A Review of Accelerometry-Based Wearable Motion Detectors for Physical Activity Monitoring

Characteristics of physical activity are indicative of one’s mobility level, latent chronic diseases and aging process. Accelerometers have been widely accepted as useful and practical sensors for wearable devices to measure and assess physical activity. This paper reviews the development of wearable accelerometry-based motion detectors. The principle of accelerometry measurement, sensor properties and sensor placements are first introduced. Various research using accelerometry-based wearable motion detectors for physical activity monitoring and assessment, including posture and movement classification, estimation of energy expenditure, fall detection and balance control evaluation, are also reviewed. Finally this paper reviews and compares existing commercial products to provide a comprehensive outlook of current development status and possible emerging technologies. Keywords: accelerometry; accelerometer; physical activity; human motion; energy expenditure; gait; fall detection 1. Introduction Physical activity (PA) is regarded as any bodily movement produced by skeletal muscles which results in an energy expenditure [1]. PA has been studied in epidemiological research for investigating human movements and the relationship to health status, especially in the area of cardiovascular OPEN ACCESS

Sensors 2010 7773 diseases, diabetes mellitus and obesity. A declining PA level represents a major factor in multiple illnesses and symptoms related to functional impairment [2]. The organization Healthy People 2020 [http://www.healthypeople.gov/HP2020/] led by the U.S. government has recognized PA as one of the leading health indicators (LHI), which are a measurement of health of a nation’s population. Various methods of subjective and objective PA assessment tools have been developed. Subjective methods, such as diaries, questionnaires and surveys, are inexpensive tools. However, these methods often depend on individual observation and subjective interpretation, which make the assessment results inconsistent [3]. Some standard tests for PA assessment also require subjective judgments. For example, the timed up-and-go test (TUG-T) is a simple test for evaluating one’s ability to perform a sequence of basic activities, and the result of the TUG-T can be a predictor for risk of falling [4]. Distinguishing postural transitions in the TUG-T, however, depends on subjective judgment that counts the time taken for each posture transition. The Berg Balance Scale (BBS), a valid measure to evaluate balance control of the elderly individuals, also requires subjective observation and determination for scoring some test items [5]. On the other hand, objective techniques use wearable, or body-fixed motion sensors, which range from switches, pedometers, actometers, goniometers, accelerometers and gyroscopes, for PA assessment. Mechanical pedometers, or so-called ―step counters‖, are the simplest wearable sensors to measure human motion. The pedometer uses a spring-loaded mass or some other switch mechanism to detect the obvious impacts produced by steps during locomotion. The number of steps during motion can be registered to estimate the distance walked and the energy expenditure. Though pedometers are cheap and simple, the major drawbacks are that pedometers cannot reflect intensity of movement and therefore result in inaccurate energy expenditure estimations [6]. PA can also be objectively measured by means of magnetic systems, optical systems, or video recording. Magnetic and optical systems for PA monitoring are costly and require complex instrumentation and environment setting. Privacy concerns are a major drawback in monitoring systems based on video recording. These systems may not be practical

and therefore result in inaccurate energy expenditure estimations [6]. PA can also be objectively measured by means of magnetic systems, optical systems, or video recording. Magnetic and optical systems for PA monitoring are costly and require complex instrumentation and environment setting. Privacy concerns are a major drawback in monitoring systems based on video recording. These systems may not be practical for monitoring subjects in free-living environments. Accelerometers are sensors which measure the accelerations of objects in motion along reference axes. Measuring PA using accelerometry is preferred because acceleration is proportional to external force and hence can reflect intensity and frequency of human movement. Accelerometry data can be used to derive velocity and displacement information by integrating accelerometry data with respect to time [7]. Some accelerometers can respond to gravity to provide tilt sensing with respect to reference planes when accelerometers rotate with objects. The resulting inclination data can be used to classify body postures (orientations). With these characteristics, accelerometry is capable of providing sufficient information for measuring PA and a range of human activities. Accelerometers have been widely accepted as useful and practical sensors for wearable devices to measure and assess PA in either clinical/laboratory settings or free-living environments [8]. Accelerometers were first investigated in the 1950s to measure gait velocity and acceleration [9]. Accelerometry measurement of human motion was studied in more detail during the 1970s due to technological advances [10]. It was also shown that accelerometers had advantages over other techniques in quantitatively measuring human movement. Micro-electromechanical system (MEMS) technology has reduced the cost of accelerometers in smaller form factors. In the meantime, sensor performance has been enhanced while the power consumption is greatly reduced. The first batch-

