Journal Logo

SPECIAL COMMUNICATIONS: Methodological Advances

Energy Expenditure Prediction Using a Miniaturized Ear-Worn Sensor


Author Information
Medicine & Science in Sports & Exercise: July 2011 - Volume 43 - Issue 7 - p 1369-1377
doi: 10.1249/MSS.0b013e3182093014


The increasing shift of emphasis toward disease prevention and health promotion in the National Health Service highlights the positive effects of physical activity on lifelong health, well-being, and disease prevention (15). It has been reported that individuals who are active are almost two times less likely to die prematurely from a heart attack than their sedentary counterparts (28). There is strong evidence to suggest that people who are physically active can reduce the risk of developing coronary heart disease, stroke, and type 2 diabetes by up to 50% and alleviate the risk of premature death by about 20%-30% (34).

Objective measurement of physical activity in free-living conditions without behavior modification plays an important role in studies providing intervention strategies to increase daily activity (34). Moreover, understanding the relationship among activity levels, disease progression, and medication requires person-specific behavior profiling that can be attained by observing activity changes over time and relating them to health and well-being. Deterioration in conditions such as heart arrhythmias (13), diabetes mellitus, and hypertension (high blood pressure affecting approximately 50 million individuals in the United States alone [5]) is gradual and often subtle over time. For these conditions, continuous observation of activity levels as well as physiological parameters can provide more accurate diagnosis and tailored treatments for each subject compared with the currently used snapshot diagnosis and management. With recent advances in sensing technology, miniaturized wireless sensors can be worn unobtrusively to provide continuous information and feedback to both clinicians and patients. One example includes postoperative monitoring where increased activity is linked to improvement after surgery (2) and could lead to earlier discharge after surgery and provide continuous monitoring at home.

These developments have also led to the use of lightweight pervasive sensing platforms for activity recognition and energy expenditure monitoring. Among these platforms, accelerometer-based systems have been widely used for estimating energy expenditure but are worn over body locations such as the waist (6), legs (11), and arms (33). Two important factors, however, can affect the use of such accelerometers for pervasive measurement of energy expenditure. The first is "wearability," which is related to sensor location, size, ergonomics, and its inference to the activities being measured. The second is the ability to derive relevant descriptors of activities, also known as features, which can best relate each type of activity to energy expenditure.

Previous studies have demonstrated that optimal sensor positions on the body depend on the type of activities being monitored. Accelerometers worn on upper extremities (wrist and arm), for example, can efficiently measure energy expenditure for arm-dominant activities and sedentary activities indoors (33). Energy expenditure for outdoor activities, on the other hand, such as walking and running shows a high correlation with accelerometer signals from lower extremities (11). Acceleration at the waist correlates well with energy expenditure for whole body movement (24). Multisite acceleration measurement (19) has shown promising results but involved using multiple sensors on the arms, waist, and legs, thus limiting its practical daily use. Combining acceleration with other types of physiological sensing, such as heart rate (37), could also provide good results in terms of energy prediction. Although these systems are suitable in laboratory settings, they could affect a user's activities in the end. This depends on the sensors' use, their ergonomics, and weight, as well as the physical condition of the wearer. Many are also difficult to use in outdoor environments where equipment maintenance is not possible.

Head-worn accelerometers have been used to study the movement of the head compared to the trunk as well as gait measurement (1). During motion, the head remains stable when compared to the trunk, and therefore, its direction of movement during activities can be more representative of the body's movement. Kavanagh et al. (17,18) used triaxial accelerometers attached to the head and trunk of subjects to collect data while they were walking and showed that accelerations detected at the head were not only more regular than those at the trunk in each direction but also associated with a greater degree of coupling between directions. Although it is a relatively stable part of the body during motion, the head moves proportionately more with increasing activity (7), which is important to consider when evaluating gait. The trunk offers a stabilizing effect and maintains head stability by regulating gait-related oscillations.

