- Email: [email protected]
Detection of physical activities using a physical activity monitor system for wheelchair users
Medical Engineering & Physics 37 (2015) 68–76
Contents lists available at ScienceDirect
Medical Engineering & Physics journal homepage: www.elsevier.com/locate/medengphy
Detection of physical activities using a physical activity monitor system for wheelchair users Shivayogi V. Hiremath a,c , Stephen S. Intille d,e , Annmarie Kelleher a,b , Rory A. Cooper a,b,f , Dan Ding a,b,f,∗ a
Department of Veterans Affairs, Human Engineering Research Laboratories, VA Pittsburgh Healthcare System, Pittsburgh, PA, United States Department of Rehabilitation Science and Technology, University of Pittsburgh, Pittsburgh, PA, United States c Department of Physical Medicine and Rehabilitation, University of Pittsburgh, Pittsburgh, PA, United States d College of Computer and Information Science, Northeastern University, Boston, MA, United States e Department of Health Sciences, Northeastern University, Boston, MA, United States f Department of Bioengineering, University of Pittsburgh, Pittsburgh, PA, United States b
a r t i c l e
i n f o
Article history: Received 15 March 2014 Received in revised form 12 September 2014 Accepted 18 October 2014 Keywords: Spinal cord injury Physical activity Wheelchair users Activity monitor system Smartphones Machine learning Classification
a b s t r a c t Availability of physical activity monitors for wheelchair users can potentially assist these individuals to track regular physical activity (PA), which in turn could lead to a healthier and more active lifestyle. Therefore, the aim of this study was to develop and validate algorithms for a physical activity monitoring system (PAMS) to detect wheelchair based activities. The PAMS consists of a gyroscope based wheel rotation monitor (G-WRM) and an accelerometer device (wocket) worn on the upper arm or on the wrist. A total of 45 persons with spinal cord injury took part in the study, which was performed in a structured university-based laboratory environment, a semi-structured environment at the National Veterans Wheelchair Games, and in the participants’ home environments. Participants performed at least ten PAs, other than resting, taken from a list of PAs. The classification performance for the best classifiers on the testing dataset for PAMS-Arm (G-WRM and wocket on upper arm) and PAMS-Wrist (G-WRM and wocket on wrist) was 89.26% and 88.47%, respectively. The outcomes of this study indicate that multi-modal information from the PAMS can help detect various types of wheelchair-based activities in structured laboratory, semi-structured organizational, and unstructured home environments. © 2014 IPEM. Published by Elsevier Ltd. All rights reserved.
1. Introduction Regular physical activity (PA) levels among persons with disabilities, for which 46% rated as performing some form of leisure-time PA in 2008, are significantly lower than the PA levels of the general population, for which 68% rated as performing some form of leisure-time PA [1]. Moreover, the obesity rate in persons with disabilities was 36% (2008); a rate much higher than the 23% in persons without disabilities [2]. Among those with disabilities are wheelchair users who lack regular PA and have reduced energy expenditure leading to even higher obesity and overweight levels
∗ Corresponding author at: University of Pittsburgh, Dept. of Rehabilitation Science and Technology, Human Engineering Research Laboratories, VA Pittsburgh Healthcare System 6425 Penn Avenue, Suite 400, Pittsburgh, PA 15206, United States. Tel.: +1 412 822 3684; fax: +1 412 822 3699. E-mail address: [email protected] (D. Ding). http://dx.doi.org/10.1016/j.medengphy.2014.10.009 1350-4533/© 2014 IPEM. Published by Elsevier Ltd. All rights reserved.
[3,4]. To address the need of achieving regular PA for wheelchair users, we have developed a physical activity monitoring system that can track regular PA levels and detect wheelchair-based activities. Research has evaluated the performance of various types of sensor-based activity monitors among persons who use wheelchairs to track movement to detect PAs [5–11]. Accelerometer-based activity monitors have been used to evaluate community living and wheelchair movement [5,6,9]. Warms et al. found that the activity counts from a wrist-worn accelerometer had low to moderate correlation (0.30–0.77, p < 0.01) with selfreported activity intensity for individual participants [5]. Coulter et al. investigated a wheel-mounted tri-axial accelerometer and found high validity of the device in detecting wheel revolutions, absolute angle and duration of movement (ICC(2,1) > 0.99, 0.99, 0.98, respectively) in wheelchair users [9]. Similarly, Sonenblum et al. used a wheel-mounted tri-axial accelerometer to detect wheelchair movement, and this device measured the
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
distance travelled with an accuracy greater than 90% for various wheelchair and wheel types, propulsion techniques, speeds, and wheelchair-related activities of daily living [6]. Tolerico et al. used another type of monitor, based on reed switches and a magnet, to find that manual wheelchair users travelled for a mean (SD) distance of 6745.3 (1937.9) m/day at a speed of 0.96 (0.17) m/s and 2457.0 (1195.7) m/day at a speed of 0.79 (0.19) m/s at the National Veterans Wheelchair Games and in the community, respectively [7]. Some of the limitations of current devices are that a single accelerometer on the wrist or a single wheel monitor cannot recognize manual wheelchair movements and upper extremity movements, respectively. Moreover, wheel monitoring devices alone cannot distinguish between selfpropulsion and external pushing. Postma et al. used a six-accelerometer based activity monitoring system and detected wheelchair propulsion from a series of activities of daily living with an overall agreement of 92%, a sensitivity of 87% and a specificity of 92% [11]. Hiremath et al. evaluated a multi-sensor based activity monitor (SenseWear, BodyMedia Inc., USA) to detect four activities: resting, wheelchair propulsion, arm-ergometry and deskwork [10]. The classification accuracy for detecting four wheelchair-related PAs was 96.3% using quadratic discriminant analysis and 94.8% using Naïve Bayes algorithms. Unfortunately, consumers cannot use any of these activity monitors to obtain real-time feedback about their mobility characteristics, as the information is usually post-processed based on the data stored in the devices. Real-time feedback of the PA level is an actionable parameter available throughout the day and that can be utilized whenever the wheelchair user has time to perform PAs. Access to this information can motivate users to increase their PA levels while controlling their energy intake. Shuger et al. conducted a randomized controlled trial in 197 sedentary overweight or obese adults to evaluate whether electronic feedback about diet and PA was more effective for weight loss [12]. The study concluded that continuous self-monitoring using sensor based technology with real-time feedback may promote weight loss in sedentary overweight or obese adults. Most of the real-time feedback systems also provide a report of PA level at the end of the day and indicate if the user had met their regular PA levels. Based on our previous research, we developed a physical activity monitoring system (PAMS) that tracks PA levels and provides feedback through smartphones [10,13,14]. The PAMS consists of two components: a gyroscope-based wheel rotation monitor (G-WRM) for capturing wheelchair wheel movement, and an accelerometer device (wocket) worn either on the upper arm or wrist to track upper arm or wrist acceleration, respectively [13,15]. The primary aim of this study was to develop and validate algorithms for PAMS to detect wheelchair based activities. The secondary aim was to evaluate the performance of individual components in the PAMS (i.e., G-WRM, wocket on the upper arm, or wocket on the wrist) as compared to using the two components in the PAMS. 2. Methods The study was approved by the Institutional Review Board of the University of Pittsburgh, US Army Medical Research & Material Command’s Human Research Protection Office, and the VA Pittsburgh Healthcare System. The study was conducted at a university laboratory, at the National Veterans Wheelchair Games (NVWG) held in Richmond, VA, USA in 2012, and in the participants’ home environments. 2.1. Participants A total of 45 persons with spinal cord injury (SCI) took part in the study. Participants were included in the study if they were 18–65
69
years of age, used a manual wheelchair (>80% of their ambulation), and had a diagnosis of SCI. Participants were excluded from the study if they were unable to tolerate sitting for 3 h, had active pelvic or thigh wounds, had a history of cardiovascular disease, or were pregnant (based on self-report). 2.2. Procedures The first part of the study was performed by 45 manual wheelchair users with SCI in the laboratory (lab) environment (N = 25) or in the semi-structured convention center environment at the NVWG (N = 20). A portion of the population who took part in the lab also participated in the study for a second time in their home environments (N = 20). 2.3. Protocol in lab or NVWG 2.3.1. Pre-activity session Before testing, a researcher explained the purpose and overall procedure of the study to the participants. After signing an informed consent, participants filled in a questionnaire that included questions on demographics (e.g., gender, ethnicity, age, injury level, and time of injury), wheelchair information (e.g., brand and model), and health and physical activity history. 2.3.2. Activity session Participants were asked to perform at least ten physical activities (PAs), other than resting, from this list of PAs that involved different parts of the body and varying levels of intensity: (1) propelling their wheelchair on a tile surface at a self-selected medium and fast pace, (2) propelling on a medium pile carpet at a selfselected medium or slow pace, (3) propelling up and down a ramp (slope of 2.7◦ , length 12.19 m) at a self-selected pace, (4) being pushed in a wheelchair on a tile surface or a medium pile carpet or up and down a ramp, (5) playing wheelchair basketball, (6) folding laundry, (7) performing deskwork involving reading and using a computer, (8) playing darts, (9) using a resistance band (Theraband), and (10) exercising on an arm ergometer at a self-selected pace and resistance. The participants chose the ten activities that they felt safe to perform, thus reducing the risk of injury. The PAs were chosen to cover a range of activities, representative of everyday activities, in wheelchair users and feasible in each of the three environments. The resting trial involved collecting the baseline data for 6 min while the participants sat still in their wheelchairs. During testing, a G-WRM was secured to the participant’s wheelchair and two wockets were worn on the participant’s upper arm and wrist. First, participants received instructions on how to perform the wheelchair-based activities. When participants wished to try out a particular trial before performing it, they were asked to do so for 1 to 2 min prior to the actual trial. All participants used their own manual wheelchairs and performed each activity for a minimum of 6 min, with at least a 3-min break between activity trials. One of the investigators noted the start and stop time for each activity trial. The activities were recorded on video, serving as a reference for subsequent timing and independent classification of the activities performed. Each testing session lasted for about 3 h. 2.4. Protocol in home environment Participants were invited to do a follow-up session if they lived within 60 miles of the lab and were willing to use the PAMS while they performed 10 daily activities and a resting trial, similar to the lab testing, in their home environment. The follow-up session was scheduled within 6 months of their testing in the lab.