Sensors 2010 7774 fabricated MEMS accelerometers were reported in 1979 [11]. Since then various research and commercial applications have used MEMS accelerometers in wearable systems for PA monitoring. This paper provides a comprehensive review on the working principles, capabilities, and various applications of accelerometry-based wearable motion detectors for PA monitoring and assessment. The authors searched for published literature after year 2000 using a range of related keywords such as ―accelerometry‖, ―accelerometer‖, ―wearable‖, ―physical activity‖, ―human motion‖, ―human movement‖, ―activity classification‖, ―energy expenditure‖, ―fall detection‖, ―balance stability‖ and ―gait‖. Selected literatures before year 2000 are also included. This paper first discusses the principles and fundamentals of accelerometry, along with different sensor placements. Various research using accelerometry-based wearable motion detectors for PA monitoring and assessment, including posture and movement classification, estimation of energy expenditure, fall detection and balance control evaluation, are then reviewed. Finally this paper reviews and compares existing commercial products to provide a comprehensive outlook of current development status and possible emerging technologies. 2. Design Fundamentals for Accelerometry-Based Wearable Motion Detectors 2.1. Accelerometry: Principles and Sensors Inertial sensors are basically force sensors to sense linear acceleration along one or several directions, or angular motion about one or several axes. The former is referred to as an accelerometer, and the later a gyroscope. The common operation principle of accelerometers is based on a mechanical sensing element which consists of a proof mass (or seismic mass) attached to a mechanical suspension system with respect to a reference frame. Inertial force due to acceleration or gravity will cause the proof mass to deflect according to Newton’s Second Law. The acceleration can be measured electrically with the physical changes in displacement of the proof mass with respect to the reference frame. Piezoresistive, piezoelectric and differential capacitive accelerometers are the most common types [12,13]. 2.1.1. Piezoresistive accelerometers The sensing element consists of a cantilever beam and its proof mass is formed by bulk- micromachining. The motion of the proof mass due to acceleration can be detected by piezoresistors in the cantilever beam and proof mass. The piezoresistors are arranged as a Wheatstone bridge to produce

differential capacitive accelerometers are the most common types [12,13]. 2.1.1. Piezoresistive accelerometers The sensing element consists of a cantilever beam and its proof mass is formed by bulk- micromachining. The motion of the proof mass due to acceleration can be detected by piezoresistors in the cantilever beam and proof mass. The piezoresistors are arranged as a Wheatstone bridge to produce a voltage proportional to the applied acceleration. Piezoresistive accelerometers are simple and low- cost. The piezoresistive accelerometers are DC-responsive that can measure constant acceleration such as gravity. The major drawbacks of piezoresistive sensing are the temperature-sensitive drift and the lower level of the output signals. 2.1.2. Piezoelectric accelerometers In a piezoelectric accelerometer, the sensing element bends due to applied acceleration which causes a displacement of the seismic mass, and results in an output voltage proportional to the applied acceleration. Piezoelectric accelerometers do not respond to the constant component of accelerations.

Sensors 2010 7775 2.1.3. Differential capacitive accelerometers The displacement of the proof mass can be measured capacitively. In a capacitive sensing mechanism, the seismic mass is encapsulated between two electrodes. The differential capacitance is proportional to the deflection of the seismic mass between the two electrodes. The advantages of differential capacitive accelerometers are low power consumption, large output level, and fast response to motions. Better sensitivity is also achieved due to the low noise level of capacitive detection. Differential capacitive accelerometers also have DC response. Currently this kind of accelerometer has widely been used in most applications, especially in mobile and portable systems and consumer electronics. 2.2. Sensor Placement Gemperle et al. [14] proposed the ergonomic guideline of ―wearability‖ to describe the interaction between the human body and wearable objects. The ―wearability map‖ was generalized to indicate the proper locations of a human body for unobtrusive sensor placement. These locations include the collar area, rear of upper arm, forearm, front and rear sides of ribcage, waist, thighs, shin, and top of the foot. These locations have common characteristics of similar area for men and women, a relatively larger continuous surface, and low movement and flexibility. The sensor placement of wearable devices refers to the locations where the sensors are placed, and how the sensors are attached to those locations. Wearable activity sensors can be placed on different parts of a human body whose movements are being studied. In many cases, it is necessary to measure the whole-body movement. Therefore, the sensors are commonly placed on the sternum [15], lower back [3], and waist [16]. Most studies adopted waist-placement for motion sensors because of the fact that the waist is close to the center of mass of a whole human body, and the torso occupies the most mass of a human body. This implies that the accelerations measured by a single sensor at this location can better represent the major human motion. From an ergonomic point of view, the torso can better bear extra weight when carrying wearable devices. Sensors or devices can be easily attached to or detached from a

