.000 0.029 0.034 0.021 0.000 0.027 0.000 0.086 0.025 0.041 0.011 0.006 0.030
0.010 0.033 0.150 0.038 0.080 0.035 0.029 0.042 0.000 0.035 0.000 0.096 0.025 0.056 0.053 0.007 0.037
0.005 0.011 0.114 0.011 0.000 0.023 0.025 0.021 0.000 0.173 0.000 0.079 0.025 0.036 0.008 0.005 0.025
Falls, n
2 2 2 1 1 1 9 04 04 04 0 0 0 0 0 0 18
Days: technical failure days excluded. Alarms: nonuse days excluded. Average usage hours/day: nonuse days excluded. Falls: number of falls reported during the test period; for detected falls see table 2. Data modifications Mod 1–3 explained in Material and Methods. Superscript numbers 1–3 indicate status at the end of the test period. 1 Dead or withdrawn due to decline in health status. 2 Withdrawal due to other reasons. 3 Completed the test period. 4 Did not fall during the test period but had a fall history before the test period.
• Mod 3 was calculated by combining modifications Mod 1 and Mod 2. To calculate the fall detection sensitivity, falls that occurred when the sensor was worn were included. Fall detection sensitivity was calculated as the percentage of falls correctly identified as falls by the system.
Analyses of Data Real-life falls were identified based on the routine reporting by the care personnel. Based on the time of the fall event, the acceleration database was analyzed to find the corresponding acceleration data. Care personnel were asked to report alarms and categorize them as fall alarms or false fall alarms. The number of fall alarms was collected from the database based on the fall label generated by the implemented fall detection system. For usage hour analyses, inactivity periods exceeding 3 h were excluded during the daytime. At nighttime, if an indication of any activity was seen, all hours were included. The false alarm rate was calculated by dividing the number of false alarms by the number of usage hours during the test period. Usage time (in hours) for 1 false alarm to occur is also reported. Average values for the test population were weighted by the usage hours. The false alarm rate was first calculated based on the raw data. The effect of minor data analysis procedures was studied offline using three different data modifications (Mod 1–Mod 3): • Mod 1 was calculated by excluding the false fall alarms related to the attachment or taking off the sensor system. This modification was performed by deleting the 15-min time period from the beginning and from the end of the daily usage period if available. • Mod 2 was calculated by excluding multiple false fall alarms in a short time period. This modification was performed by combining alarms clustered in 1-min sliding epochs.
The total length of the test period was 1,540 days, consisting of 1,105 days when the fall detection system was used by the test persons and 306 days when the sensor was not used. A total of 129 days were discarded from analyses due to technical reasons (table 1). The sensor system was mostly used during waking hours but some test persons used the system also at nighttime. Overall, 6 test persons completed the test period; 10 withdrew from the study, 5 due to death or a decline in health status and 5 for other reasons (table 1). During the test period, 18 falls were reported to have occurred among 7 of the test persons (table 1); 3 falls oc-
Fall Detection Sensitivity and False Alarm Rate
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
63
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
Results
Table 2. Description and detection of the 15 falls of the study subjects during the monitoring period
Person No./fall No.