70
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
small Bluetooth-based wireless accelerometers that capture body motion using a tri-axial capacitive micro-machined accelerometer [15]. Both the G-WRM and wocket transmit the sensor information wirelessly to an Android-based smartphone, which then processes the data every minute by using classification algorithms, and provides the user with real-time feedback. The G-WRM and wocket data are synchronized with the same reference, i.e., smartphone. Tri-axial acceleration from the wocket was sampled at 40 Hz. All devices were calibrated prior to the participant testing. The GWRM was calibrated as per the procedures discussed in Hiremath et al. [13]. The tri-axial accelerometer in the wocket was calibrated by placing the device on a level surface and orienting the device in six different directions to obtain an offset and a scaling factor for all the axes. An Android-based cellphone was secured to the participant’s waist to collect data from the G-WRM and the two wockets. 2.6. Data preparation and feature extraction
Fig. 1. Investigator propelling on a tile while the device collected data.
2.4.1. Activity session Participants were asked to perform a minimum of 10 PAs for at least 6 min per activity, and they could choose from the list of PAs performed in the laboratory environment or add new PAs that they wanted to perform in their home environments. The PAs performed in addition to the PAs mentioned included activities such as: (1) propelling in their home or community on a tile or carpet surface, (2) watching television, (3) simulating eating and cooking, (4) sweeping or vacuuming the floor, (5) making bed, (6) using dumbbells or handgrip, (7) washing dishes and laundry, and (8) performing wheelchair pushups. Each testing session lasted about 2 h. 2.5. Instrumentation and data collection The PAMS consisted of a G-WRM secured to the spokes of the wheelchair wheel on the right side and a wocket worn on the participant’s right upper arm or wrist (Fig. 1). The G-WRM is a self-enclosed rechargeable Bluetooth-based wireless device that contains a two-axis gyroscope for measuring the angular velocity of the wheelchair wheel [13]. The G-WRM measured the angular velocity, which was then converted into wheelchair velocity and distance traveled based on the wheelchair wheel’s diameter. Angular velocity from the G-WRM was sampled at 64 samples per second (64 Hz) during the testing. The value of the angular velocity was then down-sampled to 1 Hz by averaging the sampled data to capture wheelchair velocity for various PA trials. The wockets are
The data collected from the G-WRM and wockets were visually inspected by using a graphical user interface (personal computer) in order to identify any sensor malfunction or erroneous data and to identify the starting times for each device. The next step was to extract a set of time domain and frequency domain features, which are statistical measures used to distinguish between various types of wheelchair related PAs. The time domain features, such as mean, mean absolute deviation, and peaks were simple to extract and can be used to classify activities that are considerably different. The frequency domain features, such as total power between a band of frequencies, energy, and entropy provide classification models the capability to differentiate among activities based on the key frequency of movement (wheelchair propulsion and arm-ergometry). However, we need higher computation capacity when using frequency domain features compared to time domain features. Therefore, in this study we developed models with and without frequency domain features. All features were extracted based on a 1-min window size using custom programs written in MATLAB (The Mathworks, Inc., Natick, MA, USA). Following the feature extraction phase, the data were separated into training and testing datasets to enable the development and validation of PA classification models, respectively. In this study we evaluated classification models using a 10-fold-cross validation (10-fold-CV) on 80% of the participants’ data to develop and train new models, and then tested the new model on the remaining 20% of the participants’ data (80–20 CV). The 10-fold-CV is a statistical machine learning technique used as part of the development process to select the variables and optimize the model. The process involves developing a model based on nine folds of data and validating the developed model on the tenth fold of data not used for modeling. The 10-fold-CV is repeated ten times to obtain the final error, which is an average of all these repetitions. The 10-fold cross validation was used to generate the models iteratively, by selecting and adding features, to arrive at a final set of variables that had the least classification error for the 80% of the data used in training. Then once we developed and saved the final model on the training dataset we tested the final model on the testing dataset, which was not used in the building the model. The training error is presented as the 10-fold CV error for the final set of variables. The validation error was calculated only once based on the performance of the final model on the testing dataset. Training and testing datasets for the 80–20 CV were prepared using a stratified approach with injury level (paraplegia vs. tetraplegia) and gender (male vs. female) in order to allocate 80% of the participants into the training dataset and 20% into the testing dataset. Additionally we tried to make sure during the stratification of the training and testing dataset that the data was not continuous but
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
was randomly chosen from different venues and sessions. The data preparation steps resulted in a total dataset, including the training and testing datasets, of 3826 min (63 h and 46 min) from all devices. The total dataset consisted of 1555 min, 1001 min, and 1270 min of data for the lab, NVWG and Home environments, respectively.
performance of best algorithms for activity trials for individual devices or combined systems (Eq. (5)). Precision =
Recall = 2.7. Development of PA classification models The classification of wheelchair-based PAs was broken down into a two-step process as shown in Fig. 2. First, the data were classified into three classes including near-stationary PAs (distance travelled < 1.8 m/min), PAs that might involve wheelchair movement (1.8 m/min≥ distance travelled ≤12 m/min), and PAs with consistent wheelchair movement (distance travelled >12 m/min). We chose these thresholds based on the distance travelled by the participants for various PAs in this study. For this study, the PAs that might involve wheelchair movement included eating, sweeping the floor, preparing food, making the bed, cleaning the room, filing papers, playing darts, checking mail, doing laundry, and cleaning the car. The near-stationary PAs and PAs with consistent wheelchair movement were further classified into: (a) resting, arm-ergometry, and other household activities; (b) wheelchair propulsion, caretaker pushing and basketball, respectively. The other household activities in the near near-stationary PAs category included deskwork, folding clothes, using a resistance band, playing video games, and doing wheelchair pushups. This process of grouping the PAs from the total dataset into seven activities cumulated to 356 min of resting, 470 min of arm-ergometry, 1260 min of other household activities, 859 min of wheelchair propulsion, 330 min of caretaker pushing, 106 min of basketball, and 445 min of PAs that might involve wheelchair movement. We developed and evaluated various classification algorithms, including naïve bayes (NB), decision tress (DT) and support vector machines (SVM) [16] for individual devices including G-WRM, wocket on the upper arm, wocket on the wrist, and combined systems including PAMS-Arm (G-WRM and wocket on the upper arm) and PAMS-Wrist (G-WRM and wocket on the wrist). The classification models were developed using the training data (80–20 CV) to select the most appropriate features and evaluate the algorithms’ performance. As discussed previously, the G-WRM alone cannot distinguish between PAs such as wheelchair propulsion or caretaker pushing [13]. But PAMS-Arm and PAMS-Wrist can combine complementary information from the wheelchair, using G-WRM, and the upper arm or wrist movement, using wocket, to better detect wheelchair based PAs. In addition, because a wrist-wocket captures substantially more movement than the arm-wocket for light non-intensive activities such as deskwork or household activities, we performed PA detection using PAMSWrist to determine whether the wrist-wocket was a better predictor of wheelchair based PAs compared to the upper arm wocket. 2.8. Data analysis The performance of the activity classification algorithms was evaluated using performance measures including: (1) precision that indicates the proportion with which the activity is detected correctly (Eq. (1)), (2) recall (or sensitivity) that indicates the proportion of actual activities that are correctly identified (Eq. (2)), (3) specificity (or true negative rate) that indicates the proportion of activities not performed that are correctly identified (Eq. (3)), (4) accuracy that indicates the proportion of true positives and true negatives with respect to the total cases for each activity trial (Eq. (4)), and (5) overall accuracy that indicates the overall
71
true positive true positive + false positive
true positive true positive + false negative
Specificity =
Accuracy =
true negative true negative + false positive
true positive + true negative number of cases
Overall accuracy =
correctly classified cases total number of cases
(1)
(2)
(3)
(4)
(5)
3. Results Of the 45 participants, 39 were male and 6 were female with a mean (SD) age of 41.0 (12.6) years. The injury level of the participants varied from cervical level 5 to lumbar level 5, with 14 participants having injuries at or above thoracic level 3, and 31 participants having injuries at or below thoracic level 4. Twentythree of the 45 participants had an incomplete injury. The average number of years participants had used a manual wheelchair was 12.6 (8.6) years. Self-reported PA indicated that 36 participants performed some form of regular PA; 5 performed occasional PA; and 4 performed no regular PA. All 45 participants completed the study. Out of this total number, 37 finished 10 trials, 7 finished 9 trials, and 1 finished 8 trials. Most of these activity trials (701 out of 707 trials) were performed for 6 min per trial as requested, while 6 trials took between 2–4 min. Due to device malfunction of the G-WRM and wockets, as a result of a Bluetooth blockage or battery discharge, data from 24 trials (out of the 707) were lost across a range of activities. The PAMS-Arm, PAMS-Wrist, and G-WRM were able to classify the first phase of wheelchair-based PAs into near-stationary PAs, PAs that might involve wheelchair movement, and PAs with consistent wheelchair movement with an accuracy higher than 93% as they had the wheelchair wheel movement information captured by G-WRM (Table 1). The PAMS-Arm and PAMS-Wrist had reasonably high accuracy (85% and higher) for classifying the sub-activities of near-stationary PAs and PAs with consistent wheelchair movement (Table 2). In addition, the G-WRM was able to classify propulsion, basketball, and caretaker pushing with an accuracy higher than 88% (Table 2). As an example, Fig. 3 shows how the classifier model selected an acceleration-based feature (standard deviation in y direction of arm accelerometer) to detect wheelchair propulsion, caretaker pushing, and basketball activities that may have a different amount of arm movement during the activities. Classification performance analysis of the best classifiers indicated that PAMS-Arm and PAMS-Wrist had reasonably high overall classification accuracy (89.26% vs. 88.47%) with high precision and recall for most activities (Table 3). Table 4 shows the confusion matrix for PAMS-Arm, where the best classification algorithms (from Table 1) were used to classify wheelchair based PAs. The confusion matrix analysis for PAMS-Arm indicated that most of the PAs were correctly predicted (diagonal elements). In addition, the PAMS-Arm had reasonably high accuracy (85% and higher) for classifying the sub-activities of near-stationary PAs and PAs with consistent wheelchair movement using non-frequency domainbased features (Table 5).
72
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
Fig. 2. A two-step process for classifying various wheelchair-based PAs.
