//Logo Image
AuthorsChe-Chang Yang, Yeh-Liang Hsu(2008-12-19)recommendYeh-Liang Hsu(2010-06-06)
NoteThis paper is published on Telemedicine and e-Health, vol. 15, no.1, pp.1035-1045, January, 2009.

Development of a wearable motion detector for tele-monitoring and real-time identification of physical activity

Abstract

Characteristics of physical activity are indicative of one’s mobility level, latent chronic diseases and aging process. Current research has been oriented to provide quantitative assessment of physical activity with ambulatory monitoring approaches. This study presents the design of a portable microprocessor-based accelerometry measuring device to implement real-time physical activity identification. An algorithm was developed to process real-time tri-axial acceleration signals produced by human movement to identify targeted still postures, postural transitions, and dynamic movements. Fall detection was also featured in this algorithm to meet the increasing needs of elderly care in free-living environments. High identification accuracy was obtained in performance evaluation. This device is technically viable for tele-monitoring and real-time identification of physical activity, while providing sufficient information to evaluate a person’s activity of daily living and her/his status of physical mobility. Limitations regarding real-time processing and implementation of the system for tele-monitoring in the home environment were also observed.

Keywords: physical activity, fall detection, accelerometry, tele-monitoring, home telehealth

1.     Introduction

Physical activity can be regarded as any bodily movement or posture produced by skeletal muscles and resulting in energy expenditure [Capspersen et al., 1985]. Aging and various aspects of behavioral variables directly affect one’s physical activity level [Pennathur et al., 2003]. For elderly people, changes in physical activity are highly related to the transition from relatively independent living to ill and declined functional health status. This transition process is often subtle and cannot be easily perceived by care-givers or even the elderly themselves [Celler et al., 1995]. Therefore, it is important to provide objective and accurate assessment approaches to one’s long-term physical activity level in a free-living environment (e.g., home) for evaluation of quality of life and overall functional health status [Lyons et al., 2005].

The difficulties in assessing physical activity are primarily due to the subtle and complex nature of body movement, which requires accurate and reliable measuring techniques. A majority of assessment methods in the past relied on questionnaires or observation by regular visits, which may lead to inconsistent assessment results and inability to be applied in large-scale studies [Meijer et al., 1991]. Advances in sensors, microprocessors and wireless communication technologies have been the driving factors to facilitate continuous tele-monitoring of physical activity [Aminian et al., 2004]. Accelerometers have been regarded as appropriate motion sensors for the use of human movement measurement. These accelerometry-based systems can provide significant information on the details of physical activities, and further quantify the subject’s mobility level at acceptable cost.

Most traditional motion sensor-based monitoring systems capitalize on off-line data processing techniques to analyze recorded human movements [Najafi et al., 2002, 2003]. In these systems, portable data loggers with sensor units are commonly used to record the continuous data of human movement. The recorded data is then uploaded to a personal computer (PC) and is analyzed by complex signal processing techniques, such as fast Fourier transform (FFT), wavelet transform or other frequency-domain processing methods. However, these off-line systems are unable to provide real-time identification of human movements, let alone tele-monitoring of physical activities.

Activity monitoring with real-time signal processing has been developed recently. Mathie et al [2003] proposed a generic framework for automated human movement classification using accelerometry data. The movement classification within that framework does not require complex signal processing, making it applicable to embedded systems and portable devices. Karantonis et al [2006] also presented a real-time system for physical activity classification in home environments. All the movements are classified in the real-time processing of a wearable system. However, a PC-based program is still required to identify walking. The advantages of such approaches are the capability of providing not only real-time monitoring data but also the potential of instant and timely response to emergent events [Scanaill et al., 2006].

Home telehealth is an emerging research area which utilizes information and communication technologies to enable effective delivery and management of health services at a patient’s residence. Home telehealth systems can provide access to the users’ monitored data for users themselves, contributing to health self-management and reduction in in-person home visits [Bensink et al., 2006; Finkelstein et al., 2006].

