Journal of Medical and Biological Engineering, 30(4): 47-5 47 Fall Detection System for Healthcare Quality Improvement in Residential Care Facilities Chih-Ning Huang 1 Chih-Yen Chiang 1 Guan-Chun Chen 1 Steen J. Hsu Woei-Chyn Chu 1 Chia-Tai Chan 1,* 1 Institute of Biomedical Engineering, National Yang-Ming University, Taipei 11, Taiwan Department of Information Management, Minghsin University of Science and Technology, Hsinchu 304, Taiwan Received 6 Feb 010; Accepted 5 Jul 010 Abstract Falls and fall-induced injuries among elderly people have become an important public health concern in an aging society. More than 50% of those living in residential care facilities fall at least once a year, and about half of them fall more than once a year. Because fall-induced injuries result in health decline and increasing medical care cost, fall management plays an important role in the residential care facilities. In this study, we propose a fall-detecting system based on wearable sensor and real-time fall detection algorithm. We use a head-mounted tri-axial accelerometer to capture the movement data of human body and develop a fall detection method to distinguish between falls and daily activities. A ZigBee-based alarm system is also proposed. It provides location information of the user in the case of emergency. When a fall happens, the caregivers can know where the accident is and then give immediate care for the residents directly to reduce severe injury, which could improve the healthcare quality in residential care facilities. The experimental results have demonstrated the proposed scheme has high reliability and sensitivity for fall detection. It fulfills the requirements of fall detection. Keywords: Fall detection, ZigBee, Embedded system, Residential care facilities 1. Introduction Recently, the population of older adults has increased rapidly due to medical treatment improvements, especially in the developed and developing countries. The data from the International Data Base (IDB) of the U.S. Census Bureau, Population Division [1], shows that in these countries, the number of people over the age of 65 was about 10%~0% of the population in 009. Further, in the super-aged countries, like Italy, France and Japan, the number of people over 65 years old was over 0%. In Taiwan, the number of people 65 years old and above had reached 10.7% of the population in 009. With the coming of old-aged society, falls and fall induced injuries among elderly people are important public health concerns [,3]. The mean incidence of falls in nursing homes is 1.5 falls per bed per year [4]. More than 50% of those living in residential care facilities fall every year, and about half of them do so repeatedly[4,5]. For the residents, a high proportion of falls result in injuries, like longstanding pain, lacerations, and fractures, and can even lead to death. Non-injurious falls also * Corresponding author: Chia-Tai Chan Tel: +886--867371; Fax: +886--810847 E-mail: ctchan@ym.edu.tw have a negative impact on residents, with about 75% of fallers experiencing loss of confidence and/or fear of further falls that will cause loss of physical activity and function [6]. For residential care facilities, falls incur medical disputes with residents or their families, and high medical expenses. The most well-known and widely used fall management approaches are fall prevention and fall detection. The former includes regular exercise, vitamin D and calcium supplementation, withdrawal of psychotropic medication, professional environment hazard assessment and modification, setting protection, and multiple interventions [7]. Unfortunately, we cannot prevent fall accidents completely, so a real-time fall alarm to caregivers where a fall has happened is significant after incidents. The great progresses of wireless communications and microelectro-mechanical systems (MEMSs) have enabled the development of low-cost, low-power, multifunctional, tiny sensor nodes that can sense the environment and communicate with each other over short distances. It has led to an emerging area with wide range of potential applications such as healthcare. Accelerometers have been used in various studies to monitor a range of human movement [8-13]. Sensors can be associated with the ZigBee wireless nodes to create healthcare applications for elderly people in residential care facilities. Taking advantage of ZigBee features, users can move freely around the residential