Table 1 Performance of various algorithms for detecting and classifying the first phase of wheelchair based PAs into near-stationary PAs, PAs that might involve wheelchair movement, and PAs with consistent wheelchair movement. Training accuracy in %
Testing accuracy in %
Features
Model
PAMS-arm
93.56
98.01
SVM
PAMS-wrist
93.56
98.01
G-WRM
93.56
98.01
Arm wocket
78.70
78.73
Wrist wocket
80.74
79.32
Root mean square of velocity from G-WRM, mean cross rate of velocity from G-WRM Root mean square of velocity from G-WRM, mean cross rate of velocity from G-WRM Root mean square of velocity from G-WRM, mean cross rate of velocity from G-WRM Ratio of dominant frequency’s power with total power in resultant acceleration from arm wocket, mean absolute deviation with respect to median in resultant acceleration from arm wocket Mean of acceleration in x axis from wrist wocket, entropy without the DC component in resultant acceleration from wrist wocket
SVM SVM DT
DT
Abbreviations used in this table: physical activities (PAs); physical activity monitoring system (PAMS); gyroscope based wheel rotation monitor (G-WRM); accelerometer device (wocket); G-WRM and wocket on upper arm (PAMS-arm); G-WRM and wocket on wrist (PAMS-wrist); support vector machines (SVM); decision tress (DT); mean value of the waveform (DC component).
Fig. 3. Plot of the acceleration feature chosen by the classification algorithm for PAMS-arm to classify and detect propulsion, caretaker pushing, and basketball activities.
4. Discussion The classification algorithms used in the first step to detect near-stationary PAs, PAs that might involve wheelchair movement, and PAs with consistent wheelchair movement indicated
that the G-WRM was a better predictor compared to the arm or wrist wockets, as G-WRM accurately detected the presence of wheelchair movement. The key features involved in the GWRM based classification algorithm were the root mean square and mean cross rate of velocity, both related to wheelchair movement. One possible reason for obtaining higher testing accuracy than the training accuracy (98.01% vs. 93.56%) for the G-WRM based classification algorithm is the lesser variation in the testing dataset due to the reduced number of participants (Table 1). In addition, the error for the testing dataset is just a single outcome compared to the error for the training dataset which was based on the 10-fold CV. During the second phase, of classifying the sub-activities for near-stationary PAs and PAs with consistent wheelchair movement, the results indicated that the arm (Fig. 3) and wrist wockets were better predictors of PAs than the G-WRM. From the features identified by the classification algorithms we see that most of them were frequency-based features, as the PAs involved repetitive use of upper arm movements. The reasonably high classification accuracy (92% for training and 88% for testing) of the G-WRM for moving PAs was probably due to G-WRM’s ability to distinguish caretaker pushing, which was performed by the same investigator, from propulsion and basketball. This pushing of wheelchair users by one individual may have resulted in a specific type of speed or acceleration pattern that the G-WRM was sensitive to.
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
73
Table 2 Performance of various algorithms for detecting and classifying the sub-activities of near-stationary PAs and PAs with consistent wheelchair movement (moving). Nearstationary PAs consist of resting, arm-ergometry, and other household activities. PAs with consistent wheelchair movement consist of propulsion, caretaker pushing, and basketball activities. Device/system
Near-stationary/ moving
Training accuracy in %
Testing accuracy in %
Features
Model
PAMS-arm
Near-stationary
86.78
84.95
DT
PAMS-arm
Moving
95.57
93.40
PAMS-wrist
Near-stationary
90.32
86.38
PAMS-wrist
Moving
97.13
93.87
G-WRM
Near-stationary
61.97
50.18
G-WRM
Moving
92.06
87.73
Arm wocket
Near-stationary
86.78
84.95
Arm wocket
Moving
96.12
91.51
Wrist wocket
Near-stationary
90.32
86.38
Wrist wocket
Moving
97.13
93.87
2nd dominant frequency content’s power in resultant acceleration from arm wocket, standard deviation of acceleration in z axis from arm wocket, amplitude of resultant acceleration from arm wocket Standard deviation of acceleration in y axis from arm wocket, amplitude of acceleration in y axis from arm wocket, entropy content of velocity from G-WRM 3rd dominant frequency content’s power in resultant acceleration from wrist wocket, Mean of acceleration in x axis from wrist wocket, dominant frequency content’s power in resultant acceleration from wrist wocket Ratio of dominant frequency’s power with total power in resultant acceleration from wrist wocket, mean of acceleration in x axis from wrist wocket, 3rd dominant frequency content’s power in resultant acceleration from wrist wocket Root mean square of velocity from G-WRM, zero cross rate of velocity from G-WRM, entropy of velocity from G-WRM 3rd dominant frequency content’s power in velocity from G-WRM, mean cross rate of velocity from G-WRM, entropy of velocity from G-WRM 2nd dominant frequency content’s power in resultant axis from arm wocket, standard deviation of acceleration in z axis from arm-wocket, amplitude in resultant axis from arm wocket Ratio of dominant frequency’s power with total power in y axis acceleration from arm wocket, 3rd dominant frequency content’s power in y axis from wrist wocket, root mean square of x axis acceleration from arm wocket 3rd dominant frequency content’s power in resultant acceleration from wrist wocket, mean of acceleration in x axis from wrist wocket, dominant frequency content’s power for resultant acceleration from wrist wocket Ratio of dominant frequency’s power with total power in resultant acceleration from wrist wocket, mean of acceleration in x axis from wrist wocket, 3rd dominant frequency content’s power in resultant acceleration from wrist wocket
SVM
DT
DT
NB DT
DT
DT
DT
DT
Abbreviations used in this table: physical activities (PAs); physical activity monitoring system (PAMS); gyroscope based wheel rotation monitor (G-WRM); accelerometer device (wocket); G-WRM and wocket on upper arm (PAMS-arm); G-WRM and wocket on wrist (PAMS-wrist); support vector machines (SVM); decision tress (DT); Naïve Bayes (NB).
An in-depth classification performance analysis (Table 3) of the best classifiers indicated that PAMS-Arm and PAMS-Wrist had similar overall classification accuracy (89.26% vs. 88.47%), precision (lower precision for resting and may be moving) and recall (lower recall for resting and basketball). The analysis revealed that multimodal sensor information from G-WRM and wocket for PAMS-Arm and PAMS-Wrist had higher accuracy than the individual devices (G-WRM and wockets). The confusion matrix analysis (Table 4) for PAMS indicated that most of the PAs were correctly predicted (diagonal elements). Further, the analysis indicated that misclassified PAs (non-diagonal) were classified into biomechanically similar PAs; for example, resting was classified as other household PAs in stationary category, and basketball as PA that may involve wheelchair movement because of the intermittent wheelchair movement. Even though misclassification resulted in lower classification accuracies, grouping activities into similar type of wheelchair-based activities can lead to energy expenditure or metabolic equivalent of task values similar to the values obtained for correctly identified activities. It is also evident that the wockets on arm or wrist have higher accuracy than the G-WRM because the arm plays a major role in all the type of PAs studied here. In addition, the overall classification performance of arm and wrist wockets (70.38% vs. 74.55%) was almost identical, with the wrist wocket having a slightly higher accuracy. These results, along with the results from the PAMS-Arm and PAMS-wrist above, indicate
that the users have the option of wearing the wocket on their arm or wrist based on their comfort or preference. However, the users will have to provide this detail to the Android based application prior to using PAMS so that appropriate classification models can be chosen to detect wheelchair based PAs. Since many of the classifier algorithms used frequency-based features to detect and classify PAs, we also evaluated the classification performance for non-frequency domain based features, which are computationally less expensive to generate. The analysis (Tables 2 and 5) indicated that the classification accuracy for PAMSArm was similar for near-stationary PAs (0.87 vs. 0.87 for training and 0.85 vs. 0.81 for testing) and PAs with consistent wheelchair movement (0.96 vs. 0.95 for training and 0.93 vs. 0.94 for testing). Comparing the features chosen by the classifier algorithms for frequency and non-frequency based features indicated that the classifier algorithms picked slightly different features with a similar type of movement information, such as rate vs. frequency, and change in distance vs. entropy. For example, the classifier chose the mean cross rate of acceleration in resultant direction using non-frequency based feature instead of the power of second dominant frequency for non-moving PAs; and the mean velocity of the G-WRM was chosen instead of the entropy of the velocity of the G-WRM feature for moving PAs. As the use of frequency based features did not seem to improve the accuracy, we plan to use non-frequency based features when programing the smartphones.
74
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
Table 3 Classification performance in terms of precision, recall or sensitivity, specificity (or true negative rate), accuracy for each activity and overall accuracy for the best classifiers on the validation dataset for PAMS-arm, PAMS-wrist, G-WRM, arm wocket, and wrist wocket. Device/system
First level activity
Sub-activity trials
PAMS-arm
Near-stationary PAs
Resting Arm-ergometry Other household activities Propulsion Caretaker pushing Basketball May be moving Resting Arm-ergometry Other household activities Propulsion Caretaker pushing Basketball May be moving Resting Arm-ergometry Other household activities Propulsion Caretaker pushing Basketball May be moving Resting Arm-ergometry Other household activities Propulsion Caretaker pushing Basketball May be moving Resting Arm-ergometry Other household activities Propulsion Caretaker pushing Basketball May be moving
PAMS-wrist
G-WRM
Arm wocket
Wrist wocket
PAs with consistent wheelchair movement May be moving Near-stationary PAs PAs with consistent wheelchair movement May be moving Near-stationary PAs PAs with consistent wheelchair movement May be moving Near-stationary PAs PAs with consistent wheelchair movement May be moving Near-stationary PAs PAs with consistent wheelchair movement May be moving
Precision 0.88 0.98 0.81 0.95 0.96 1.00 0.57 0.74 0.91 0.87 0.94 0.94 0.94 0.57 N/A N/A 0.50 0.91 0.90 0.62 0.57 0.88 0.81 0.63 0.89 0.67 0.88 0.21 0.25 0.97 0.76 0.98 0.42 0.88 0.24
Recall
Specificity
Accuracy (%)
Overall accuracy (%)
0.47 0.92 0.96 0.99 0.94 0.57 1.00 0.68 0.93 0.88 0.97 0.96 0.53 1.00 0.00 0.00 1.00 0.90 0.79 0.60 1.00 0.47 0.89 0.79 0.65 0.55 0.50 0.83 0.02 0.93 0.84 0.86 0.83 0.23 0.75
0.99 1.00 0.91 0.98 1.00 1.00 0.98 0.98 0.98 0.95 0.98 0.99 1.00 0.98 1.00 1.00 0.62 0.97 0.99 0.98 0.98 0.99 0.95 0.82 0.97 0.97 1.00 0.92 0.99 0.99 0.90 0.99 0.88 1.00 0.94
94.43 98.21 92.84 98.41 99.01 97.42 98.21 94.83 97.22 93.04 97.61 99.01 97.02 98.21 90.66 81.71 72.56 95.03 97.22 95.43 98.21 94.43 94.23 81.51 88.47 93.24 96.62 92.25 90.26 98.21 88.27 95.83 87.67 95.23 93.64
89.26
88.47
65.41
70.38
74.55
Abbreviations used in this table: TN physical activity monitoring system (PAMS); gyroscope based wheel rotation monitor (G-WRM); accelerometer device (wocket); G-WRM and wocket on upper arm (PAMS-arm); G-WRM and wocket on wrist (PAMS-wrist).