A centralized framework is used in most home telehealth systems, in which a centralized database is used for data storage and analysis. Figure 1 shows the structure of the decentralized home telehealth system developed by the authors [Hsu et al., 2007]. Instead of using a centralized database that gathers data from many households, a single household is the fundamental unit for sensing, data transmission, storage and analysis. The core of the system is the “Distributed Data Server (DDS)” inside a household, which is a thin server designed specifically for the decentralized home telehealth system. As shown in Figure 1, sensing data from sensors embedded in the home environment are transmitted to the DDS, and are then processed and stored in the Multi-Media-Card (MMC) of the DDS. Authorized remote users can request data from the DDS using an Internet web browser (running Java applets) or a Visual Basic (VB) program. Event-driven messages (mobile phone text messages or e-mails) can be sent to specified caregivers when an urgent situation is detected.

Figure 1. The structure of the decentralized home telehealth system

There are several advantages of the decentralized structure over the traditional centralized database structure:

(1)    The scale of the home telehealth system is much smaller, which makes it economically viable and acceptable to the end-users.

(2)    Instead of sending the health monitoring data to a centralized database in a home healthcare provider, health monitoring data are stored within the household. Only authorized caregivers can access the data. Privacy is better protected.

(3)    The route from the sensor to server is much shorter. Data transmission is easier and more reliable. When the Internet communication fails, the local system can still function normally and keep collecting data. Thus data integrity is better preserved.

(4)    The home telehealth provider only needs to maintain an application server to provide email, short message, and Java services. The DDS can also be used as a gateway if a centralized database is needed.

Several tele-monitoring applications have been developed based on this decentralized structure by the authors, including a portable device for tele-monitoring of snoring and obstructive sleep apnea syndrome (OSAS) [Cheng et al., 2008], a portable device for tele-monitoring of physical activities during sleep [Cheng et al., 2008], and a telepresence robot for interpersonal communication with the elderly in a home environment [Tsai et al., 2007].

This study presents the design of a portable microprocessor-based motion detector to implement tele-monitoring and real-time identification of physical activity based on the decentralized home telehealth platform described above. An algorithm is developed for real-time physical activity identification in free-living environments, including 3 still postures (sitting still, standing still and lying still), 4 postural transitions (sit-to-stand, stand-to-sit, lie-to-sit and sit-to-lie), and 2 dynamic movements (turning on bed and walking). In addition, possible falls can also be detected by this algorithm. Good identification accuracy was obtained in the performance evaluation. Limitations regarding real-time processing and tele-monitoring of physical activities were also observed during the performance evaluation and implementation of the system in the home environment.

2.     System development

2.1 Instrumentation design

Referring to Figure 1, the “sensor” in the physical activity monitoring system developed in this research is a wearable motion detector shown in Figure 2. This device mainly consists of a PIC (programmable interface controller) microprocessor (PIC18F6722, Microchip Inc.), an RF encoder/transmitter (PT2262, Princeton Technology, TWS-CS-2, Wenshing Electronics Co., LTD, Taipei Taiwan) and a motion sensor module. The device is packaged in a 100mm×60mm×25mm exterior (excluding the battery cartridge) and is 140g in total weight. The motion sensor module utilizes a miniature MEMS-fabricated tri-axial accelerometer (KXM52-1050, Kionix Inc., Ithaca, NY), which functions on the principle of differential capacitance in response to motion and constant gravity to measure both acceleration and inclination. The motion detector is attached to the pants’ belt at waist level. Waist level has been considered as an appropriate position for motion sensor attachment because it is close to the center of gravity of the body [Bouten et al., 1997; Meijer et al, 1991; Karantonis et al., 2006].

As shown in Figure 2, the analog signals of the accelerometer outputs are initially low-pass filtered at 50Hz to limit the signal bandwidth in response before the following 60Hz analog-to-digital conversion (ADC) data sampling process. Real-time computation is processed in the PIC microprocessor to cyclically identify one of the targeted physical activities. Each identified activity is immediately transmitted to the DDS via wireless RF amplitude shift keying (ASK) 433.92MHz for data storage.