and the torso occupies the most mass of a human body. This implies that the accelerations measured by a single sensor at this location can better represent the major human motion. From an ergonomic point of view, the torso can better bear extra weight when carrying wearable devices. Sensors or devices can be easily attached to or detached from a belt around waist level. Therefore, waist-placement causes less constraint in body movement and discomfort can be minimized as well. A range of basic daily activities, including walking, postures and activity transitions can be classified according to the accelerations measured from a waist-worn accelerometer [16-18]. An approach using a chest-worn accelerometer was presented to detect respiratory and snoring features for apnea diagnosis during sleep [19]. Accelerometers can also be attached to wrists, thigh, or ankles. Sleep time duration can be determined from a wrist-worn accelerometer [20] and activity levels during sleep can be measured [21]. Ankle-attached accelerometers can significantly reflect gait-related features during locomotion or walking. Steps, travel distance, velocity, and energy expenditure can be estimated by an ankle-worn accelerometer [22,23]. A special placement in which an accelerometer unit integrated into hearing aid housing was used for detecting falls [24]. The rationale of this sensor placement was based on the author’s hypothesis that the individual intends to protect the head against higher acceleration caused by abnormal activities. Accelerometers have also been placed at the top of head for measuring balance during walking [25,26].

Sensors 2010 7776 Another consideration for sensor placement is how to attach sensors to the human body. Wearable sensors can be directly attached to the skin [15,24], or with some form of indirect attachment by using straps, pant belts and wristbands, or other accessories [20,22,25,26]. Sensors and wearable devices can also be integrated into clothing [27]. In principal, the accelerometers or motion sensors should be securely fitted and attached to the human body in order to prevent relative motion between the sensors and the parts of the human body. Loose attachment or unsecured fit causes vibration and displacement of the wearable systems, and this is liable to produce extraneous signal artifacts and to degrade sensing accuracy. 3. Capabilities of Wearable Systems Using Accelerometry Measurement Accelerometers can be used in ambulatory monitoring to continuously measure long-term activities of subjects in a free-living environment. The recorded longitudinal activity data can be used to identify postures and to classify several daily movements which are related to an individual’s functional status. Signal analysis and algorithm are used to classify daily human movements that are of interest, and adverse activity, such as falls can be detected as well. Important features extracted from posture sway and gait pattern have also been studied for the purposes of evaluating risks of falling and mobility. In addition, energy expenditure is the typical application featured by most commercially available accelerometers. 3.1. Posture and Movement Classification Movement classification using accelerometry-based methodologies has been widely studied. Approaches to movement classification can be threshold-based or using statistical classification schemes. Threshold-based movement classification takes advantage of known knowledge and information about the movements to be classified. It uses a hierarchical algorithm structure (like decision tree) to discriminate between activity states. A set of empirically-derived thresholds for each classification subclass are required. Kiani et al. [28] presented a systematic approach to movement classification based on a hierarchical decision tree that enables automatic movement detection and classification. Mathie et al. [29] further presented a generic classification framework consisting of a hierarchical binary tree for classifying postural transitions, falling, walking, and other movements using signals from a

thresholds for each classification subclass are required. Kiani et al. [28] presented a systematic approach to movement classification based on a hierarchical decision tree that enables automatic movement detection and classification. Mathie et al. [29] further presented a generic classification framework consisting of a hierarchical binary tree for classifying postural transitions, falling, walking, and other movements using signals from a wearable triaxial accelerometer. This modular framework also allows modifying individual classification algorithm for particular purposes. Tilt sensing is a basic function provided by accelerometers which respond to gravity or constant acceleration. Therefore, human postures, such as upright and lying, can be distinguished according to the magnitude of acceleration signals along sensitive axes from only one accelerometer worn at the waist and torso [16,17]. However, the single-accelerometer approach has difficulty in distinguishing between standing and sitting as both are upright postures, although a simplified scheme with tilt threshold to distinguish standing and sitting has been proposed [16]. Standing and sitting postures can be distinguished by observing different orientations of body segments where multiple accelerometers are attached. For example, two accelerometers can be attached to the torso and thigh to distinguish standing and sitting postures from static activities [30-32]. Trunk tilt variation due to sit-stand postural transitions can be measured by integrating the signal from a gyroscope attached to the chest of the