Time
Description
Fall detection
1/1
02:00
The person was trying to get up from her bed but the blanket was entangled around her legs; she slid on the floor; no injuries reported
Detected
1/2
19:20
The person was found lying on the floor of her room after a fall; she had gotten entangled in a blanket when moving from the bed; resulted in bruises in the left side rib cage
Detected
2/3
12:35
The person was found on the floor near the toilet seat; resulted in bruises and pain in her neck
Detected
2/4
14:45
The person was found lying on the floor; resulted in a hip fracture
Detected
3/5
02:00
The person was found lying on the floor next to her bed; no injuries reported
Detected
7/6
04:00
The person was found sitting on the floor; she reported that she had dragged herself from the toilet to the bedroom; no injuries reported
Detected
7/7
04:30
The person was found lying on the floor in her room; she had been on her way to the toilet when 1 leg gave way; no injuries reported
Detected
7/8
09:30
The person was walking and fell down in the joined living room; she had been helped back up; resulted in bruises on her right knee
Detected
7/9
15:00
The person fell down to her knees next to the toilet seat because 1 leg gave up on her; no injuries reported
Detected
7/10
17:39
The person fell in her room; she had crawled back next to the bed and was able to get up by herself; no injuries reported
Detected
7/11
18:15
The person was found lying on the floor in her room; she had been changing clothes when she lost her balance and fell; no injuries reported
Detected
7/12
23:30
The care personnel heard a noise from the room and found the person sitting on the floor next to the toilet room door; no injuries reported
Detected
7/13
19:50
The person was walking to her room and fell in the doorway; she was helped back up; no injuries reported
Data collected, but no fall alarm
4/14
16:30
The person was found sitting on the floor next to her bed; not known if the person has fallen down, but leg radiographed indicated no fracture
Data not collected
5/15
14:30
The person was found lying on the floor; he was alone on his way to the toilet; no injuries were reported
Data not collected
64
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
person having fallen in the doorway. Based on the acceleration data (fig. 1b) the person would have fallen into a seated end posture; however, this was not mentioned on the fall report. Most of the detected falls occurred during the night or afternoon (table 2). Only a minority of the false fall alarms were documented by the care personnel. Based on the personal communication and available alarm reports, most of the false fall alarms were reported to be related to clothing and to taking Kangas/Korpelainen/Vikman/Nyberg/ Jämsä
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
curred at times when the person did not have the sensor on. The implemented fall detection system detected 12 falls from 6 persons (table 2) out of 15 falls where the system was used during the accident, thus resulting in a fall detection sensitivity of 80.0%. Examples of acceleration data are presented in figure 1a–d. In 2 of the nondetected falls the person was found sitting or lying on the floor with no more details available. The third nondetected fall (fall 7/13 in table 2) was described as the
5
4
4
3
3
2
2
1
1
0
0
–1
1
2
3
b
–1
5
5
4
4
3
3
2
2
1
1
0
0
–1
c
0
0
1
2
–1
3
Time (s)
0
1
0
1
d
2
3
2
3
Time (s)
Fig. 1. Examples of fall and nonfall acceleration signals. Acceleration as a sum vector (solid line), and low-pass-filtered posture calculated from vertical acceleration (dashed line) [24, 26] from fall 7/7, resulting in an online fall alarm (a), fall 7/13, resulting in data collection but no fall alarm due to a nonlying end posture (b), false
fall alarm reported to have occurred when the test subject removed the sensor belt and placed it on the table (c) and nonfall-related activity, in which data collection was activated but the algorithm did not generate a fall alarm (d).
off or attaching the sensor at the waist (data not shown). An example figure of a false fall alarm is presented in figure 1c. During the field test, the fall detection system generated 812 fall alarms of which 52 occurred on days that were discarded due to technical reasons and 12 were fall related, thus resulting in 748 false fall alarms (table 1). The maximum number of daily false fall alarms per person was 10 and the corresponding median and average values were 0 and 0.61, respectively A total of 15,500 h of data were analyzed for false fall alarm rates. Based on the raw data, the false alarm rate was 0.049 false alarms per usage hour, equating to 20.4 usage hours per false fall alarm (table 1). The false alarm rate was reduced by around 40% when 15-min time periods from the beginning and ending of the daily usage period were excluded from the analyses (table 1). When fall alarms which were clustered within a 1-min time period were combined, the false alarm rate was reduced by 25% to 0.037 false alarms per usage hour, equating to 27.0 usage
hours per false fall alarm (table 1). When combining modifications Mod 2 and Mod 3, the false alarm rate was reduced to 0.025 false alarms per usage hour (40.0 usage hours per false alarm). Test persons who either had a fall history during the previous 6 months or had fallen during the test period had the trend of lower average false fall alarm rate based on the raw data (22.7 usage hours per false alarm) compared with the nonfaller group (16.7 usage hours per false alarm). After data modifications the trend remained, with the result for data modification Mod 3 being 46.1 and 38.9 usage hours per false alarm for the faller and nonfaller groups, respectively.