Figure 2. The functional diagram and the configuration of the wearable motion detector

2.2 General processes of the algorithm

As described earlier, the algorithm developed in this research aims to identify 3 still postures (lying still, sitting still, and standing still), 4 postural transitions (sit-to-stand, stand-to-sit, sit-to-lie and lie-to-sit), and walking. Detection of possible falls is also featured in this algorithm. Figure 3 shows the main algorithm flowchart. There are 5 processing steps: sampling (Cx), pre-processing (Px), dynamic postural transition identification (DBx), still posture identification (DAx) and possible fall detection (DCx). Each data sample is processed in the time domain to generate one of the 9 events described above. In the case where there is no definite result determined by the algorithm, the event will be recorded as an uncertain movement or an uncertain posture.

Referring to Figure 3, acceleration and trunk tilt data in x-, y-, and z-axis are collected in the data sampling process which consists of the primary stage C1 (0.5s) and the secondary stage C2 (1.5s). The use of this dual-stage data sampling strategy ensures that the data of a dynamic event can be captured within the sampling interval. After the primary data sampling stage C1, decision D1 determines whether any sign of dynamic movement exists by inspecting apparent acceleration changes in each axis. If no dynamic movement is detected (D1=No), the 0.5s trunk tilt data is used to identify one of the three possible still postures in the processes DAx. If a dynamic movement is detected (D1=Yes), the secondary data sampling stage C2 is activated to collect the subsequent 1.5s data which follows the previously buffered 0.5s data in C1. The median-filtering (length n=3) and moving-averaging are also applied to the C1-C2 combined data to reduce signal spikes and high frequency noises, respectively. The resulting processed data P represents a dynamic event for the following step-by-step identification processes.

 

Figure 3. Main algorithm flowchart

2.3 Identifying dynamic activities

The “slope mapping” technique which functions as a high-pass filter is used in many steps in the algorithm to register whether there are apparent changes in the data sequences. This process maps the data P into a binary sequence C according to Equation (1) and (2). The entry ci in C is assigned “1” if the “slope” si between two consecutive data points  and  exceeds a specific threshold sthr. Otherwise, ci is assigned “0”. Figure 4 shows an example of applying the slope mapping technique to a trunk tilt data of a sit-to-stand transition. Note that the binary sequence is expressed as a histogram to highlight the entries in “1”.

        , i=1,2,3,…,149                                                                (1)

        , i=1,2,3,…,150                                                    (2)

Figure 4. An example of applying slope mapping to a trunk tilt data

The process D3 and DA1 examine the trunk tilt values according to the trunk tilt orientation shown in Figure 5. Decision D3 is used to examine the end posture (trunk tilt) state from the data segment of 1.5s to 2.0s out of P. An upright state here is defined as the trunk tilt within the range of 60°forward and backward with reference to vertical. Decision DA1 determines a lying posture which is defined as the trunk tilt within the range of 60°relative to the horizontal, or within the range of 60° to 120° as illustrated in Figure 5.

Figure 5. Illustration of the trunk tilt orientation

(1)   Identifying sit-stand postural transitions

The acceleration in vertical direction is used to identify sit-stand postural transitions because the y-axis signal is more sensitive to external motions through empirical tests. Coherent properties of the acceleration patterns were also observed in an extensive test with 15 recruited ostensibly healthy subjects of various ages (11 males and 4 females, 5 young subjects under 35 years old, 5 middle-aged subjects between 35 to 65, and 5 elder subjects over 65). The acceleration patterns of sit-stand postural transitions can be characterized by its peak order, peak distance and peak amplitudes. Figure 6 shows an example of a vertical acceleration pattern during sit-to-stand and stand-to-sit postural transitions in slow, normal and fast motion. In this figure, it can be observed that a positive acceleration peak appears first followed by a negative acceleration peak for sit-to-stand. A reverse order for stand-to-sit was also observed. The peak amplitude is defined as the maximum positive or negative acceleration in the data sample, and the peak distance is the time interval (in seconds) between the positive peak and the negative peak. Faster transition generates shorter peak distance with higher peak amplitudes, and slow transition generates longer peak distance with lower peak amplitudes.