In this work, we will use a miniature lightweight (7.4-g weight) ear-worn sensor for energy expenditure estimation in a population of healthy adults during normal lifestyle and sporting activities. The sensor described has previously been used for gait monitoring (1,23), postoperative activity recognition (2), and sports performance analysis (27). A novel technique has been developed to extract subtle yet important features from the sensor for accurate prediction of activity types for estimating energy expenditure. Our study is motivated by the need of a nonintrusive and accurate measurement tool and methods for continuous monitoring of daily activities and exercise of patients.


Participants and Settings

A total of 25 healthy participants (18 men and 7 women) were recruited for this study (mean ± SD age = 29.96 ± 4.53 yr). Ethical approval was obtained from St. Mary's Research Ethics Committee (08/H0712/36). All subjects gave written consent before taking part in the experiment. Physical characteristics of the volunteers are summarized in Supplementary Table 1 ( The activities chosen represent a list of lifestyle and sporting activities, similar to those used previously (8). The subjects refrained from intense physical activities in the last 2 h before taking part in the experiment. The subjects performed each of the following activities for 5 min with a rest period of 2 min in between. The cycling data for one subject in the study are incomplete because of exhaustion. Four participants could not complete the full period (5 min) of fast running. The activities are as follows:

  1. Lying down
  2. Standing
  3. Computer work
  4. Vacuuming
  5. Going up and down stairs
  6. Slow walking on a treadmill (5 km·h−1)
  7. Brisk walking on a treadmill (6.2 km·h−1)
  8. Slow running on a treadmill (9.5 km·h−1)
  9. Fast running on a treadmill (12 km·h−1)
  10. Cycling (the subjects were asked to cycle at a comfortable speed)
  11. Rowing (the subjects were asked to attempt rowing at a 2:25 split)


The ear-worn activity recognition sensor

The e-AR (ear-worn activity recognition) sensor is based on the Body Sensor Networks (BSN) platform (22,36). It is a lightweight (5.6 × 3.5 × 1.0 cm3, 7.4 g) miniature sensor that can be worn discretely behind the ear as shown in Figure 1. A three-axis MEMS (Micro-Electro-Mechanical-Systems) accelerometer is embedded into the sensor for capturing the mobility and activity information of the subject and wirelessly transmitted to a receiver connected to a laptop/tablet computer in real time. A sampling rate of 50 Hz was used in all experiments conducted in this study. Supplementary Figure 1 shows an example of 1 min (signal voltage levels) of acceleration data for a subject performing the 11 activities given above (

A, The ear-worn activity recognition sensor. B, Mean squared error for each activity shown versus feature number. MSE drops to a minimum when the optimal number of features is added then rises again owing to the addition of irrelevant features.

Indirect calorimetry

The Cosmed K4b2 system was used in this study as the reference measurement, and it was worn by all participants while performing the activities in the experiment. It has been shown to provide good repeatability for measuring mean minute ventilation (E), oxygen uptake (V˙O2), and carbon dioxide production (V˙CO2) (25). It was used previously for validation of similar activity measurements (8) and is robust enough for outdoor experiments. Calibration was performed before each test according to the manufacturer's instructions. All experiments were performed indoors with ambient temperatures between 17°C and 21°C. The V˙O2 (mL·min−1) values were converted to V˙O2 (mL·kg−1·min−1) adjusting for the subjects' weight and eventually to METs by dividing by 3.5.

Modeling Approach

Task-known and task-blind prediction.

By task-known prediction in this work, we refer to predicting energy expenditure using features extracted from sensor data while manually labeling the activities performed. In task-blind prediction, the algorithm has to predict the activities performed first then estimate the energy expenditure. Knowledge of both activities performed and energy expenditure provides a clearer picture of a person's behavior profile compared with each modality alone. Supplementary Figure 2 shows the method used for both task-known and task-blind prediction using "raw" acceleration signals from the e-AR sensor to predict METs values ( Data from both the e-AR sensor and the K4b2 were synchronized and resampled to 100 Hz; as the K4b2 provides breath-by-breath data, so does the resampling covered areas between the breaths. Feature extraction was then applied to the accelerometer data, which is discussed further below.