Table 4 Confusion matrix for PAMS-arm on validation dataset (20% of participants’ data not used for training) using the best algorithms to classify wheelchair based PAs. True\predicted
Resting
Arm-ergometry
Other household activities
Propulsion
Caretaker Pushing
Basketball
May be moving
Resting Arm-ergometry Other household activities Propulsion Caretaker pushing Basketball May be moving
22 0 3 0 0 0 0
0 85 2 0 0 0 0
24 6 135 0 0 1 0
0 0 0 134 2 5 0
0 0 0 1 44 1 0
0 0 0 0 0 17 0
1 1 0 0 1 6 12
Abbreviations used in this table: physical activities (PAs); physical activity monitoring system (PAMS); gyroscope based wheel rotation monitor (G-WRM); accelerometer device (wocket); G-WRM and wocket on upper arm (PAMS-arm).
Our algorithms’ overall classification accuracy for the detection of wheelchair based PAs (PAMS-Arm: 93.6% for training and 98.0% for testing), and the detection of PAs involving wheelchair movement (PAMS-Arm: 95.6% for training and 93.5% for testing) was
similar to Sonenblum et al.’s detection of wheelchair movement for various wheelchair-related activities of daily living (90–96%) [6]. This resemblance, in addition to our previous validation of GWRM [13], indicates that our system can measure wheelchair use in
Table 5 Performance of classification algorithms for PAMS-Arm using 80–20 CV to detect and classify near-stationary PAs and PAs with consistent wheelchair movement (moving) using non-frequency domain-based features. Near-stationary PAs consist of resting, arm-ergometry, and other household activities. PAs with consistent wheelchair movement consist of propulsion, caretaker pushing, and basketball activities. Near-stationary/moving
Training accuracy in %
Testing accuracy in %
Features
Model
Near-Stationary
85.61
86.02
SVM
Moving
95.84
93.87
Mean cross rate of resultant acceleration from arm-wocket, mean cross rate of acceleration in z axis from arm-wocket, Amplitude of acceleration in z axis from arm wocket Standard deviation of resultant acceleration from arm-wocket, Amplitude of acceleration in x axis from arm wocket, Mean of velocity from G-WRM
DT
Abbreviations used in this table: physical activities (PAs); physical activity monitoring system (PAMS); gyroscope based wheel rotation monitor (G-WRM); accelerometer device (wocket); G-WRM and wocket on upper arm (PAMS-arm); support vector machines (SVM); decision tress (DT).
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
activities of daily living. The advantage of our system is that it can track both wheelchair and arm movements towards tracking PA. Additionally, the classification performance results (overall accuracy: 96%, sensitivity: 99% and specificity: 98%) for detecting wheelchair propulsion using PAMS-Arm were slightly higher than the results of Postma et al. who detected wheelchair propulsion compared to other activities (overall agreement: 92%, sensitivity: 87% and specificity: 92%) [11]. The advantage of our system is the reduced number of devices: G-WRM on wheel and arm wocket instead of the six-accelerometers based activity monitoring system used by Postma et al. The overall classification accuracies for PAMS-Arm (89.3%) and PAMS-Wrist (88.5%) were lower compared to those in our previous study, during which we used the SenseWear activity monitor to detect resting, wheelchair propulsion, arm-ergometry and deskwork activities (96.3%) [10]. The lower classification accuracy was probably due to testing of a larger number of wheelchair-based PAs, performed at a self-chosen pace obtained in structured, semi-structured and unstructured natural environments. One of the limitations of this study was that a large percentage of our participants self-reported that they were physically active on a regular (N = 36) or occasional basis (N = 5). Thus, the PA levels reported here are significantly higher compared to those in Washburn et al., who indicated that out of their participants only 13–16% of persons with SCI reported consistent PA [17], and the majority reported virtually no regular PA [18]. However, our high PA levels might have been inflated due to self-reported PA levels, which tend to have a social acceptability bias. Another limitation of this study is the inclusion of a large cohort of veterans with SCI tested during the NVWG (N = 20). The majority of the veterans who took part in this study probably had a better standard of care regarding assistive technology; however, they still had current PA and health levels similar to the general population. Future studies should recruit a greater percentage of manual wheelchair users from the general population and identify additional wheelchair based activities to develop new classification models. Additionally, our study only recruited individuals with SCI to limit the influence of various disabilities on wheelchair related PAs. Recruiting individuals with other disabilities would enable us to develop classification algorithms to quantify PA levels among different groups of persons with similar disabilities. Another limitation of the study is that we were unable to evaluate the validity of PAMS in all three environments separately due to the limited dataset. Future studies should evaluate the validity of PAMS in various settings separately. Moreover, the classification models developed here were based on arm acceleration and wheelchair velocity, both movementbased variables. Future work should develop PAMS to incorporate other forms of physiologic sensing, such as galvanic skin response, skin temperature, near body temperature, and heart rate, in order to detect resistance-based PAs. We are in the process of evaluating and developing PAMS to estimate energy expenditure so that it can be reliably deployed into longitudinal testing of one to two weeks in the community. In future, we plan to study specific movement related features, which could assist researchers and clinicians with quantifying upper arm movement associated with carpal tunnel syndrome or overuse syndrome.
5. Conclusion The outcomes of this study indicate that multi-modal information obtained from PAMS can help detect various types of wheelchair based activities in structured laboratory, semistructured NVWG, and unstructured home environments with reasonably high accuracy. The study also indicated that the overall
75
classification performance of PAMS was higher than the individual components (G-WRM, wocket on the upper arm or wrist). Furthermore, technologies such as PAMS can provide real-time feedback about physical activity levels and can motivate individuals to achieve a healthy lifestyle. Ethical approval The study was approved by the Institutional Review Board of the University of Pittsburgh (PRO10060368), US Army Medical Research & Material Command’s Human Research Protection Office (HRPO log number: A-16284.1), and the VA Pittsburgh Healthcare System (MIRB/PRO#: 02967). Acknowledgements The work is supported by the Department of Defense (W81XWH-10-1-0816). SVH’s work on this article was funded through the Switzer Research Fellowship (H133F110032) awarded by the National Institute on Disability and Rehabilitation Research, Department of Education. The work is also supported by the Human Engineering Research Laboratories, VA Pittsburgh Healthcare System. The contents do not represent the views of the Department of Veterans Affairs or the United States Government. The authors thank their colleagues at the Human Engineering Research Laboratories, especially Alejandra Manoela Ojeda, Monsak Socharoentum, Natthasit Wongsirikul, and Matthew Hannan for their input and effort during development and data collection. Conflict of interest statement The authors have no conflict of interest. References [1] U.S. Department of Health and Human Services. Healthy People 2020. Washington, DC: U.S. Department of Health and Human Services; 2011. Available: http://www.healthypeople.gov/2020/topicsobjectives2020/overview.aspx? topicid=9 [2] Centers for Disease Control and Prevention. Overweight and Obesity: Among People with Disabilities. Centers for Disease Control and Prevention; 2010. [3] Rimmer JH, Yamaki K, Davis BM, Wang E, Vogel LC. Obesity and overweight prevalence among adolescents with disabilities. Prev Chronic Dis 2011;8:1–6. [4] Buchholz AC, Pencharz PB. Energy expenditure in chronic spinal cord injury. Clin Nutr Metab Care 2004;7:635–9. [5] Warms CA, Belza BL. Actigraphy as a measure of physical activity for wheelchair users with spinal cord injury. Nurs Res 2004;53:136–43. [6] Sonenblum SE, Sprigle S, Caspall J, Lopez R. Validation of an accelerometerbased method to measure the use of manual wheelchairs. Med Eng Phys 2012;34:781–6. [7] Tolerico ML, Ding D, Cooper RA, Spaeth DM, Fitzgerald SG, Cooper R, et al. Assessing mobility characteristics and activity levels of manual wheelchair users. J Rehabil Res Dev 2007;44:561–72. [8] Lee M, Zhu W, Hedrick B, Fernhall B. Estimating MET values using the ratio of HR for persons with paraplegia. Med Sci Sports Exerc 2010;42:985–90. [9] Coulter EH, Dall PM, Rochester L, Hasler JP, Granat MH. Development and validation of a physical activity monitor for use on a wheelchair. J Spinal Cord 2011;49:445–50. [10] Hiremath SV, Ding D, Farringdon J, Vyas N, Cooper RA. Physical activity classification utilizing SenseWear activity monitor in manual wheelchair users with spinal cord injury. Spinal Cord 2013;51:705–9. [11] Postma K, Berg-Emons van den HJG, Bussmann JBJ, Sluis TAR, Bergen MP, Stam HJ. Validity of the detection of wheelchair propulsion as measured with an activity monitor in patients with spinal cord injury. Spinal Cord 2005;43:550–7. [12] Shuger SL, Barry VW, Sui X, McClain A, Hand GA, Wilcox S, et al. Electronic feedback in a diet- and physical activity-based lifestyle intervention for weight loss: a randomized controlled trial. Int J Behav Nutr Phys Act 2011;8:41:1–8. [13] Hiremath SV, Ding D, Cooper RA. Development and evaluation of a gyroscope based wheel rotation monitor for manual wheelchair users. Spinal Cord Med 2013;36:347–56. [14] Hiremath SV, Ding D, Farringdon J, Cooper RA. Predicting energy expenditure of manual wheelchair users with spinal cord injury using a multi-sensor based activity monitor. Arch Phys Med Rehabil 2012;93:1937–43.
76
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
[15] Intille SS, Albinali F, Mota S, Kuris B, Botana P, Haskell WL. Design of a wearable physical activity monitoring system using mobile phones and accelerometers. In: Proceedings of Annual International Conference of the IEEE Engineering in Medicine and Biology Society. 2011. p. 3636–9. [16] Witten IH, Frank E. Data mining practical machine learning tools and techniques. 2nd ed. San Francisco, CA: Morgan Kaufmann; 2005.
[17] Washburn R, Hedrick BN. Descriptive epidemiology of physical activity in university graduates with locomotor disabilities. Int J Rehabil Res 1997;20: 275–87. [18] Fernhall B, Heffernan K, Jae SY, Hedrick B. Health implications of physical activity in individuals with spinal cord injury: a literature review. J Health Hum Serv Adm 2008;30:468–502.