Sensors 2010 7777 subject [33]. Sit-stand postural transitions can be identified according to the patterns of vertical acceleration from an accelerometer worn at the waist [17]. Acceleration signals can be used to determine walking in ambulatory movement. Walking can be identified by frequency-domain analysis [16,34]. It is characterized by a variance of over 0.02 g in vertical acceleration and frequency peak within 1–3 Hz in the signal spectrum [34]. Discrete wavelet transform is used to distinguish walking on a level ground and walking on a stairway [18]. Movement classification using statistical schemes utilize a supervised machine learning procedure, which associates an observation (or features) of movement to possible movement states in terms of the probability of the observation. Those schemes include, for example, k-nearest neighbor (kNN) classification [31,35], support vector machines (SVM) [36,37], Naive Bayes classifier [38,39], Gaussian mixture model (GMM) [40] and hidden Markov model (HMM) [41,42]. Naive Bayes classifier determines activities according to the probabilities of the signal pattern of the activities. In GMM approach, the likelihood function is not a typical Gaussian distribution. The weights and parameters describing probability of activities are obtained by the expectation-maximization algorithm. Transitions between activities can be described as a Markov chain that represents the likelihood (probability) of transitions between possible activities (states). The HMM is applied to determine unknown states at any time according to observable activity features (extracted from accelerometry data) corresponding to the states. After the HMM is trained by example data, it can be used to determine possible activity state transitions. 3.2. Estimation of Energy Expenditure Energy expenditure (EE) can be estimated by measuring physical activities. The doubly labeled water method (DLW) and indirect calorimetry that measures oxygen uptake, carbon dioxide production and cardiopulmonary parameters are regarded as the gold-standard references of EE. Though accurate, gas analyzers for indirect calorimetry are expensive and they require specialized skills to operate. The isotopes analysis and production for DLW method are costly and are not suitable for large-scale studies [43]. Accelerometers provide an alternative method of estimating energy expenditure in a free-living environment. EE due to physical activity can be better

the gold-standard references of EE. Though accurate, gas analyzers for indirect calorimetry are expensive and they require specialized skills to operate. The isotopes analysis and production for DLW method are costly and are not suitable for large-scale studies [43]. Accelerometers provide an alternative method of estimating energy expenditure in a free-living environment. EE due to physical activity can be better predicted from the acceleration integral in anterior-posterior direction of an accelerometer [44], though vertical acceleration is most sensitive to major activities like walking or running. The signal integral of triaxial acceleration outputs has been found to have linear relationship with the metabolic energy expenditure due to several daily activities [45]. Commercial accelerometers usually convert the magnitude of accelerations to provide ―activity counts‖ per defined period of time (epoch). The activity counts represent the estimated intensity of measured activities during each time period. Therefore, the recorded activity counts can be compared with questionnaires, or more accurately, the DLW method [46] or indirect calorimetry to estimate the energy expenditure due to activities [47]. Several regression equations can be derived or validated for different accelerometers to better match exact EE of physical activities among subjects. Factors affecting the accuracy of EE estimation using accelerometry are the location and attachment of accelerometers, external vibration, gravitational artifact, and the types of activity performed in a free-living environment. Sensor attachment to trunk, lower back or second lumbar vertebra is preferred because the trunk represents the major part of body mass and moves with most activities. Gravitational

Sensors 2010 7778 effect is also relatively small on this body segment [45]. On the other hand, waist-mounted accelerometers are unable to measure upper limb movement and have inaccurate EE estimation when the subjects carry different loads of weight during activity [29]. Moreover, EE during walking may be inaccurately estimated when the locomotion is not horizontal, e.g., slope climbing and walking up and downstairs. A barometer that measures the atmosphere pressure was integrated with a triaxial accelerometer [34]. This approach can use the added information of altitude changes to determine movement with vertical displacement, such as taking elevator, walking upstairs and downstairs. 3.3. Fall Detection and Balance Control Evaluation Fall-related injuries cause fracture and trauma which remarkably deteriorate the health and functional status of elderly people, leading to living dependence and higher risk of morbidity and mortality. Falls can be conceptually deemed as a rapid postural change from upright to reclining position to ground, or some lower level not as a consequence of sustaining a violent blow, loss of consciousness, sudden onset of paralysis as in stroke or an epileptic seizure [48]. The first approach to fall detection using accelerometry is published by Williams et al. [49], and a fall detector was presented after a number of pilot studies [50]. In its design implementation, the fall detector consisted of two piezoelectric shock sensors to detect the impact and a mercury tilt switch to identify the orientation. A two-stage detection process which detects both impact (acceleration) and orientation was used to better eliminate false alarms. The two-stage detection process firstly screens if any impact greater than a certain threshold exists (the first stage). A fall emergency is registered after the first stage if the reclining posture remains unchanged (the wearer does not get up) for a specific period of time. This design implementation led to the product commercialization of the fall detector by Tunstall Group [http://www.tunstall.co.uk/]. Similar approaches have been incorporated into fall detection algorithms using a waist-mounted accelerometer [16,17]. Lindemann et al. [24] evaluated a fall detector that was fixed behind the ear. Two high-g (50 g) accelerometers were orthogonally

This paper reviews the development of wearable accelerometry-based motion detectors.