Feature extraction.

Features describe different characteristics of the acceleration signals collected; these include the frequency, statistical distribution, and variation of the data. A total of 44 of such features were extracted from the three-dimensional acceleration signals per minute of activity. Given the nature of the activities selected, such as climbing up and down stairs, a window of 1 min was selected to make sure representative features of the whole activity are included. In the climbing activity, for example, subjects stopped between staircases whereas they rowed at different speeds during the rowing exercise. It is possible to choose a smaller window for activities with a fixed rate, such as walking on a treadmill, but a large window was chosen for overall consistency. The features that can be extracted are summarized in Table 1. The data from the first minute of each activity were discarded to allow for changes in heart rate. The features extracted have been chosen to incorporate time-frequency changes of the acceleration signals, as well as statistical features that are relevant (29).

Features extracted from the e-AR sensor's three-dimensional acceleration signal.

Feature selection.

The aim of feature selection is to find the most "relevant" set of features for predicting energy expenditure with the lowest errors (14). The nearest-neighbor-based feature selection method was used in this study (26), using the k-nearest-neighbor (kNN) (10) algorithm estimator. This is a machine learning algorithm aiming at using kNN for regression (prediction of outputs from a set of inputs) and searching in the feature space, where subsets with the least estimation error were found.


Classification consists of assigning labels to a new point based on prior information. In this article, we aim to classify activity type by using a method that combines feature weighting with classifier design, known as boosting. Feature weighting indicates assigning weights to features reflecting their relevance in predicting the output, so important features get higher weights. Boosting combines classifiers (sets of simple rules) to "boost" prediction performance. The Adaboost (Adaptive boosting) algorithm (3,30) uses a simple classifier (chosen as a 1 − nearest-neighbor algorithm in this work) repeatedly. The importance of examples in the data set is indicated in each round by a distribution of weights over all samples. The weight of each correctly classified sample is decreased, thus allowing the subsequent steps to focus more on the difficult examples that could cause interclass confusion. The final classifier combines a weighted set of individual classifiers that can provide better classification over the data set. Details of Adaboost implementation can be found in Bartlett and Freund (3) and Schapire and Singer (30). In this work, the Adaboost algorithm was used on each activity (task) class separately, thus leading to one-versus-all classification. An ensemble of Adaboost classifiers was then formed for overall classification combining the strength of individual classifiers.

Data Analysis

Task-known prediction.

To observe the importance of each feature in predicting energy expenditure, leave-one-out cross validation was applied to the data set. For each activity, each subject's data were used as a test set, and all other subjects were used as a training set for the regression algorithm per activity. The feature selection algorithm was applied to each activity, leading to a weighting for each feature. Inputs were all features per activity, and outputs were the METs values. The relevant features were used in the kNN regression algorithm to predict the METs values for each activity per subject. Mean squared errors (MSE) between the predicted and measured METs values were calculated to quantify the prediction error.

Task-blind prediction

For task-blind prediction, the activity type is unknown. The classification algorithm is therefore aimed at predicting the activity first and then estimating the energy expenditure of the activity. This is particularly important for unrestricted, normal daily living scenarios where the types of activity may be unknown. Leave-one-out cross validation was used for classifying activity. Once the activity was classified using the Adaboost algorithm, the most relevant features for that particular activity (as determined in task-known prediction) were used in the kNN algorithm to predict the METs values per minute. MSE between the predicted and measured METs values were then calculated. Although task-blind prediction is of more interest to both researchers and clinicians, task-known prediction is necessary in our framework to find out which features are most relevant per activity. Online task-blind prediction can then be performed by first detecting the activity (using the classifier), then using the relevant features to estimate the energy expenditure.

Statistical analysis.

Statistical analysis was performed using the Matlab Statistics Toolbox (version; Mathworks, Inc., Cambridge, UK). The following performance assessment measures were calculated for both task-blind and task-known prediction:

  • Mean (with 95% confidence interval)
  • SD
  • Median
  • Minimum/maximum values.