48 J. Med. Biol. Eng., Vol. 30. No. 4 010 environment. In this study, we proposed a location-aware fall detection system using tri-axial accelerometers as fall-detecting sensors and IEEE 80.15.4 ZigBee nodes as indoor position engines. The proposed system can provide emergency alarm and position information to caregivers through ZigBee WSN as soon as residents fall, which can improve the healthcare quality in residential care facilities, and makes the residents more confident regarding daily activities.. Materials and methods.1 Participants Two females and three males participated in the fall-detection experiment. The basic characteristics of these participants are shown in Table 1. Table 1. The basic characteristics of the research subject. Age (y) Body height (cm) Body weight (kg) Females 4 ± 0 16.5 ±.1 57 ± 5.66 Males 6.67 ± 3. 171.33 ± 5.69 68.33 ± 6.35. Fall detection mechanism The assessment of unintentional fall is difficult due to the subtle and complex nature of body movement, which requires accurate and reliable measuring techniques. The sensor placement is also critical, such as hip, trunk, wrist or head, since it will result in different signal patterns. Diverse research communities have investigated the related literature and proposed many different mechanisms [8-11], but the efficiency of fall detection and the reliability of posture recognition are always challenges. Careful consideration of the movements for unintentional falls and daily activities is essential for designing a successful fall-detection algorithm...1 Movement characteristics of unintentional falls and daily activities This study focused on the unintentional falls caused by faint or weakness that make residents fall on the ground unconsciously from their daily activities. We classified falls into 8 major types, each fall has 1~3 kinds of directions. The falling directions including front fall, posterior fall and lateral fall at either left or right side, were executed in our experiments. On the other side, we also selected 7 types of daily movement with normal or fast speed. The falls and daily activities are shown on Table... Wearable sensor location Several portable accelerometer systems have been designed to detect falls [8-11], but these fall detection systems cannot fulfill the requirements of the efficiency of fall detection and the reliability of posture recognition. According to the wearing position, we summarize the four latest accelerometer-based fall detection schemes as follows. The first one used a tri-axial accelerometer (Analog Devices ADXL150EM-3) on the chest [8]. However, it made erroneous detections easily when people tilt seriously. The second scheme proposed a wrist-mounted fall Table. Falls and daily activities movement characteristics. Daily activities Characteristics Stand From sit From squat Sit Normal Fast Lie on the bed Normal Fast Walk Normal Fast Jump On the ground On the bed Go up and down stairs Normal Fast Run (18 meters) Normal speed Fall Characteristics Stand Front Posterior Lateral Sit to stand Front Posterior Lateral Stand to sit Front Posterior Lateral Walk Front Posterior Lateral Stoop Front Posterior Lateral Jump Front Posterior Lateral Walk backward Posterior Lateral Lie on the bed (30 cm height) Turn the body then fall to the ground detector that used two sensors (Analog Device ADXL0) to measure the acceleration [9]. Because of the complex movement of the wrist, the accuracy was just about 65%, and only three kinds of falls could be detected. The third system employed a tri-axial accelerometer (KXM5-1050, Kionix, Inc.) worn on the waist to provide highly precise identification of postures and dynamic movements [10]. But the high time-complexity of its algorithm resulted in erroneous detections during rapid daily movements. Lindemann et al. s pilot study proposed integrating two accelerometers (ADXL50QC) into a hearing-aid housing [11]. The system could recognize 7 kinds of falls and 5 kinds of daily movement accurately. It is a considerable design because the sensors are placed at the head level and the accuracy of fall recognition was raised beyond that of previous studies [8-10]. However, the integration period was set to 1.5 seconds to calculate the reference velocity. The period setting was an empirical value that the authors defined as the