Figure 6. Examples of vertical acceleration patterns of slow, normal and fast stand-to-sit transition

As shown in the main algorithm flowchart in Figure 2, the following decisions DB1 and DB2 identify a sit-stand postural transition if a dynamic movement with upright trunk posture is recognized (D2=Yes, D3=Yes). Figure 7 further illustrates the decomposed procedure of the decisions DB1 and DB2 which embodies several required thresholds and parameters for distinguishing sit-stand transitions. The first step (R1) is to check the type of peak orders. The following two steps (R2 and R3) check the peak distance (DP) and peak amplitudes (positive peak VP and negative peak VN) to further confirm whether the dynamic event is indeed a sit-to-stand or stand-to-sit transition. The threshold values used in this identification process were experimentally obtained in the test with 15 subjects mentioned above. R1, R2 or R3 must be all satisfied to identify sit-to-stand or stand-to-sit transition.

 

Figure 7. The decomposed flowchart of sit-stand postural transition identification

(2)   Identifying walking movement

Figure 8 shows an example of the acceleration pattern in vertical direction at level walking. As shown in this figure, walking movement is characterized by its fast and repeated oscillating changes in the vertical acceleration. Previous studies revealed that the frequencies of normal level walking are between 0.7Hz to 4Hz with the peak acceleration values between 0.4g to -0.3g [Aminian et al., 2002; Mathie et al., 2003; Nyan et al., 2006; Sekine et al., 2000]. Those parameters were adopted in this identification process. The decision DB3 also uses slope mapping which processes the vertical acceleration data to identify walking movement if DB2 fails to recognize a stand-to-sit postural transition.

 

Figure 8. An example of acceleration patterns in vertical direction at level walking

(3)   Identifying lie-sit postural transitions

Tilt data and its end posture state (trunk orientation) are required to identify and distinguish lie-sit type postural transitions. By applying a slope mapping technique, a lie-to-sit transition produces positive increments in trunk tilt variation and ends with upright posture. On the contrary, a sit-to-lie transition produces negative increments in trunk tilt variation and ends with lying posture. Note that reverse stance (handstand) posture is not considered in the trunk tilt orientation. Decision DB4 identifies whether a lie-to-sit transition if walking movement is not recognized in DB3. For a dynamic movement with recognized lying end state (D2=Yes, D3=No), DB5 is used to identify sit-to-lie transition if the sign of fall is not recognized (DC1=No).

2.4 Identifying possible falls and still postures

A fall can be intuitively regarded as a movement accompanied by unusually high acceleration peaks in a short time interval, which is followed by a period of lying still posture. Therefore, fall detection in this algorithm is designed with two identification stages. For the first identification stage, the tri-axial acceleration data of the dynamic (D2=Yes), non-upright (D3=No) posture event is processed in the decision DC1 to determine whether a sign of fall exists. The sign of fall is defined in this algorithm to indicate an existence of peak acceleration greater than Vf=±1.5g captured in the tri-axial acceleration data. For the second identification stage, the algorithm starts to register the duration of lying still posture in the following identification loop. Decision DA1 examines the trunk tilt data to determine a lying posture. The previously recognized sign of fall event can be raised to indicate a possible fall if prolonged TD=20s period of lying still is detected.

Identifying and distinguishing sitting still or standing still posture requires not only the trunk tilt orientation data, but also the information of previously identified postural transitions or walking movement. If a still event is not lying posture (DA1=No) and its previous event is either a sit-to-stand transition or walking movement, this event can be recognized as standing still (DA3=Yes). On the other hand, the event is sitting still if the previous event is either a stand-to-sit or lie-to-sit transition (DA2=Yes).

3.     Performance evaluation of the system

3.1 Basic performance test: sensitivity and specificity

In order to evaluate the performance of the device in identifying still postures and dynamic activities, 10 subjects were recruited for the laboratory-based test. Sensitivity and specificity for identifying postural transitions (sit-stand transitions, lie-sit transitions) and walking were recorded. In the sensitivity test, the subjects were asked to perform separate test items according to the facilitator’s instruction. The test items include the reciprocal postural transitions (sit-stand transitions and sit-lie transitions) and continuous walking movements. The subjects repeated each postural transition 20 times. In addition, the sensitivity of walking was obtained from 50s of continuous walking steps performed by each subject.