Contents lists available at ScienceDirect
Medical Engineering & Physics journal homepage: www.elsevier.com/locate/medengphy
Detection of physical activities using a physical activity monitor system for wheelchair users Shivayogi V. Hiremath a,c , Stephen S. Intille d,e , Annmarie Kelleher a,b , Rory A. Cooper a,b,f , Dan Ding a,b,f,∗ a
Department of Veterans Affairs, Human Engineering Research Laboratories, VA Pittsburgh Healthcare System, Pittsburgh, PA, United States Department of Rehabilitation Science and Technology, University of Pittsburgh, Pittsburgh, PA, United States c Department of Physical Medicine and Rehabilitation, University of Pittsburgh, Pittsburgh, PA, United States d College of Computer and Information Science, Northeastern University, Boston, MA, United States e Department of Health Sciences, Northeastern University, Boston, MA, United States f Department of Bioengineering, University of Pittsburgh, Pittsburgh, PA, United States b
a r t i c l e
i n f o
Article history: Received 15 March 2014 Received in revised form 12 September 2014 Accepted 18 October 2014 Keywords: Spinal cord injury Physical activity Wheelchair users Activity monitor system Smartphones Machine learning Classification
a b s t r a c t Availability of physical activity monitors for wheelchair users can potentially assist these individuals to track regular physical activity (PA), which in turn could lead to a healthier and more active lifestyle. Therefore, the aim of this study was to develop and validate algorithms for a physical activity monitoring system (PAMS) to detect wheelchair based activities. The PAMS consists of a gyroscope based wheel rotation monitor (G-WRM) and an accelerometer device (wocket) worn on the upper arm or on the wrist. A total of 45 persons with spinal cord injury took part in the study, which was performed in a structured university-based laboratory environment, a semi-structured environment at the National Veterans Wheelchair Games, and in the participants’ home environments. Participants performed at least ten PAs, other than resting, taken from a list of PAs. The classification performance for the best classifiers on the testing dataset for PAMS-Arm (G-WRM and wocket on upper arm) and PAMS-Wrist (G-WRM and wocket on wrist) was 89.26% and 88.47%, respectively. The outcomes of this study indicate that multi-modal information from the PAMS can help detect various types of wheelchair-based activities in structured laboratory, semi-structured organizational, and unstructured home environments. © 2014 IPEM. Published by Elsevier Ltd. All rights reserved.
1. Introduction Regular physical activity (PA) levels among persons with disabilities, for which 46% rated as performing some form of leisure-time PA in 2008, are significantly lower than the PA levels of the general population, for which 68% rated as performing some form of leisure-time PA [1]. Moreover, the obesity rate in persons with disabilities was 36% (2008); a rate much higher than the 23% in persons without disabilities [2]. Among those with disabilities are wheelchair users who lack regular PA and have reduced energy expenditure leading to even higher obesity and overweight levels
∗ Corresponding author at: University of Pittsburgh, Dept. of Rehabilitation Science and Technology, Human Engineering Research Laboratories, VA Pittsburgh Healthcare System 6425 Penn Avenue, Suite 400, Pittsburgh, PA 15206, United States. Tel.: +1 412 822 3684; fax: +1 412 822 3699. E-mail address: [email protected] (D. Ding). http://dx.doi.org/10.1016/j.medengphy.2014.10.009 1350-4533/© 2014 IPEM. Published by Elsevier Ltd. All rights reserved.
[3,4]. To address the need of achieving regular PA for wheelchair users, we have developed a physical activity monitoring system that can track regular PA levels and detect wheelchair-based activities. Research has evaluated the performance of various types of sensor-based activity monitors among persons who use wheelchairs to track movement to detect PAs [5–11]. Accelerometer-based activity monitors have been used to evaluate community living and wheelchair movement [5,6,9]. Warms et al. found that the activity counts from a wrist-worn accelerometer had low to moderate correlation (0.30–0.77, p < 0.01) with selfreported activity intensity for individual participants [5]. Coulter et al. investigated a wheel-mounted tri-axial accelerometer and found high validity of the device in detecting wheel revolutions, absolute angle and duration of movement (ICC(2,1) > 0.99, 0.99, 0.98, respectively) in wheelchair users [9]. Similarly, Sonenblum et al. used a wheel-mounted tri-axial accelerometer to detect wheelchair movement, and this device measured the
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
distance travelled with an accuracy greater than 90% for various wheelchair and wheel types, propulsion techniques, speeds, and wheelchair-related activities of daily living [6]. Tolerico et al. used another type of monitor, based on reed switches and a magnet, to find that manual wheelchair users travelled for a mean (SD) distance of 6745.3 (1937.9) m/day at a speed of 0.96 (0.17) m/s and 2457.0 (1195.7) m/day at a speed of 0.79 (0.19) m/s at the National Veterans Wheelchair Games and in the community, respectively [7]. Some of the limitations of current devices are that a single accelerometer on the wrist or a single wheel monitor cannot recognize manual wheelchair movements and upper extremity movements, respectively. Moreover, wheel monitoring devices alone cannot distinguish between selfpropulsion and external pushing. Postma et al. used a six-accelerometer based activity monitoring system and detected wheelchair propulsion from a series of activities of daily living with an overall agreement of 92%, a sensitivity of 87% and a specificity of 92% [11]. Hiremath et al. evaluated a multi-sensor based activity monitor (SenseWear, BodyMedia Inc., USA) to detect four activities: resting, wheelchair propulsion, arm-ergometry and deskwork [10]. The classification accuracy for detecting four wheelchair-related PAs was 96.3% using quadratic discriminant analysis and 94.8% using Naïve Bayes algorithms. Unfortunately, consumers cannot use any of these activity monitors to obtain real-time feedback about their mobility characteristics, as the information is usually post-processed based on the data stored in the devices. Real-time feedback of the PA level is an actionable parameter available throughout the day and that can be utilized whenever the wheelchair user has time to perform PAs. Access to this information can motivate users to increase their PA levels while controlling their energy intake. Shuger et al. conducted a randomized controlled trial in 197 sedentary overweight or obese adults to evaluate whether electronic feedback about diet and PA was more effective for weight loss [12]. The study concluded that continuous self-monitoring using sensor based technology with real-time feedback may promote weight loss in sedentary overweight or obese adults. Most of the real-time feedback systems also provide a report of PA level at the end of the day and indicate if the user had met their regular PA levels. Based on our previous research, we developed a physical activity monitoring system (PAMS) that tracks PA levels and provides feedback through smartphones [10,13,14]. The PAMS consists of two components: a gyroscope-based wheel rotation monitor (G-WRM) for capturing wheelchair wheel movement, and an accelerometer device (wocket) worn either on the upper arm or wrist to track upper arm or wrist acceleration, respectively [13,15]. The primary aim of this study was to develop and validate algorithms for PAMS to detect wheelchair based activities. The secondary aim was to evaluate the performance of individual components in the PAMS (i.e., G-WRM, wocket on the upper arm, or wocket on the wrist) as compared to using the two components in the PAMS. 2. Methods The study was approved by the Institutional Review Board of the University of Pittsburgh, US Army Medical Research & Material Command’s Human Research Protection Office, and the VA Pittsburgh Healthcare System. The study was conducted at a university laboratory, at the National Veterans Wheelchair Games (NVWG) held in Richmond, VA, USA in 2012, and in the participants’ home environments. 2.1. Participants A total of 45 persons with spinal cord injury (SCI) took part in the study. Participants were included in the study if they were 18–65
69
years of age, used a manual wheelchair (>80% of their ambulation), and had a diagnosis of SCI. Participants were excluded from the study if they were unable to tolerate sitting for 3 h, had active pelvic or thigh wounds, had a history of cardiovascular disease, or were pregnant (based on self-report). 2.2. Procedures The first part of the study was performed by 45 manual wheelchair users with SCI in the laboratory (lab) environment (N = 25) or in the semi-structured convention center environment at the NVWG (N = 20). A portion of the population who took part in the lab also participated in the study for a second time in their home environments (N = 20). 2.3. Protocol in lab or NVWG 2.3.1. Pre-activity session Before testing, a researcher explained the purpose and overall procedure of the study to the participants. After signing an informed consent, participants filled in a questionnaire that included questions on demographics (e.g., gender, ethnicity, age, injury level, and time of injury), wheelchair information (e.g., brand and model), and health and physical activity history. 2.3.2. Activity session Participants were asked to perform at least ten physical activities (PAs), other than resting, from this list of PAs that involved different parts of the body and varying levels of intensity: (1) propelling their wheelchair on a tile surface at a self-selected medium and fast pace, (2) propelling on a medium pile carpet at a selfselected medium or slow pace, (3) propelling up and down a ramp (slope of 2.7◦ , length 12.19 m) at a self-selected pace, (4) being pushed in a wheelchair on a tile surface or a medium pile carpet or up and down a ramp, (5) playing wheelchair basketball, (6) folding laundry, (7) performing deskwork involving reading and using a computer, (8) playing darts, (9) using a resistance band (Theraband), and (10) exercising on an arm ergometer at a self-selected pace and resistance. The participants chose the ten activities that they felt safe to perform, thus reducing the risk of injury. The PAs were chosen to cover a range of activities, representative of everyday activities, in wheelchair users and feasible in each of the three environments. The resting trial involved collecting the baseline data for 6 min while the participants sat still in their wheelchairs. During testing, a G-WRM was secured to the participant’s wheelchair and two wockets were worn on the participant’s upper arm and wrist. First, participants received instructions on how to perform the wheelchair-based activities. When participants wished to try out a particular trial before performing it, they were asked to do so for 1 to 2 min prior to the actual trial. All participants used their own manual wheelchairs and performed each activity for a minimum of 6 min, with at least a 3-min break between activity trials. One of the investigators noted the start and stop time for each activity trial. The activities were recorded on video, serving as a reference for subsequent timing and independent classification of the activities performed. Each testing session lasted for about 3 h. 2.4. Protocol in home environment Participants were invited to do a follow-up session if they lived within 60 miles of the lab and were willing to use the PAMS while they performed 10 daily activities and a resting trial, similar to the lab testing, in their home environment. The follow-up session was scheduled within 6 months of their testing in the lab.