duration from the body s initial contact on the ground to the body is at rest. This static period is obviously not suitable in all kinds of unintentional falls and could be mistaken, possibly leading to an erroneous judgment. We attempt to distinguish more kinds of falls and daily activities for the elderly residents. As mentioned above, we put the fall detection sensor at head level and propose a novel algorithm to provide a robust and precise falls detection that can inform the accidents to the relevant personnel in real time...3 Fall detection algorithm Accelerometers have been used in various studies to monitor a range of human movements [8-13]. The preferred sensor placement is at head-level, which is the most sensitive part of fall detection. It is human nature to protect the head against high acceleration. Therefore, high head-level acceleration means the body is doing an unintentional activity like falling. Sensors placed at the head-level not only eliminate unnecessary interference from the body movement but also raise the detection sensitivity. According to the sensor s placement, the x-axis is frontal direction, the y-axis is vertical side, and the z-axis is equivalent to sagittal side. A 6 G tri-axial accelerometer
Fall Detection System for Healthcare 49 Start Acceleration in x-,y-, z- direction No No No SVM a > 6G? No S h > G? SVM ai - SVM ai+1 <=0.045G, i = 0 ~59 V ref > m/s? Alarm Figure 1. The fall detection algorithm. (MMA760Q, ± 6 G) is used to measure three axial accelerations and transform the values into voltages, and then store the values in the embedded flash memory. The sample rate of the accelerometer is 00 Hz. The sum-vector of axial accelerations is denoted as SVM a (sum vector magnitude of accelerations) which was adopted as the threshold in fall detection. Let SVM a be defined as: SVM a a a a (1) x y z where a x, a y and a z are the accelerations of the x-axis, y-axis, and z-axis, respectively. It is used to describe the spatial variation of acceleration during the falling interval. Because the SVM a of normal daily activities are all under 6 times the acceleration of gravity (G) [1,13], the value of the first threshold of SVM a is 6 G. The variation of acceleration on the horizontal x-z plane, S h, which indicates the body is tilting forward or backward, is calculated by: S h x z a a () During the fall, the acceleration on the horizontal plane will be larger than G. In order to distinguish falling from ordinary daily activities, the threshold S h is set to G. Moreover, the reference velocity is defined as: T t V ti SVM 1 dt (3) ref Trs a where T ti is the time when the body begins to tilt and T rs is the time that body is at rest. Before the integration, the acceleration component that is due to gravity must be removed from SVM a. It is worthy of note that the reference velocity interval ranges from T ti to T rs. T ti is the first time that S h is larger than G, and then we use the continuous 60 data points of stable SVM a to estimate whether the faller is at rest or not. Severe injuries might occur if the reference velocity of the head is more than m/s because of the fragility of the head. The threshold of the V ref is set to m/s, which can distinguish some daily activities involving violent motions from falls. The algorithm is shown in Figure 1. First it calculates SVM a continuously; as soon as SVM a is lager than 6 G, the proposed system will give alarm directly. If S h is bigger than G, it will use the continuous 60 data points of stable SVM a to estimate whether the faller is at rest or not. If the faller is at rest, it will go into the integral transforms of V ref. Finally, the proposed system will give alarm when V ref is over m/s..3 Indoor location awareness Advances in ubiquitous computing technologies have successfully supported the development of location-based service. Location awareness is one of the widespread context-aware applications that can know where incidents have happened and respond immediately. Several techniques have been developed for indoor positioning systems, such as radiofrequency, ultrasound, infrared and magnetic field. This has lead to various positioning methods such as trilateration, multilateration, and received signal strength indicator (RSSI), etc. [14,15]. In this study, we adopted the TI CC431ZDK as a solution for wireless indoor positioning engine [16]. The ZigBee