To look at agreement between measured and predicted values, the following performance assessment measures were calculated:

  • Estimate of mean difference with the associated 95% confidence intervals.
  • The SE (which for an unbiased estimator can be considered equal to RMSE).
  • The upper and lower limits of agreement as well as their 95% confidence intervals.

Classification results are shown as success rates (percentage of correct classification/total number of samples) and confusion matrices indicating the percent of correct classification per activity class.


Task-known prediction of energy expenditure.

Figure 1B shows examples of the METs prediction error changes with the number of features selected. It is evident that not all features contribute toward positive discrimination of pattern classes; the MSE values drop to a minimum and then increase (deteriorate), largely due to the inclusion of poor, irrelevant features. This is a common problem in pattern recognition. The number of features that provided the minimum MSE was selected as the optimal number for predicting a particular activity (10 features for stair climbing in Fig. 1B, for example). This approach is more computationally efficient while optimizing prediction accuracy and data storage size, allowing for data processing performed on the sensor.

To compare with previous work using the ActiGraph combined with a similar activity routine, measured and predicted METs for task-known prediction are shown in Table 2. They are compared to the results obtained using the regression equation developed in Crouter et al. (8) and with the ActiGraph Freedson METs equation (8). It should be noted that the values for walking and running cannot be directly compared between the two studies because the treadmill speeds were different. Table 3 shows the mean difference (with a 95% confidence interval for the bias), the SD, and the lower/upper limits of agreement (including their 95% confidence interval). The SE is acceptable for most activities given the intersubject variability for activities such as vacuuming and stair walking. METs values for cycling are generally difficult to predict in studies using the ActiGraph (8), but these have been well predicted in this study. The confidence intervals for the bias are higher for activities that subjects were doing at their own pace; this is evident for vacuuming = −0.37 to 0.03 METs, stairs = −0.56 to 0.52 METs, and cycling = −0.15 to 0.79 METs, for example. It is worth noting that, in this study, the values of the measured V˙O2 and the corresponding estimated METs data for lying down seem to be higher than those of computer work and standing. This is because subjects were asked to lie down as soon as they started performing the protocol before reaching to their resting state, so interpretation of the data must take this context into consideration.

METs values for task-known prediction of the cross-validation group.
Task-known and task-blind prediction.

Task-blind prediction of energy expenditure.

For task-blind prediction, the number of activity types was reduced from 11 to 8. In this case, fast and slow running were combined together, as were fast and slow walking. Standing still and doing computer work were also combined because they are both sedentary activities. Table 4 shows the measured and predicted METs values for each type of activity. The last column also shows the success rates in classifying activity per activity type. Table 5 presents a confusion matrix between actual and predicted activity classes. Activities that have large variations between subjects, such as vacuuming or rowing, were more difficult to classify. However, the activities are generally well classified with a success rate for all activities of 88.99% and the following rates per activity type: lying down = 89.62%, standing/computer work = 99.10%, vacuuming = 76.60%, stairs = 89.13%, walking = 85.11%, running = 98.96%, and cycling = 79.79%.

Task-blind prediction.
Confusion matrix between actual and predicted activity classes for task-blind prediction.

Table 3 shows the mean difference, the SD, as well as the lower and upper limits of agreement between the measured and the predicted METs values for task-blind prediction. In general, the mean difference shows acceptable values with a maximum underestimation of −1.18 METs and an overestimation of 0.83 METs for fast running and slow running, respectively. The mean difference values for low-level activities (lying down, standing, and computer work) are low indicating a good prediction (0.11, 0.14, and −0.16 METs, respectively) with narrow 95% confidence intervals and a low SE. The mean difference values show an overestimation for vacuuming and standing (with positive confidence intervals) and an underestimation for running (with negative confidence intervals). In general, the values for the mean difference and the confidence interval for the bias are acceptable for use within daylong monitoring of activities of daily living.