70
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
small Bluetooth-based wireless accelerometers that capture body motion using a tri-axial capacitive micro-machined accelerometer [15]. Both the G-WRM and wocket transmit the sensor information wirelessly to an Android-based smartphone, which then processes the data every minute by using classification algorithms, and provides the user with real-time feedback. The G-WRM and wocket data are synchronized with the same reference, i.e., smartphone. Tri-axial acceleration from the wocket was sampled at 40 Hz. All devices were calibrated prior to the participant testing. The GWRM was calibrated as per the procedures discussed in Hiremath et al. [13]. The tri-axial accelerometer in the wocket was calibrated by placing the device on a level surface and orienting the device in six different directions to obtain an offset and a scaling factor for all the axes. An Android-based cellphone was secured to the participant’s waist to collect data from the G-WRM and the two wockets. 2.6. Data preparation and feature extraction
Fig. 1. Investigator propelling on a tile while the device collected data.
2.4.1. Activity session Participants were asked to perform a minimum of 10 PAs for at least 6 min per activity, and they could choose from the list of PAs performed in the laboratory environment or add new PAs that they wanted to perform in their home environments. The PAs performed in addition to the PAs mentioned included activities such as: (1) propelling in their home or community on a tile or carpet surface, (2) watching television, (3) simulating eating and cooking, (4) sweeping or vacuuming the floor, (5) making bed, (6) using dumbbells or handgrip, (7) washing dishes and laundry, and (8) performing wheelchair pushups. Each testing session lasted about 2 h. 2.5. Instrumentation and data collection The PAMS consisted of a G-WRM secured to the spokes of the wheelchair wheel on the right side and a wocket worn on the participant’s right upper arm or wrist (Fig. 1). The G-WRM is a self-enclosed rechargeable Bluetooth-based wireless device that contains a two-axis gyroscope for measuring the angular velocity of the wheelchair wheel [13]. The G-WRM measured the angular velocity, which was then converted into wheelchair velocity and distance traveled based on the wheelchair wheel’s diameter. Angular velocity from the G-WRM was sampled at 64 samples per second (64 Hz) during the testing. The value of the angular velocity was then down-sampled to 1 Hz by averaging the sampled data to capture wheelchair velocity for various PA trials. The wockets are
The data collected from the G-WRM and wockets were visually inspected by using a graphical user interface (personal computer) in order to identify any sensor malfunction or erroneous data and to identify the starting times for each device. The next step was to extract a set of time domain and frequency domain features, which are statistical measures used to distinguish between various types of wheelchair related PAs. The time domain features, such as mean, mean absolute deviation, and peaks were simple to extract and can be used to classify activities that are considerably different. The frequency domain features, such as total power between a band of frequencies, energy, and entropy provide classification models the capability to differentiate among activities based on the key frequency of movement (wheelchair propulsion and arm-ergometry). However, we need higher computation capacity when using frequency domain features compared to time domain features. Therefore, in this study we developed models with and without frequency domain features. All features were extracted based on a 1-min window size using custom programs written in MATLAB (The Mathworks, Inc., Natick, MA, USA). Following the feature extraction phase, the data were separated into training and testing datasets to enable the development and validation of PA classification models, respectively. In this study we evaluated classification models using a 10-fold-cross validation (10-fold-CV) on 80% of the participants’ data to develop and train new models, and then tested the new model on the remaining 20% of the participants’ data (80–20 CV). The 10-fold-CV is a statistical machine learning technique used as part of the development process to select the variables and optimize the model. The process involves developing a model based on nine folds of data and validating the developed model on the tenth fold of data not used for modeling. The 10-fold-CV is repeated ten times to obtain the final error, which is an average of all these repetitions. The 10-fold cross validation was used to generate the models iteratively, by selecting and adding features, to arrive at a final set of variables that had the least classification error for the 80% of the data used in training. Then once we developed and saved the final model on the training dataset we tested the final model on the testing dataset, which was not used in the building the model. The training error is presented as the 10-fold CV error for the final set of variables. The validation error was calculated only once based on the performance of the final model on the testing dataset. Training and testing datasets for the 80–20 CV were prepared using a stratified approach with injury level (paraplegia vs. tetraplegia) and gender (male vs. female) in order to allocate 80% of the participants into the training dataset and 20% into the testing dataset. Additionally we tried to make sure during the stratification of the training and testing dataset that the data was not continuous but
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
was randomly chosen from different venues and sessions. The data preparation steps resulted in a total dataset, including the training and testing datasets, of 3826 min (63 h and 46 min) from all devices. The total dataset consisted of 1555 min, 1001 min, and 1270 min of data for the lab, NVWG and Home environments, respectively.
performance of best algorithms for activity trials for individual devices or combined systems (Eq. (5)). Precision =
Recall = 2.7. Development of PA classification models The classification of wheelchair-based PAs was broken down into a two-step process as shown in Fig. 2. First, the data were classified into three classes including near-stationary PAs (distance travelled < 1.8 m/min), PAs that might involve wheelchair movement (1.8 m/min≥ distance travelled ≤12 m/min), and PAs with consistent wheelchair movement (distance travelled >12 m/min). We chose these thresholds based on the distance travelled by the participants for various PAs in this study. For this study, the PAs that might involve wheelchair movement included eating, sweeping the floor, preparing food, making the bed, cleaning the room, filing papers, playing darts, checking mail, doing laundry, and cleaning the car. The near-stationary PAs and PAs with consistent wheelchair movement were further classified into: (a) resting, arm-ergometry, and other household activities; (b) wheelchair propulsion, caretaker pushing and basketball, respectively. The other household activities in the near near-stationary PAs category included deskwork, folding clothes, using a resistance band, playing video games, and doing wheelchair pushups. This process of grouping the PAs from the total dataset into seven activities cumulated to 356 min of resting, 470 min of arm-ergometry, 1260 min of other household activities, 859 min of wheelchair propulsion, 330 min of caretaker pushing, 106 min of basketball, and 445 min of PAs that might involve wheelchair movement. We developed and evaluated various classification algorithms, including naïve bayes (NB), decision tress (DT) and support vector machines (SVM) [16] for individual devices including G-WRM, wocket on the upper arm, wocket on the wrist, and combined systems including PAMS-Arm (G-WRM and wocket on the upper arm) and PAMS-Wrist (G-WRM and wocket on the wrist). The classification models were developed using the training data (80–20 CV) to select the most appropriate features and evaluate the algorithms’ performance. As discussed previously, the G-WRM alone cannot distinguish between PAs such as wheelchair propulsion or caretaker pushing [13]. But PAMS-Arm and PAMS-Wrist can combine complementary information from the wheelchair, using G-WRM, and the upper arm or wrist movement, using wocket, to better detect wheelchair based PAs. In addition, because a wrist-wocket captures substantially more movement than the arm-wocket for light non-intensive activities such as deskwork or household activities, we performed PA detection using PAMSWrist to determine whether the wrist-wocket was a better predictor of wheelchair based PAs compared to the upper arm wocket. 2.8. Data analysis The performance of the activity classification algorithms was evaluated using performance measures including: (1) precision that indicates the proportion with which the activity is detected correctly (Eq. (1)), (2) recall (or sensitivity) that indicates the proportion of actual activities that are correctly identified (Eq. (2)), (3) specificity (or true negative rate) that indicates the proportion of activities not performed that are correctly identified (Eq. (3)), (4) accuracy that indicates the proportion of true positives and true negatives with respect to the total cases for each activity trial (Eq. (4)), and (5) overall accuracy that indicates the overall
71
true positive true positive + false positive
true positive true positive + false negative
Specificity =
Accuracy =
true negative true negative + false positive
true positive + true negative number of cases
Overall accuracy =
correctly classified cases total number of cases
(1)
(2)
(3)
(4)
(5)
3. Results Of the 45 participants, 39 were male and 6 were female with a mean (SD) age of 41.0 (12.6) years. The injury level of the participants varied from cervical level 5 to lumbar level 5, with 14 participants having injuries at or above thoracic level 3, and 31 participants having injuries at or below thoracic level 4. Twentythree of the 45 participants had an incomplete injury. The average number of years participants had used a manual wheelchair was 12.6 (8.6) years. Self-reported PA indicated that 36 participants performed some form of regular PA; 5 performed occasional PA; and 4 performed no regular PA. All 45 participants completed the study. Out of this total number, 37 finished 10 trials, 7 finished 9 trials, and 1 finished 8 trials. Most of these activity trials (701 out of 707 trials) were performed for 6 min per trial as requested, while 6 trials took between 2–4 min. Due to device malfunction of the G-WRM and wockets, as a result of a Bluetooth blockage or battery discharge, data from 24 trials (out of the 707) were lost across a range of activities. The PAMS-Arm, PAMS-Wrist, and G-WRM were able to classify the first phase of wheelchair-based PAs into near-stationary PAs, PAs that might involve wheelchair movement, and PAs with consistent wheelchair movement with an accuracy higher than 93% as they had the wheelchair wheel movement information captured by G-WRM (Table 1). The PAMS-Arm and PAMS-Wrist had reasonably high accuracy (85% and higher) for classifying the sub-activities of near-stationary PAs and PAs with consistent wheelchair movement (Table 2). In addition, the G-WRM was able to classify propulsion, basketball, and caretaker pushing with an accuracy higher than 88% (Table 2). As an example, Fig. 3 shows how the classifier model selected an acceleration-based feature (standard deviation in y direction of arm accelerometer) to detect wheelchair propulsion, caretaker pushing, and basketball activities that may have a different amount of arm movement during the activities. Classification performance analysis of the best classifiers indicated that PAMS-Arm and PAMS-Wrist had reasonably high overall classification accuracy (89.26% vs. 88.47%) with high precision and recall for most activities (Table 3). Table 4 shows the confusion matrix for PAMS-Arm, where the best classification algorithms (from Table 1) were used to classify wheelchair based PAs. The confusion matrix analysis for PAMS-Arm indicated that most of the PAs were correctly predicted (diagonal elements). In addition, the PAMS-Arm had reasonably high accuracy (85% and higher) for classifying the sub-activities of near-stationary PAs and PAs with consistent wheelchair movement using non-frequency domainbased features (Table 5).
72
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
Fig. 2. A two-step process for classifying various wheelchair-based PAs.