nodes are divided into three types, a tracking node, reference nodes and a gateway. Tracking node can receive the signals (coordinates X, Y, and RSSI) from neighbor reference nodes, calculate its location, and send out its location information. Reference nodes that are set in the indoor environment fixedly give the tracking node its position information and deliver the fall signal. The fall signal can be delivered to the embedded system through a gateway which communicates the entire networks. For the indoor location procedure, based on the TI ZigBee location engine, the RSSI value is determined by a function: RSSI ( 10 N log 10 d A) (4) where the value A is the average RSSI measured one meter long between the sender and the receiver by all directions, the N is a parameter that describes how the RSSI value decreases and the value d is the distance between the sender and the receiver. As soon as a fall happens, the system will start the indoor location procedure. First of all, we must find the suitable environment parameters, A and N. The ZigBee indoor location procedure includes four stages, as shown in Figure : (1) The tracking node sends out a 1-hop broadcast to get the addresses of all reference nodes that are within the range the RSSI signal
50 J. Med. Biol. Eng., Vol. 30. No. 4 010 reaches. () The tracking node sends out a 1-hop blast broadcast message to these reference nodes. (3) These reference nodes feedback the mean RSSI values that they have received and their own coordinates to the tracking node. (4) The tracking node uses the mean RSSI of these reference nodes to calculate its own position, then it sends the calculated relative coordinate back to the ZigBee gateway. ZigBee reference nodes Fall happens ZigBee tracking nodes Event information ZigBee gateway 3. Results 3.1 Hardware We have integrated the modular tri-axial accelerometer with battery and TI ZigBee tag which contains a CC431 micro-controller unit (MCU) and an antenna into a fall detection sensor. The architecture diagram of the fall detection sensor is shown in Figure 4. The proposed fall detection sensor is about 4 cm (length) 4 cm (width) 1.5 cm (height). Figure 5(a) is the modular circuit board of the fall detection sensor; the upper one is the TI ZigBee tag, and the other is the tri-accelerometer with battery board. Comparing with the TI ZigBee tag with battery board, as shown in Figure 5(b), the plane size of the fall detection sensor is one third, and by changing the TI ZigBee tag s antenna into the chip antenna, the height is reduced by a factor of 6. Calculates its own position Alarm Figure. Signaling of indoor location awareness scheme..4 The scenario of fall detection system In residential care facilities, most residents are frail and aged, which increases the frequency of falls happening when the residents are single. Since fall accidents cannot be completely avoided in residential care facilities, we developed a fall detection system to deliver fall alarms and the accident locations for improving the quality of care facilities. The scenario and signal delivery are shown in Figure 3. When a single resident falls in room 6, the fall detection sensor detects the event and triggers the position engine. After calculating the position information, the tracking node delivers the location information and fall event to the gateway through reference nodes. Finally, the system sends the alarm and lets the caregivers know by displaying the message XXX has fallen in Room 6! on the graphical user interface. Then caregivers give help immediately, which can reduce severe injury and medical treatment cost. RF transceiver 8051 core Figure 4. The architecture diagram of the detector. Figure 5. (a) The modular design of fall detection sensor. (b) The fall detection sensor compared to TI ZigBee tag with battery board. 3. Fall detection experiment Figure 3. The scenario and signal delivery of fall detection system. 3..1 Experiments on daily activities The SVM a comparisons between daily activities of 6 types are shown in Figure 6. Obviously, the SVM a of daily activities are smaller than 6 G. Some specific movements, such as jumping and fast lying down on the bed, can easily cause