Fall Detection Sensitivity and False Alarm Rate
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
Discussion
In this study, we analyzed the sensitivity and false alarm rate of an accelerometer-based fall detector in reallife conditions among 16 older persons living in elderly care units. A total of 15,500 h of real-life data were moni65
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
Acceleration (g) Acceleration (g)
a
5
66
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
According to the Technology Acceptance Model [32], the actual usage of a device or service depends on the experience of its ease of use and usefulness. The ease of use is a design issue. The usefulness, however, is very much dependent on the alarm accuracy. It could be expected that for a frequent faller, the high sensitivity rate would motivate the use of the detector for the person wearing it, care staff and relatives, reducing fear and anxiety as well as the risk of long lies after a fall. For persons with a low risk of falls, on the other hand, even very low false alarm rates could heavily outnumber possible true alarms, causing nuisance and disruptions in daily activities and reducing the detector’s usefulness. The rate of 1 false alarm per 40 usage hours in this study may still be considered high and further development and a further reduction of false alarm rates is therefore important. This could be achieved simply by providing the user with the possibility to deactivate the alarm when needed and with more sophisticated data analyses for online algorithm improvements. In this study, 1 fall was recorded which was not recognized as a fall by the system. According to the signal, this was due to the requirement for a horizontal end posture of the fall in the algorithm. Based on the description of the event, the person had been walking and fell in the doorway. The acceleration data suggest that the fall ended in a sitting posture. Changing the threshold for posture detection might improve the fall detection sensitivity, although it might also have an effect on specificity. The exact definition of a fall event is vague and varies between studies. Two commonly used definitions, one used by the World Health Organization and another by the Kellogg International Work Group [33], do not require a horizontal end posture. This was also the case in a proposal for a fall model where a fall was defined as an unexpected event in which the person comes to rest on the ground, floor or lower level [34]. However, it has been shown that combining posture detection with impact and preimpact velocity improves the fall detection specificity and reduces the false alarm rate compared with impact and velocity only. Bourke et al. [20] monitored the posture change before and after a fall, without the criteria of a lying end posture. In some fall detection algorithms, the horizontal end posture for the lying posture after the fall-associated impact has been used as one criterion for a fall [26, 35]. This study has some limitations. Falls were monitored by the care personnel through their routine protocol at the care units. In most cases, the faller was alone at the time of the event and she/he was found sitting or lying Kangas/Korpelainen/Vikman/Nyberg/ Jämsä
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
tored. The results showed that the fall detection system detected 12 out of 15 real-life falls, having a sensitivity of 80%, with a false alarm rate as low as 0.025 alarms per usage hour. These data suggest that automatic accelerometric fall detection systems might offer a tool for improving safety among older people. A limited number of reports on fall detection sensitivity and specificity in real-life conditions are published. In the present study, the false alarm rate of 0.025 alarms per usage hour was in agreement with the range of 0.04–0.21 false alarms per hour presented previously [20]. Another study reported a group of 8 older people using fall detectors for a total of 168 days, resulting in a fall detection sensitivity of 62.5% and 25 false fall alarms [30]. In a recent study, 13 published fall detection algorithms were applied and tested on acceleration data from 29 real-life falls collected from 15 older persons [31]. Our previously published algorithms [24, 26] applied here were among those tested. The results from Bagalá et al. [31] show a high fall detection specificity close to 100% for our fall detection algorithms, as has been shown in our earlier studies with offline data analyses with experimental falls [26]. The fall detection sensitivity of 80% with real-life falls presented in the present study is lower than that in our previous study with experimental falls (92.5%) [26] but higher than the fall detection sensitivity of 48% from Bagalá et al. [31]. Their falls database included more subjects whose end posture was not lying, which may explain the difference. Furthermore, the false alarm rate in this study is lower than that generated in 24-hour recordings from 3 fallers showing values around 6 false alarms per day for our algorithms [31]. The data analyses between studies – offline data analysis [31] versus implemented real-time software in our study – are different. Also, the study populations are different: 15 test persons suffering from progressive supranuclear palsy, thus prone to recurrent falls, versus a population of 16 people including both fallers and nonfallers in our study. Based on the discussions with the care personnel, most of the false fall alarms were associated with the time periods when the sensor was being attached or taken off. Data processing involving deleting these 15-min time periods at the beginning and end of the daily usage period had a positive effect on the false alarm rate. The data processing with the clustering of alarms within a 1-min time period resulted in a lower false alarm rate than the nonmodified data. Combining these two modifications decreased the false alarm rate by approximately 50% compared with nonmodified data.
on the floor. In 1 case, the fall was discovered based on the fall alarm, even though the person was able to recover by herself. In practice, these kinds of falls with recovery may have occurred even more. Thus, the exact number of falls per person might be inaccurate. Generally, it has been estimated that at least 30% of falls remain unreported [36]. This study includes data from institutionalized older persons only. However, the monitored time period was longer than in previous studies [19, 20], and both fall detection sensitivity and false alarm rates were analyzed. Unfortunately, the compliance of care personnel to use the phone for SMS fall alarm messages remained low, also resulting in low compliancy in reporting false fall alarms. However, the acceleration signal for false alarms is available for further studies to improve fall detection accuracy. In conclusion, we have evaluated a real-time fall detection algorithm among 16 older persons during 15,500 h of recording. The results showed a fall detection sensitivity of 80%, with 1 false alarm per 40 usage hours. In the
future, more advanced algorithms to further reduce the false alarm rate would increase the applicability of the system.