Table 1 Performance of various algorithms for detecting and classifying the first phase of wheelchair based PAs into near-stationary PAs, PAs that might involve wheelchair movement, and PAs with consistent wheelchair movement. Training accuracy in %
Testing accuracy in %
Features
Model
PAMS-arm
93.56
98.01
SVM
PAMS-wrist
93.56
98.01
G-WRM
93.56
98.01
Arm wocket
78.70
78.73
Wrist wocket
80.74
79.32
Root mean square of velocity from G-WRM, mean cross rate of velocity from G-WRM Root mean square of velocity from G-WRM, mean cross rate of velocity from G-WRM Root mean square of velocity from G-WRM, mean cross rate of velocity from G-WRM Ratio of dominant frequency’s power with total power in resultant acceleration from arm wocket, mean absolute deviation with respect to median in resultant acceleration from arm wocket Mean of acceleration in x axis from wrist wocket, entropy without the DC component in resultant acceleration from wrist wocket
SVM SVM DT
DT
Abbreviations used in this table: physical activities (PAs); physical activity monitoring system (PAMS); gyroscope based wheel rotation monitor (G-WRM); accelerometer device (wocket); G-WRM and wocket on upper arm (PAMS-arm); G-WRM and wocket on wrist (PAMS-wrist); support vector machines (SVM); decision tress (DT); mean value of the waveform (DC component).
Fig. 3. Plot of the acceleration feature chosen by the classification algorithm for PAMS-arm to classify and detect propulsion, caretaker pushing, and basketball activities.
4. Discussion The classification algorithms used in the first step to detect near-stationary PAs, PAs that might involve wheelchair movement, and PAs with consistent wheelchair movement indicated
that the G-WRM was a better predictor compared to the arm or wrist wockets, as G-WRM accurately detected the presence of wheelchair movement. The key features involved in the GWRM based classification algorithm were the root mean square and mean cross rate of velocity, both related to wheelchair movement. One possible reason for obtaining higher testing accuracy than the training accuracy (98.01% vs. 93.56%) for the G-WRM based classification algorithm is the lesser variation in the testing dataset due to the reduced number of participants (Table 1). In addition, the error for the testing dataset is just a single outcome compared to the error for the training dataset which was based on the 10-fold CV. During the second phase, of classifying the sub-activities for near-stationary PAs and PAs with consistent wheelchair movement, the results indicated that the arm (Fig. 3) and wrist wockets were better predictors of PAs than the G-WRM. From the features identified by the classification algorithms we see that most of them were frequency-based features, as the PAs involved repetitive use of upper arm movements. The reasonably high classification accuracy (92% for training and 88% for testing) of the G-WRM for moving PAs was probably due to G-WRM’s ability to distinguish caretaker pushing, which was performed by the same investigator, from propulsion and basketball. This pushing of wheelchair users by one individual may have resulted in a specific type of speed or acceleration pattern that the G-WRM was sensitive to.
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
73
Table 2 Performance of various algorithms for detecting and classifying the sub-activities of near-stationary PAs and PAs with consistent wheelchair movement (moving). Nearstationary PAs consist of resting, arm-ergometry, and other household activities. PAs with consistent wheelchair movement consist of propulsion, caretaker pushing, and basketball activities. Device/system
Near-stationary/ moving
Training accuracy in %
Testing accuracy in %
Features
Model
PAMS-arm
Near-stationary
86.78
84.95
DT
PAMS-arm
Moving
95.57
93.40
PAMS-wrist
Near-stationary
90.32
86.38
PAMS-wrist
Moving
97.13
93.87
G-WRM
Near-stationary
61.97
50.18
G-WRM
Moving
92.06
87.73
Arm wocket
Near-stationary
86.78
84.95
Arm wocket
Moving
96.12
91.51
Wrist wocket
Near-stationary
90.32
86.38
Wrist wocket
Moving
97.13
93.87
2nd dominant frequency content’s power in resultant acceleration from arm wocket, standard deviation of acceleration in z axis from arm wocket, amplitude of resultant acceleration from arm wocket Standard deviation of acceleration in y axis from arm wocket, amplitude of acceleration in y axis from arm wocket, entropy content of velocity from G-WRM 3rd dominant frequency content’s power in resultant acceleration from wrist wocket, Mean of acceleration in x axis from wrist wocket, dominant frequency content’s power in resultant acceleration from wrist wocket Ratio of dominant frequency’s power with total power in resultant acceleration from wrist wocket, mean of acceleration in x axis from wrist wocket, 3rd dominant frequency content’s power in resultant acceleration from wrist wocket Root mean square of velocity from G-WRM, zero cross rate of velocity from G-WRM, entropy of velocity from G-WRM 3rd dominant frequency content’s power in velocity from G-WRM, mean cross rate of velocity from G-WRM, entropy of velocity from G-WRM 2nd dominant frequency content’s power in resultant axis from arm wocket, standard deviation of acceleration in z axis from arm-wocket, amplitude in resultant axis from arm wocket Ratio of dominant frequency’s power with total power in y axis acceleration from arm wocket, 3rd dominant frequency content’s power in y axis from wrist wocket, root mean square of x axis acceleration from arm wocket 3rd dominant frequency content’s power in resultant acceleration from wrist wocket, mean of acceleration in x axis from wrist wocket, dominant frequency content’s power for resultant acceleration from wrist wocket Ratio of dominant frequency’s power with total power in resultant acceleration from wrist wocket, mean of acceleration in x axis from wrist wocket, 3rd dominant frequency content’s power in resultant acceleration from wrist wocket
SVM
DT
DT
NB DT
DT
DT
DT
DT
Abbreviations used in this table: physical activities (PAs); physical activity monitoring system (PAMS); gyroscope based wheel rotation monitor (G-WRM); accelerometer device (wocket); G-WRM and wocket on upper arm (PAMS-arm); G-WRM and wocket on wrist (PAMS-wrist); support vector machines (SVM); decision tress (DT); Naïve Bayes (NB).
An in-depth classification performance analysis (Table 3) of the best classifiers indicated that PAMS-Arm and PAMS-Wrist had similar overall classification accuracy (89.26% vs. 88.47%), precision (lower precision for resting and may be moving) and recall (lower recall for resting and basketball). The analysis revealed that multimodal sensor information from G-WRM and wocket for PAMS-Arm and PAMS-Wrist had higher accuracy than the individual devices (G-WRM and wockets). The confusion matrix analysis (Table 4) for PAMS indicated that most of the PAs were correctly predicted (diagonal elements). Further, the analysis indicated that misclassified PAs (non-diagonal) were classified into biomechanically similar PAs; for example, resting was classified as other household PAs in stationary category, and basketball as PA that may involve wheelchair movement because of the intermittent wheelchair movement. Even though misclassification resulted in lower classification accuracies, grouping activities into similar type of wheelchair-based activities can lead to energy expenditure or metabolic equivalent of task values similar to the values obtained for correctly identified activities. It is also evident that the wockets on arm or wrist have higher accuracy than the G-WRM because the arm plays a major role in all the type of PAs studied here. In addition, the overall classification performance of arm and wrist wockets (70.38% vs. 74.55%) was almost identical, with the wrist wocket having a slightly higher accuracy. These results, along with the results from the PAMS-Arm and PAMS-wrist above, indicate
that the users have the option of wearing the wocket on their arm or wrist based on their comfort or preference. However, the users will have to provide this detail to the Android based application prior to using PAMS so that appropriate classification models can be chosen to detect wheelchair based PAs. Since many of the classifier algorithms used frequency-based features to detect and classify PAs, we also evaluated the classification performance for non-frequency domain based features, which are computationally less expensive to generate. The analysis (Tables 2 and 5) indicated that the classification accuracy for PAMSArm was similar for near-stationary PAs (0.87 vs. 0.87 for training and 0.85 vs. 0.81 for testing) and PAs with consistent wheelchair movement (0.96 vs. 0.95 for training and 0.93 vs. 0.94 for testing). Comparing the features chosen by the classifier algorithms for frequency and non-frequency based features indicated that the classifier algorithms picked slightly different features with a similar type of movement information, such as rate vs. frequency, and change in distance vs. entropy. For example, the classifier chose the mean cross rate of acceleration in resultant direction using non-frequency based feature instead of the power of second dominant frequency for non-moving PAs; and the mean velocity of the G-WRM was chosen instead of the entropy of the velocity of the G-WRM feature for moving PAs. As the use of frequency based features did not seem to improve the accuracy, we plan to use non-frequency based features when programing the smartphones.
74
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
Table 3 Classification performance in terms of precision, recall or sensitivity, specificity (or true negative rate), accuracy for each activity and overall accuracy for the best classifiers on the validation dataset for PAMS-arm, PAMS-wrist, G-WRM, arm wocket, and wrist wocket. Device/system
First level activity
Sub-activity trials
PAMS-arm
Near-stationary PAs
Resting Arm-ergometry Other household activities Propulsion Caretaker pushing Basketball May be moving Resting Arm-ergometry Other household activities Propulsion Caretaker pushing Basketball May be moving Resting Arm-ergometry Other household activities Propulsion Caretaker pushing Basketball May be moving Resting Arm-ergometry Other household activities Propulsion Caretaker pushing Basketball May be moving Resting Arm-ergometry Other household activities Propulsion Caretaker pushing Basketball May be moving
PAMS-wrist
G-WRM
Arm wocket
Wrist wocket
PAs with consistent wheelchair movement May be moving Near-stationary PAs PAs with consistent wheelchair movement May be moving Near-stationary PAs PAs with consistent wheelchair movement May be moving Near-stationary PAs PAs with consistent wheelchair movement May be moving Near-stationary PAs PAs with consistent wheelchair movement May be moving
Precision 0.88 0.98 0.81 0.95 0.96 1.00 0.57 0.74 0.91 0.87 0.94 0.94 0.94 0.57 N/A N/A 0.50 0.91 0.90 0.62 0.57 0.88 0.81 0.63 0.89 0.67 0.88 0.21 0.25 0.97 0.76 0.98 0.42 0.88 0.24
Recall
Specificity
Accuracy (%)
Overall accuracy (%)
0.47 0.92 0.96 0.99 0.94 0.57 1.00 0.68 0.93 0.88 0.97 0.96 0.53 1.00 0.00 0.00 1.00 0.90 0.79 0.60 1.00 0.47 0.89 0.79 0.65 0.55 0.50 0.83 0.02 0.93 0.84 0.86 0.83 0.23 0.75
0.99 1.00 0.91 0.98 1.00 1.00 0.98 0.98 0.98 0.95 0.98 0.99 1.00 0.98 1.00 1.00 0.62 0.97 0.99 0.98 0.98 0.99 0.95 0.82 0.97 0.97 1.00 0.92 0.99 0.99 0.90 0.99 0.88 1.00 0.94
94.43 98.21 92.84 98.41 99.01 97.42 98.21 94.83 97.22 93.04 97.61 99.01 97.02 98.21 90.66 81.71 72.56 95.03 97.22 95.43 98.21 94.43 94.23 81.51 88.47 93.24 96.62 92.25 90.26 98.21 88.27 95.83 87.67 95.23 93.64
89.26
88.47
65.41
70.38
74.55
Abbreviations used in this table: TN physical activity monitoring system (PAMS); gyroscope based wheel rotation monitor (G-WRM); accelerometer device (wocket); G-WRM and wocket on upper arm (PAMS-arm); G-WRM and wocket on wrist (PAMS-wrist).