This study describes the use of a novel miniaturized and lightweight e-AR sensor for activity detection and inferring human energy expenditure. By using optimal feature selection for each activity, more accurate predictions of energy expenditure were derived than those of previous work (8). The values for METs predicted using our algorithm and the measured METs using the K4b2 show good agreement for task-known prediction (Table 3).

This work was also extended to investigate energy expenditure prediction when activity is unknown (task-blind prediction). The advantage of using task-blind prediction is that it provides both an indication of the activity performed and a prediction of energy expenditure. The knowledge of activities performed during long periods is important in both rehabilitation and well-being studies. Increased activity over time is highly indicative of recovery for cardiac rehabilitation (16) and sports injury rehabilitation, for example. The results for activity detection predicting expenditure are highly encouraging for most activities considered in this work with an average success rate of 88.99%. Activities with high variations in energy expenditure depending on speed are harder to classify.

The values for METs predicted using the proposed algorithm and the measured METs using the K4b2 show good agreement for both task-known and task-blind prediction as given in Tables 3 and 4. It is somehow difficult to compare this work with previous work using different sensors, calorimetric validation and activity routines. However, the values of mean bias and limits of agreement indicate the validity of the sensor for predicting METs for a range of activities.

Compared with recently published studies in energy expenditure using accelerometers, such as Bonomi et al. (4), Staudenmayer et al. (32) and Langer et al. (21), this work presents a much lighter sensor (7.4 vs 34.8 g [5] and 70 g [21,31]) with comparable results in energy expenditure prediction. The lightweight and stability of this sensor may enhance patient compliance and lead to better wearability during long periods. Although pedometers were used in several clinical settings to get an overview of energy expenditure, they have proven to be inaccurate (9) and only capable of providing a crude measure of general physical activity (20) rather than energy expenditure. The reason for this is that pedometers can only provide step counts, which cannot capture sedentary activities (35), such as cooking or activities that include a stable upper-body position such as cycling. Accelerometers, on the other hand, such as the one used in the e-AR sensor, provide means of capturing movements in three dimensions, which renders the possibility of providing a detailed analysis of behavior, from activity type to energy expenditure per minute.

Although MEMS-based accelerometers were previously used for energy expenditure estimation in (6,8,12,19,29), these studies used a fixed set of features for all types of activities without analyzing their relationship to energy expenditure per activity. Repetitive activities, such as running, cycling and rowing, for example, show a good correlation between frequency-related features and energy expenditure, whereas activities such as lying down and doing domestic chores depend less on such features. Assigning appropriate weights to relevant features would lead to better rates of activity recognition and energy prediction. In this work, we extracted a large range of features (44 features detailed in Table 1), then assessed how relevant they are for each type of activity. This is more efficient in terms of computational cost and managing data size by discarding redundant features. Results for both task-known and task-blind prediction show better accuracy for low-level activities that are generally poorly detected with an ActiGraph (8). High-level activities are also well predicted (as shown in Table 3).

The main limitation of this study is the number of subjects taking part with a similar age range due to practical constraints of the experiment. It could be argued that energy expenditure for similar activities could be age related, and thus further, larger studies with mixed age cohort will need to be examined. Another limitation of the study is that there are more male than female subjects. This is somehow remedied by using a leave-one-out method to train the classifiers, which means that the training set would always contain both males and females, although obviously the training set has more males (17) than females (7). We have also performed the same analysis described in the methods section on the data set grouped by gender. Thus, males and females were considered separately and METs were predicted using task-known and task-blind prediction as above. The results of task-known prediction are shown in Supplementary Tables 2 ( and 3 ( for males and females, respectively. Compared with all subjects combined for task-known prediction (Table 3), we note that the males show lower SE and closer limits of agreement. The female data set shows the opposite, as the number of participants is quite low (n = 7), which would not give the model enough training examples. Task-blind prediction for the gender-separated population shows values of SE, mean difference, and limits of agreement for males (Supplementary Table 4; http://links/ that are close to those values listed in Table 3 (combined population; http://links/ However, the values for females are much larger (Supplementary Table 5;, with large errors for running (most likely because of subjects who held on to the treadmill during fast running). Because the method presented in this work uses the nearest-neighbor approach, it is important to have enough subjects for training. Thus, the larger the training set, the better new samples are predicted. It is worth noting that for population studies, the consideration of the effect of age, gender, and subject's health status on energy expenditure could lead to more representative results, and the analysis framework shown in this article can be readily generalized to such studies. More activities of daily living could also be included such as cooking, washing, and cleaning. Although periods of activity are relatively short compared with similar studies (16,33), the experimental setup is notably simpler, and one of the significant strengths of the current study is that it could be easily used to observe subjects outdoors, performing extreme sports such as climbing (27) and during long periods.