Fall Detection System for Healthcare 51 erroneous judgment of fall. Since the V ref of jumping and fast lying on the bed are 1.98 m/s and 1.69 m/s, respectively, it will make a correct fall-detection judgment by using our proposed algorithm. Table 3 shows the times of experiment on each daily activity which demonstrates a 100% correctness of returning judgment on daily activities. 7 6 5 4 (G) 3 1 0 Male Female Stand Sit Lie Lie on the bed bed Walk Jump Go Go up up and down and Run down stairs stairs Figure 6. The mean and variation of SVM a on daily activities. Table 3. The daily activities experiment results. Daily activities Times Thresholds Error Males Females < G < m/s detections Stand 6 4 9 1 0 Sit 6 4 9 1 0 Lie on the bed 6 4 7 3 0 Walk 6 4 7 3 0 Jump 6 4 4 6 0 Go up and down stairs 6 4 8 0 Run (18 meters) 3 3 0 (width) corridor. There are two reference nodes in the corridor, four reference nodes in the corner of classroom, and two reference nodes on the border between Region 1 and Region, as shown in Figure 7. Table 5 shows the error distance with different environment parameters. The minimum error distance is about 0.5~1.06 m when A is set to 71 and N is equal to 1.75. The actual condition of our experiment is that the participant falls near the door of the Region, the location awareness fall detection system can detect the position of the fall incident and display the information on the screen of the embedded system. In addition, the system also correctly detects the fall event on the door side of Region, as shown in Figure 7. For the purpose of verifying the accuracy of indoor location awareness, one participant fell on the boundary between Region 1 and Region, where there was no partition between the two regions. Approximately 1 m error distance is displayed in Figure 8. According to the application environment, residential care facilities, it can tolerate the error in the region space. Table 5. The error distance with different environment parameters. Error distance (m) A = 71 N = 1.75 N = 1.50 N = 1.750 N = 1.875 A = 66 A = 71 A = 76 1.5 0.75 0.7 0.85 0.75 1.08 ±1.4 ±0.3 ±0.60 ±0.40 ±0.3 ±0.8 3.. Fall experiments Table 4 lists the results of fall experiments indicating precise fall detection mechanism could detect falls precisely. The table shows a 100% correctness of the returning fall-detection judgment. Except for falling from lying on the bed, the probability of detection on the other seven types of falls through the first threshold 6 G exceeded 50%. Even though some fall postures were hard to detect, the algorithm could perfectly distinguish the fall from daily activities based on the threshold S h and V ref. Fall Table 4. The fall experiment results. Times Thresholds Error detections Males Females > 6G > G & > m/s Stand 9 6 1 3 0 Sit to stand 9 6 8 7 0 Stand to sit 9 6 1 3 0 Walk 9 6 7 8 0 Stoop 9 6 10 5 0 Jump 9 6 10 5 0 Walk backward 6 4 6 4 0 Lie on the bed* 3 1 4 0 * 30 cm height 3..3 Location-aware fall detection system The experimental environment is a 14 m (length) 8.65 m (width) space that includes a classroom separated into two 7 m (length) 5.65 m (width) regions and a 14 m (length) 3 m Figure 7. The result of location-aware fall detection in Region. Figure 8. The result of location-aware fall detection on the border between Region 1 and Region.
5 J. Med. Biol. Eng., Vol. 30. No. 4 010 4. Discussion Modular design can make the development of fall detection sensors more convenient. In future work, we will integrate the two modules into a circuit that can decrease the size of the fall detection sensor into a hearing-aid housing. The location awareness system is accomplished by a ZigBee wireless sensor network. To solve the sensor nodes power consumption problem, we have designed an event trigger mode in which the ZigBee tags wake up from sleep mode when the fall event triggers the location engine. It can save the power consumption significantly. In our experiment, the young subjects fell onto a soft mat and tried not to get hurt by their natural reaction. It is noted that real falls are almost always onto hard surfaces, so the value of acceleration in a real fall is higher than that in the experimental situation. The threshold setting in fall detection algorithm will be suited to the real conditions. Furthermore, we expect to perform the daily activities measurements for elder people in the future work, which could verify the proposed fall detection algorithm. To provide accurate fall detection, we put the sensor at head level. However, not all residents need to wear the hearing aid. In order to make the user more comfortable, we have tried to change the wearing position, like putting the sensor on the waist band. Changing the wearing position, the thresholds of fall-detection algorithm need to be re-evaluated in future work. Due to the influence from the changing of the environment, the drift of RSSI signal affects the positioning seriously, as shown in Figure 7. Although the position difference was tolerated in the experimental environment, a more accurate position can let caregivers arrive at the incident location directly. An adaptive tracking algorithm might be able to raise the accuracy of the location awareness scheme. 