Table 4 Confusion matrix for PAMS-arm on validation dataset (20% of participants’ data not used for training) using the best algorithms to classify wheelchair based PAs. True\predicted
Resting
Arm-ergometry
Other household activities
Propulsion
Caretaker Pushing
Basketball
May be moving
Resting Arm-ergometry Other household activities Propulsion Caretaker pushing Basketball May be moving
22 0 3 0 0 0 0
0 85 2 0 0 0 0
24 6 135 0 0 1 0
0 0 0 134 2 5 0
0 0 0 1 44 1 0
0 0 0 0 0 17 0
1 1 0 0 1 6 12
Abbreviations used in this table: physical activities (PAs); physical activity monitoring system (PAMS); gyroscope based wheel rotation monitor (G-WRM); accelerometer device (wocket); G-WRM and wocket on upper arm (PAMS-arm).
Our algorithms’ overall classification accuracy for the detection of wheelchair based PAs (PAMS-Arm: 93.6% for training and 98.0% for testing), and the detection of PAs involving wheelchair movement (PAMS-Arm: 95.6% for training and 93.5% for testing) was
similar to Sonenblum et al.’s detection of wheelchair movement for various wheelchair-related activities of daily living (90–96%) [6]. This resemblance, in addition to our previous validation of GWRM [13], indicates that our system can measure wheelchair use in
Table 5 Performance of classification algorithms for PAMS-Arm using 80–20 CV to detect and classify near-stationary PAs and PAs with consistent wheelchair movement (moving) using non-frequency domain-based features. Near-stationary PAs consist of resting, arm-ergometry, and other household activities. PAs with consistent wheelchair movement consist of propulsion, caretaker pushing, and basketball activities. Near-stationary/moving
Training accuracy in %
Testing accuracy in %
Features
Model
Near-Stationary
85.61
86.02
SVM
Moving
95.84
93.87
Mean cross rate of resultant acceleration from arm-wocket, mean cross rate of acceleration in z axis from arm-wocket, Amplitude of acceleration in z axis from arm wocket Standard deviation of resultant acceleration from arm-wocket, Amplitude of acceleration in x axis from arm wocket, Mean of velocity from G-WRM
DT
Abbreviations used in this table: physical activities (PAs); physical activity monitoring system (PAMS); gyroscope based wheel rotation monitor (G-WRM); accelerometer device (wocket); G-WRM and wocket on upper arm (PAMS-arm); support vector machines (SVM); decision tress (DT).
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
activities of daily living. The advantage of our system is that it can track both wheelchair and arm movements towards tracking PA. Additionally, the classification performance results (overall accuracy: 96%, sensitivity: 99% and specificity: 98%) for detecting wheelchair propulsion using PAMS-Arm were slightly higher than the results of Postma et al. who detected wheelchair propulsion compared to other activities (overall agreement: 92%, sensitivity: 87% and specificity: 92%) [11]. The advantage of our system is the reduced number of devices: G-WRM on wheel and arm wocket instead of the six-accelerometers based activity monitoring system used by Postma et al. The overall classification accuracies for PAMS-Arm (89.3%) and PAMS-Wrist (88.5%) were lower compared to those in our previous study, during which we used the SenseWear activity monitor to detect resting, wheelchair propulsion, arm-ergometry and deskwork activities (96.3%) [10]. The lower classification accuracy was probably due to testing of a larger number of wheelchair-based PAs, performed at a self-chosen pace obtained in structured, semi-structured and unstructured natural environments. One of the limitations of this study was that a large percentage of our participants self-reported that they were physically active on a regular (N = 36) or occasional basis (N = 5). Thus, the PA levels reported here are significantly higher compared to those in Washburn et al., who indicated that out of their participants only 13–16% of persons with SCI reported consistent PA [17], and the majority reported virtually no regular PA [18]. However, our high PA levels might have been inflated due to self-reported PA levels, which tend to have a social acceptability bias. Another limitation of this study is the inclusion of a large cohort of veterans with SCI tested during the NVWG (N = 20). The majority of the veterans who took part in this study probably had a better standard of care regarding assistive technology; however, they still had current PA and health levels similar to the general population. Future studies should recruit a greater percentage of manual wheelchair users from the general population and identify additional wheelchair based activities to develop new classification models. Additionally, our study only recruited individuals with SCI to limit the influence of various disabilities on wheelchair related PAs. Recruiting individuals with other disabilities would enable us to develop classification algorithms to quantify PA levels among different groups of persons with similar disabilities. Another limitation of the study is that we were unable to evaluate the validity of PAMS in all three environments separately due to the limited dataset. Future studies should evaluate the validity of PAMS in various settings separately. Moreover, the classification models developed here were based on arm acceleration and wheelchair velocity, both movementbased variables. Future work should develop PAMS to incorporate other forms of physiologic sensing, such as galvanic skin response, skin temperature, near body temperature, and heart rate, in order to detect resistance-based PAs. We are in the process of evaluating and developing PAMS to estimate energy expenditure so that it can be reliably deployed into longitudinal testing of one to two weeks in the community. In future, we plan to study specific movement related features, which could assist researchers and clinicians with quantifying upper arm movement associated with carpal tunnel syndrome or overuse syndrome.
5. Conclusion The outcomes of this study indicate that multi-modal information obtained from PAMS can help detect various types of wheelchair based activities in structured laboratory, semistructured NVWG, and unstructured home environments with reasonably high accuracy. The study also indicated that the overall
75
classification performance of PAMS was higher than the individual components (G-WRM, wocket on the upper arm or wrist). Furthermore, technologies such as PAMS can provide real-time feedback about physical activity levels and can motivate individuals to achieve a healthy lifestyle. Ethical approval The study was approved by the Institutional Review Board of the University of Pittsburgh (PRO10060368), US Army Medical Research & Material Command’s Human Research Protection Office (HRPO log number: A-16284.1), and the VA Pittsburgh Healthcare System (MIRB/PRO#: 02967). Acknowledgements The work is supported by the Department of Defense (W81XWH-10-1-0816). SVH’s work on this article was funded through the Switzer Research Fellowship (H133F110032) awarded by the National Institute on Disability and Rehabilitation Research, Department of Education. The work is also supported by the Human Engineering Research Laboratories, VA Pittsburgh Healthcare System. The contents do not represent the views of the Department of Veterans Affairs or the United States Government. The authors thank their colleagues at the Human Engineering Research Laboratories, especially Alejandra Manoela Ojeda, Monsak Socharoentum, Natthasit Wongsirikul, and Matthew Hannan for their input and effort during development and data collection. Conflict of interest statement The authors have no conflict of interest. References [1] U.S. Department of Health and Human Services. Healthy People 2020. Washington, DC: U.S. Department of Health and Human Services; 2011. Available: http://www.healthypeople.gov/2020/topicsobjectives2020/overview.aspx? topicid=9 [2] Centers for Disease Control and Prevention. Overweight and Obesity: Among People with Disabilities. Centers for Disease Control and Prevention; 2010. [3] Rimmer JH, Yamaki K, Davis BM, Wang E, Vogel LC. Obesity and overweight prevalence among adolescents with disabilities. Prev Chronic Dis 2011;8:1–6. [4] Buchholz AC, Pencharz PB. Energy expenditure in chronic spinal cord injury. Clin Nutr Metab Care 2004;7:635–9. [5] Warms CA, Belza BL. Actigraphy as a measure of physical activity for wheelchair users with spinal cord injury. Nurs Res 2004;53:136–43. [6] Sonenblum SE, Sprigle S, Caspall J, Lopez R. Validation of an accelerometerbased method to measure the use of manual wheelchairs. Med Eng Phys 2012;34:781–6. [7] Tolerico ML, Ding D, Cooper RA, Spaeth DM, Fitzgerald SG, Cooper R, et al. Assessing mobility characteristics and activity levels of manual wheelchair users. J Rehabil Res Dev 2007;44:561–72. [8] Lee M, Zhu W, Hedrick B, Fernhall B. Estimating MET values using the ratio of HR for persons with paraplegia. Med Sci Sports Exerc 2010;42:985–90. [9] Coulter EH, Dall PM, Rochester L, Hasler JP, Granat MH. Development and validation of a physical activity monitor for use on a wheelchair. J Spinal Cord 2011;49:445–50. [10] Hiremath SV, Ding D, Farringdon J, Vyas N, Cooper RA. Physical activity classification utilizing SenseWear activity monitor in manual wheelchair users with spinal cord injury. Spinal Cord 2013;51:705–9. [11] Postma K, Berg-Emons van den HJG, Bussmann JBJ, Sluis TAR, Bergen MP, Stam HJ. Validity of the detection of wheelchair propulsion as measured with an activity monitor in patients with spinal cord injury. Spinal Cord 2005;43:550–7. [12] Shuger SL, Barry VW, Sui X, McClain A, Hand GA, Wilcox S, et al. Electronic feedback in a diet- and physical activity-based lifestyle intervention for weight loss: a randomized controlled trial. Int J Behav Nutr Phys Act 2011;8:41:1–8. [13] Hiremath SV, Ding D, Cooper RA. Development and evaluation of a gyroscope based wheel rotation monitor for manual wheelchair users. Spinal Cord Med 2013;36:347–56. [14] Hiremath SV, Ding D, Farringdon J, Cooper RA. Predicting energy expenditure of manual wheelchair users with spinal cord injury using a multi-sensor based activity monitor. Arch Phys Med Rehabil 2012;93:1937–43.
76
S.V. Hiremath et al. / Medical Engineering & Physics 37 (2015) 68–76
[15] Intille SS, Albinali F, Mota S, Kuris B, Botana P, Haskell WL. Design of a wearable physical activity monitoring system using mobile phones and accelerometers. In: Proceedings of Annual International Conference of the IEEE Engineering in Medicine and Biology Society. 2011. p. 3636–9. [16] Witten IH, Frank E. Data mining practical machine learning tools and techniques. 2nd ed. San Francisco, CA: Morgan Kaufmann; 2005.
[17] Washburn R, Hedrick BN. Descriptive epidemiology of physical activity in university graduates with locomotor disabilities. Int J Rehabil Res 1997;20: 275–87. [18] Fernhall B, Heffernan K, Jae SY, Hedrick B. Health implications of physical activity in individuals with spinal cord injury: a literature review. J Health Hum Serv Adm 2008;30:468–502.