Acknowledgments The authors would like to thank the subjects who participated in the study and the staff at the care units. Peter Wallström and Anders Lundbäck, CareTech AB, are acknowledged for technical contributions. This study was supported by the European Regional Development Fund of the European Union under the Interreg IV A North program, the Regional Council of Lapland, the County Administrative Board of Norrbotten, the Norrbotten County Council, the Norrbottens Forskningsråd, the Finnish Funding Agency for Technology and Innovations, the Academy of Finland, National Semiconductor Finland, CareTech AB, and Kalix Electropolis AB.
Disclosure Statement The authors have no conflicts of interest.
References
Fall Detection Sensitivity and False Alarm Rate
8 Lehtola S, Koistinen P, Luukinen H: Falls and injurious falls late in home-dwelling life. Arch Gerontol Geriatr 2006;42:217–224. 9 Melander-Wikman A, Jansson M, Hallberg J, Mortberg C, Gard G: The lighthouse alarm and locator trial – a pilot study. Technol Health Care 2007;15:203–212. 10 Bueno-Cavanillas A, Padilla-Ruiz F, JimenezMoleon JJ, Peinado-Alonso CA, Galvez-Vargas R: Risk factors in falls among the elderly according to extrinsic and intrinsic precipitating causes. Eur J Epidemiol 2000;16:849–859. 11 Fleming J, Brayne C; Cambridge City over-75s Cohort (CC75C) Study Collaboration: Inability to get up after falling, subsequent time on floor, and summoning help: prospective cohort study in people over 90. BMJ 2008;337:a2227. 12 Heinbuchner B, Hautzinger M, Becker C, Pfeiffer K: Satisfaction and use of personal emergency response systems. Z Gerontol Geriatr 2010;43:219–223. 13 Noury N, Rumeau P, Bourke AK, ÓLaighin G, Lundy JE: A proposal for the classification and evaluation of fall detectors. IRBM 2008; 29:340–349. 14 Bourke AK, O’Brien JV, Lyons GM: Evaluation of a threshold-based tri-axial accelerometer fall detection algorithm. Gait Posture 2007;26:194–199. 15 Karantonis DM, Narayanan MR, Mathie M, Lovell NH, Celler BG: Implementation of a real-time human movement classifier using a triaxial accelerometer for ambulatory moni-
16
17
18 19
20
21
22
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
toring. IEEE Trans Inf Technol Biomed 2006; 10:156–167. Lindemann U, Hock A, Stuber M, Keck W, Becker C: Evaluation of a fall detector based on accelerometers: a pilot study. Med Biol Eng Comput 2005;43:548–551. Benocci M, Tacconi C, Farella E, Benini L, Chiari L, Vanzago L: Accelerometer-based fall detection using optimized ZigBee data streaming. Microelectronics J 2010;41:703–710. Tamura T: Wearable accelerometer in clinical use. Conf Proc IEEE Eng Med Biol Soc 2005; 7:7165–7166. Bourke AK, van de Ven PW, Chaya AE, OLaighin GM, Nelson J: Testing of a longterm fall detection system incorporated into a custom vest for the elderly. Conf Proc IEEE Eng Med Biol Soc 2008;2008:2844–2847. Bourke AK, van de Ven P, Gamble M, O’Connor R, Murphy K, Bogan E, et al: Evaluation of waist-mounted tri-axial accelerometer based fall-detection algorithms during scripted and continuous unscripted activities. J Biomech 2010;43:3051–3057. Robinovitch SN, Brumer R, Maurer J: Effect of the ‘squat protective response’ on impact velocity during backward falls. J Biomech 2004;37:1329–1337. Kangas M, Vikman I, Nyberg L, Korpelainen R, Lindblom J, Jämsä T: Comparison of reallife accidental falls in older people with experimental falls in middle-aged test subjects. Gait Posture 2012;35:500–505.