For the specificity test, 5 test items were arranged into phase A (sit-to-stand and sit-to-lie), phase B (stand-to-sit and walking) and phase C (lie-to-sit). According to the algorithm, when a subject is sitting still, his possible posture transition from that state is either sit-to-stand transition or sit-to-lie transition. Therefore in phase A, the subjects were initially in the sitting still posture and they can perform any movement other than the sit-to-stand or sit-to-lie transitions. Similarly in phase B, the subjects were initially standing still, and then performed any movements other than stand-to-sit transition or walking movement. In phase C, the subjects performed any movements other than lie-to-sit transition as they were initially lying. Each subject performed a total of 50 movements Note that during the test the subjects were allowed to vary their movement speeds or time intervals arbitrarily, and they were informed to avoid performing ambiguous movements. For example, in phase A, the subjects were asked to avoid performing a movement similar to sit-to-lie transition even though they subjectively considered it a free movement.

Good identification accuracy was obtained in the performance evaluation. Table 1 shows the sensitivity and specificity obtained from 200 and 500 data samples respectively. Note that the evaluation did not include sitting still or standing still postures because both still postures are associated with the results of previously identified postural transitions or walking.

Table 1. Evaluation of sensitivity and specificity of the algorithm in identifying still postures and dynamic movements

Posture/activity

Sensitivity (%)

Specificity (%)

Sit-to-stand

92.2

91.5

stand-to-sit

95.6

88.5

Sit-to-lie

92.2

99.5

Lie-to-sit

95.6

88.0

Walking

98.9

99.5

3.2 Demonstration of long-term tele-monitoring at home

The physical activity monitoring system developed in this research was implemented in a home environment to demonstrate the possibility of long-term tele-monitoring of physical activity at home. Figure 9 is the activity chronograph which shows the continuously recorded activity data from a test subject at home in a typical day. This figure contains 6.75-hour data in the 15-hour monitoring period. The tester could not wear the sensor continuously in some situations, such as taking a shower, going outside, etc., therefore the vertical time axis is not continuous. In this figure, lines of various lengths represent respective events (event 1 to 9) identified by the system. For example, the subject was lying from 4:00 to 9:00, and mostly sitting from 9:00 to 12:00.

The recorded activity data can be statistically classified as in Figure 9, which shows the numbers and percentages of each dynamic activity or static posture within the entire monitoring period. The subject also made a report to register his actual activities during this period, and the monitored data reveals good correlation with the report. While the chronograph in Figure 9 shows the pattern of the subject’s daily activity, the quantitative data in Figure 10 can be used to form indices of the mobility level of the subject.

Event number 1: lying still, no.2: sitting still, no.3: standing still, no.4: walking,

no.5: sit-to-stand, no.6: stand-to-sit, no.7: lie-to-sit, no.8: sit-to-lie, no.9: possible fall

Figure 9. Example of activity chronograph of the continuously recorded data in the ambulatory monitoring at home

Figure 10. Statistical results of the recorded data

4.     Discussion

The real-time system using a single wearable motion detector for physical activity identification was presented in this paper. A miniature tri-axial accelerometer was used as the motion sensor due to its superiority of accurate acceleration/tilt sensing, fast response and low power consumption. The identification algorithm was specifically developed for microprocessor-based wearable systems to identify the target human static postures, level walking and dynamic postural transitions. The potential capability of fall detection was also presented. Good identification accuracy was obtained in the performance evaluation, though the nature of actual human movements in daily living is more complex than what is considered and characterized in the algorithm.

This system is based on the decentralized home telehealth system structure described in the first section, which can be easily implemented in a home environment and is economically viable and acceptable to the end-users. This system can be used for long-term physical activity and mobility monitoring in the home environment. As depicted in Figure 1, alert messages will be sent out to the caregivers in the form of mobile phone text messages (short message service, SMS) or e-mails if potential urgent situations are detected (falls, lack of movement, etc.). Moreover, multiple sensors (wearable motion detectors) with different sensor identification numbers can be connected to the same DDS in the situation when there are more than one users at home.

During the performance evaluation and implementation of the system in the home environment, the following limitations regarding real-time processing and tele-monitoring of physical activities were also observed:

(1)      Tri-axial accelerometer

An advanced capacitance tri-axial accelerometer was used to measure both the acceleration and trunk tilt of the human body. For such a gravity-responsive accelerometer, the most precise tilt sensing can be maintained when the accelerometer is at static, or under constant acceleration. Degradation of tilt sensing accuracy in changing acceleration magnitude is inevitable [Elble, 2005]. Therefore, the acceleration data acquired from the wearable motion detector cannot be used to precisely interpret the complex nature of human motions. In spite of this constraint, tilt sensing and acceleration measurement using one tri-axial accelerometer is still valid for normal physical activity because the resulting sensor outputs still preserve apparent characteristics for either trunk tilt or acceleration patterns.