5. Conclusions Long-term care in residential care facilities has become a more popular care model for geriatrics. More than half of facility residents fall every year, and there is a high rate of recurrence in residential care facilities. Reliable and effective fall management is an important topic for improving the healthcare quality in residential care facilities. In this work, we have proposed a fall detection system that provides the immediate position information to the caregivers as soon as the fall happens. The system consists of two subsystems, fall detection system and location awareness system. The experimental results have demonstrated that the system provids high accuracy, reliability and sensitivity for fall detection. The accuracy of the location information is bounded in the tolerated region. The proposed system fulfills the requirements of delivering critical information to caregivers and can improve the healthcare quality in residential care facilities. Acknowledgements This work was supported in part by grants from the National Science Council (NSC 98-18-E-41-004) and the Ministry of Economic Affairs (96-EC-17-A-31-I5-001). References [1] International data base (IDB) of the U. S. Census Bureau, Population Division, available: http://www.census.gov/ipc/www/idb/informationgateway.php [] P. Kannus, J. Parkkari, S. Koskinen, S. Niemi, M. Palvanen, M. Järvinen and I. Vuori, Fall-induced injuries and deaths among older adults, JAMA-J. Am. Med. Assoc., 81: 1895-1899, 1999. [3] P. Kannus, J. Parkkari, S. Niemi and M. Palvanen, Fall-induced deaths among elderly people, Am. J. Public Health, 95: 4-44, 005. [4] L. Z. Rubenstein, K. R. Josephson and A. S. Robbins, Falls in the nursing home, Ann. Intern. Med., 11: 44-451, 1994. [5] M. Butler, N. Kerse and M. Todd, Circumstances and consequences of falls in residential care: the New Zealand story, J. N. Z. Med. Assoc., 117: 10, 004. [6] K. Hill, J. Schwarz, A. Kalogeoppulos and S. Gibson, Fear of falling revisited, Arch. Phys. Med. Rehab., 77: 105-109, 1996. [7] P. Kannus, H. Sievanen, M. Palvanen, T. Jarvinen and J. Parkkari, Prevention of falls and consequent injuries in elderly people, Lancet, 366: 1885-1893, 005. [8] C. N. Huang, C. Y. Chiang, J. S. Chang, Y. C. Chou, Y. X. Hong, S. J. Hsu, W. C. Chu and C. T. Chan, Location-aware fall detection system for medical care quality improvement, Proc. IEE Int. Conf. on Multimedia and Ubiquitous Engineering, 477-480, 009. [9] D. Thomas, J. Heinz, R. Michael and W. Stefan, SPEEDY: a fall detector in a wrist watch, Proc. IEEE Int. Symp. on Wearable Computers, 184-187, 003. [10] C. C. Yang and Y. L. Hsu, Development of a portable system for physical activity assessment in a home environment, Int. Comput. Symp. 006, Taipei, Taiwan, 006. [11] U. Lindemann, A. Hock, M. Stuber, W. Keck and C. Becker, Evaluation of a fall detector based on accelerometers: a pilot study, Med. Biol. Eng. Comput., 43: 548-551, 005. [1] C. C. Wang, C. Y. Chiang, P. Y. Lin, Y. C. Chou, I. T. Kuo, C. N. Huang and C. T. Chan, Development of a fall detecting system for the elderly residents, Proc. IEEE Int. Conf. of Bioinformatics and Biomedical Engineering, 1359-136, 008. [13] C. V. Bouten, K. T. Koekkoek, M. Verduin, R. Kodde and J. D. Janssen, A triaxial accelerometer and portable data processing unit for the assessment of daily physical activity, IEEE Trans. Biomed. Eng., 44: 136-147, 1997. [14] H. Liu, H. Darabi, P. Banerjee and J. Liu, Survey of wireless indoor positioning techniques and systems, IEEE Trans. Syst. Man Cybern. Part C-Appl. Rev., 37: 1067-1080, 007. [15] Y. Gu, A. Lo, I. Niemegeers, A survey of indoor positioning systems for wireless personal networks, IEEE Communication Surveys & Tutorial, 11: 13-3, 009.