67
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
1 Tinetti ME, Speechley M, Ginter SF: Risk factors for falls among elderly persons living in the community. N Engl J Med 1988; 319: 1701–1707. 2 Luukinen H, Koski K, Hiltunen L, Kivelä SL: Incidence rate of falls in an aged population in northern Finland. J Clin Epidemiol 1994; 47:843–850. 3 Campbell AJ, Borrie MJ, Spears GF, Jackson SL, Brown JS, Fitzgerald J: Circumstances and consequences of falls experienced by a community population 70 years and over during a prospective study. Age Ageing 1990;19:136–141. 4 Jensen J, Lundin-Olsson L, Nyberg L, Gustafson Y: Falls among frail older people in residential care. Scand J Public Health 2002; 30: 54–61. 5 Fletcher PC, Hirdes JP: Risk factors for falling among community-based seniors using home care services. J Gerontol A Biol Sci Med Sci 2002;57:M504–M510. 6 Pluijm SM, Smit JH, Tromp EA, Stel VS, Deeg DJ, Bouter LM, et al: A risk profile for identifying community-dwelling elderly with a high risk of recurrent falling: results of a 3-year prospective study. Osteoporos Int 2006; 17: 417–425. 7 Fleming J, Matthews FE, Brayne C; Cambridge City over-75s Cohort (CC75C) Study Collaboration: Falls in advanced old age: recalled falls and prospective follow-up of over90-year-olds in the Cambridge City over-75s Cohort Study. BMC Geriatr 2008;8:6.
68
28 Katz S, Ford AB, Moskowitz RW, Jaffe MW: Studies of illness in the aged. The index of ADL: a standardized measure of biological and psychosocial function. J Am Med Assoc 1963;185:914–919. 29 Podsiadlo D, Richardson S: The timed ‘Up & Go’: a test of basic functional mobility for frail elderly persons. J Am Geriatr Soc 1991; 39: 142–148. 30 Bloch F, Gautier V, Noury N, Lundy JE, Poujaud J, Claessens YE, et al: Evaluation under real-life conditions of a stand-alone fall detector for the elderly subjects. Ann Phys Rehabil Med 2011;54:391–398. 31 Bagalá F, Becker C, Cappello A, Chiari L, Aminian K, Hausdorff JM, et al: Evaluation of accelerometer-based fall detection algorithms on real-world falls. PLoS One 2012; 7:e37062. 32 Holden RJ, Karsh B: The Technology Acceptance Model: its past and its future in health care. J Biomed Informatics 2010;43:159–172.
Gerontology 2015;61:61–68 DOI: 10.1159/000362720
33 Gibson MJ, Andres RO, Isaacs B, Radebaugh T, Worm-Petersen J: The prevention of falls in later life. A report of the Kellogg International Work Group on the prevention of falls by the elderly. Danish Med Bullet 1987; 34(suppl 4):1–24. 34 Becker C, Schwickert L, Mellone S, Bagala F, Chiari L, Helbostad JL, et al: Proposal for a multiphase fall model based on real-world fall recordings with body-fixed sensors. Z Gerontol Geriatr 2012;45:707–715. 35 Chao PK, Chan HL, Tang FT, Chen YC, Wong MK: A comparison of automatic fall detection by the cross-product and magnitude of tri-axial acceleration. Physiol Meas 2009;30:1027–1037. 36 Lord SR, Sherrington C, Menz H, Close J: Falls in older people. Risk factors and strategies for prevention. Cambridge, Cambridge University Press, 2007.
Kangas/Korpelainen/Vikman/Nyberg/ Jämsä
Downloaded by: Stockholms Universitet 198.143.54.1 - 8/14/2015 1:18:19 AM
23 Klenk J, Becker C, Lieken F, Nicolai S, Maetzler W, Alt W, et al: Comparison of acceleration signals of simulated and real-world backward falls. Med Eng Phys 2011;33:368–373. 24 Kangas M, Konttila A, Lindgren P, Winblad I, Jämsä T: Comparison of low-complexity fall detection algorithms for body attached accelerometers. Gait Posture 2008;28:285–291. 25 Kangas M, Konttila A, Winblad I, Jämsä T: Determination of simple threshold for accelerometry-based parameters for fall detection. Conf Proc IEEE Eng Med Biol Soc 2007; 2007:1367–1370. 26 Kangas M, Vikman I, Wiklander J, Lindgren P, Nyberg L, Jämsä T: Sensitivity and specificity of fall detection in people aged 40 years and over. Gait Posture 2009;29:571–574. 27 Folstein MF, Folstein SE, McHugh PR: ‘Minimental state’. A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res 1975;12:1891–1898.