(2)      The wearable motion detector

The wearable motion detector was designed to be clipped onto a waist/pant belt. However, carrying the device might limit some movements in lying posture and therefore further influences the wearer’s comfort, and the tester could not wear the device in some situations, such as taking a shower, going outside, etc. This has been an important usability issue for fall detector in general, as these are often the situations where users are at great risk of fall at home (getting out of the shower, getting out of bed to go to the bathroom at night, etc.) Usually a special designed waist belt which contains the fall detector (the wearable motion detector in this research) is provided in the situation when the user cannot clip the fall detector to a pant belt. In addition, reduction in the size and weight of the device is also an important design issue in improving the comfort of the users.

Moreover, power capacity is an important consideration for the practicality of such wearable devices. To extend the time for use, a battery cartridge of larger power capacity can be selected. Proper selection of sampling rate in the microprocessor is also important in order to maintain optimal power longevity. For example, when the subject is at static state, the microprocessor should be set at a lower sampling rate because higher signal resolution is not necessary.

(3)      Real-time identification

The algorithm developed in this study has the ability to provide real-time status of physical activity. The microprocessor has sufficient computation capability for the real-time processing. However, because each physical activity must be identified in continuous intervals with discrete short samples, real-time systems usually cannot provide the identification accuracy higher than that of the off-time analysis systems [Karantonis, 2006]. The limitation concerning the identification accuracy of both off-line and real-time systems is the complex nature of human movement. People usually perform mixed and combined movements in their normal activities of living, even the diverse characteristics of one activity item can be observed from the same person. Advances in signal processing techniques such as self-adaptive parameter tuning can be further employed to enhance the identification accuracy.

The system developed in this research has been technically proven to be feasible for continuous tele-monitoring of physical activity at home. This real-time system can provide sufficient information in evaluating the subject’s activities of daily living (ADL) pattern and his/her physical mobility level. The use of the Ethernet-accessible DDS at home also enables remote real-time monitoring via the Internet. This home tele-health system has potential applications in rehabilitation, elderly care and personal health management.

Reference

Aminian, K., Najafi, B., Bula, C. and Leyvraz, P. F., 2002. “Spatial-temporal Parameters of Gait Measured by an Ambulatory System Using Miniature Gyroscope,” Journal of Biomechanics, vol. 35, pp. 689-699.

Aminian, K., and Najafi, B., 2004. “Capturing Human Motion Using Body-fixed Sensors: Outdoor Measurement and Clinical Applications,” Comp. Anim. Virtual Worlds, Vol. 15, pp. 75-94.

Bensink M., Hailey, D., Wootton, R., 2006. “A systematic Review of Successes and Failures in Home Telehealth: Preliminary Results,” Journal of Telemedicine and Telecare, Vol.12, pp. 8-16.

Bouten, C. V., Koekkoek, K. T. M., Verduin, M., Kodde, R., and Janssen, J. D., 1997. “A Triaxial Accelerometer and Portable Data Processing Unit for the Assessment of Daily Physical Activity,” IEEE Trans. Biomed. Eng., Vol. 44, No. 3, pp.136-147.

Capspersen, C. J., Powell, K. E., and Christenson, G. M., 1985. “ Physical Activity, Exercise and Physical Fitness: Definitions and Distinctions for Health Related Research,” Public Health Rep., Vol. 110, pp. 126-131.

Celler, B.G., Earnshaw, W., Ilsar, E.D., Betbeder-Matibet, L., Harris, M.F., Clark, R., Hesketh, T., and Lowell, N.H., 1995. “Remote Monitoring of Health Status of the Elderly at Home. A Multidisciplinary Project on Aging at the University of New South Wales,” International Journal of Biomedical Computing, Vol. 40, pp. 147-155.

Cheng, C. M., Hsu, Y. L., Young, C. M., Wu, C. H., 2008. “Development of a Portable Device for Tele-monitoring of Snoring and OSAS Symptoms,” Telemedicine and e-Health, Vol. 14, No. 1, pp. 55-68.

Cheng, C. M., Hsu, Y. L., Young, C. M., 2008. “Development of a Portable Device for Tele-monitoring of Physical Activities During Sleep,” Telemedicine and e-Health, Vol.14, pp.1001-1013.

Elble, J. R., 2005. “Gravitational Artifact in Accelerometric Measurements of Tremor,” Clinical Neurophysiology, Vol. 116, pp. 1638-1643.

Finkelstein, S. M., Speedie, S. M., and Potthoff, S., 2006. “,Home Telehealth Improves Clinical Outcomes at Lower Cost for Home Healthcare,” Telemedicine and e-health Vol. 12, No. 2, pp. 128-136.

Hsu, Y. L., Yang, C. C., Tsai, T. C., Cheng, C. M. and Wu, C. H., 2007, “Development of a Decentralized Home Telehealth Monitoring System”, Telemedicine and e-Health, Vol. 13, No.1, pp. 69-78.

Karantonis, D. M., Narayanan, M. R., Mathie, M., Lovell, N. H., and Celler, B. G., 2006. “Implementation of a Real-Time Human Movement Classifier Using a Triaxial Accelerometer for Ambulatory Monitoring,” IEEE Trans. Info. Tech Biomed., Vol. 10, No. 1, pp. 156-167.

Lyons, G. M., Culhane, K. M., Hilton, D., Grace, P. A., and Lyons, D., 2005. “A Description of an Accelerometer-based Mobility Monitoring Technique,” Medical Engineering & Physics, Vol. 27, pp. 497-504.

Mathie, M. J., Coster, A. C. F., Lovell, N. H. Celler, B. G., 2003. “Detection of Daily Physical Activities Using a Triaxial Accelerometer,” Medical & Biological Engineering & Computing, Vol. 41, pp. 296-301.

Meijer, G. A. L., Westerterp, K. R., Verhoeven, F. M. H., Koper, H. B. M., and Hoor, F. T., 1991. “Methods to Assess Physical Activity with Special Reference to Motion Sensors and Accelerometers, IEEE Trans. On Biomed. Eng., Vol. 38, No.3, pp. 221-229.

Najafi, B., and Aminian, K., 2002. “Measurement of Stand-Sit and Sit-Stand Transitions Using a Miniature Gyroscope and Its Application in Fall Risk Evaluation in the Elderly,“ IEEE Trans. Biomed. Eng., Vol. 49, pp. 843-851.

Najafi, B., Aminian, K., Paraschiv-Ionescu, A., Loew, F., Bula, C. J., and Robert P., 2003. “Ambulatory system for Human Motion Analysis Using a Kinematic Sensor: Monitoring of Daily Physical Activity in the Elderly,” IEEE Trans. On Biomed. Eng., Vol. 50, No. 6, pp. 711-723.

Nyan, M. N., Tay, F. E. H., Seah, K. H. W., and Sitoh, Y. Y., 2006. “Classification of a Gait Patterns in the Time-frequency Domain,” Journal of Biomechanics, Vol. 39, pp. 2647-2656.

Pennathur, A., Magham, R., Contreras, L R., and Dowling, W., 2003. “Daily Living Activities in Older Adults: Part I- A Review of Physical Activity and Dietary Intake Assessment Methods”, International Journal of Industrial Ergonomics, Vol. 32, pp. 389-404.

Scanaill, C. N., Carew, S., Barralon, P., Noury, N., Lyons D., and Lyons, G. M., 2006. “A Review of Approaches to Mobility Telemonitoring of the Elderly in Their Living Environment,” Annals of biomed. Eng., Vol. 34, No.4, pp. 547-563.

Sekine, M., Tamura T., Togawa T., Fukui Y., 2000. “Classification of Waist-acceleration Signals in a Continuous Walking Record,” Medical Eng. and Phy., Vol. 22, pp. 285-291.

Tsai, T. C., Hsu, Y. L., Ma, A. I., King, T., Wu, C. H., 2007. “Developing a Telepresence Robot for Interpersonal Communication with the Elderly in a Home Environment,” Telemedicine and e-Health, Vol. 13, No. 4, pp. 407-424.