The use of miniature sensors, such as the e-AR sensor, for providing continuous unobtrusive monitoring could have an important effect on well-being monitoring during long periods. We have recently used these sensors for weeklong monitoring of patients with chronic obstructive pulmonary disease, and the patient compliance was very high (31). Combining these sensors with efficient algorithms that can provide behavior profiling and activity recognition could provide a nonintrusive method to quantify sports performance in outdoor environments without affecting athletic performance during training.

This work was supported by the EU project Wirelessly Accessible Sensor Populations and by the UK Technology Strategy Board project Smart and Aware Pervasive Healthcare Environments.

The authors thank all subjects who took part in this study, Dr. Christian Cook from UK Sports, Dr. Blair Crewther, and the Medicine & Science in Sports & Exercise® reviewers for their valuable feedback. The authors also thank Dr. Jin-Fei Zhang for his help with data collection.

The e-AR sensor is now provided commercially by Sensixa, a spin-off company from Imperial College London. The corresponding author has the right to grant on behalf of all authors and does grant on behalf of all authors an exclusive license (or nonexclusive) for government employees.

The results of this study do not constitute endorsement by the American College of Sports Medicine.


1. Atallah L, Aziz O, Lo B, Yang GZ. Detecting walking gait impairment with an ear-worn sensor. In: Body Sensor Networks 2009, San Francisco, USA. Washington, DC: IEEE Computer Society; 2009. p. 177-82.
2. Aziz O, Atallah L, Lo B, et al. A pervasive body sensor network for measuring postoperative recovery at home. Surg Innov. 2007;14(2):83-90.
3. Bartlett PL, Freund Y. Adaboost is consistent. In: Scholkopf B, Platt J, Hoffman T, editors. Advances in Neural Information Processing Systems, Vol. 19. Proceedings of the 20th Annual Conference (NIPS 2006). Cambridge (MA): MIT Press; 2007. p. 105-12.
4. Bonomi AG, Plasqui G, Goris AH, Westerterp KR. Improving assessment of daily energy expenditure by identifying types of physical activity with a single accelerometer. J Appl Physiol. 2009;107:655-61.
5. Chobanian AV, Bakris GL, Black HR, et al. The Seventh Report of the Joint National Committee on Prevention, Detection, Evaluation, and Treatment of High Blood Pressure: the JNC 7 report. JAMA. 2003;289:2560-72.
6. Choi JH, Hyun JL, Hwang T, Kim JP, Park JC, Shin K. Estimation of activity energy expenditure: accelerometer approach. In: Proceedings of the 27th Annual International Conference of the Engineering in Medicine and Biology Society. Washington, DC; IEEE Computer Society; 2005. IEEE-EMBS; 2005; p. 3830-3.
7. Cromwell RL, Newton RA, Carlton LG. Horizontal plane head stabilization during locomotor tasks. J Mot Behav. 2001;33:49-58.
8. Crouter SE, Clowers KG, Bassett DR Jr. A novel method for using accelerometer data to predict energy expenditure. J Appl Physiol. 2006;100:1324-31.
9. Crouter SE, Schneider PL, Karabulut M, Basset DR. Validity of 10 electronic pedometers for measuring steps, distance and energy cost. Med Sci Sports Exerc. 2003;35(8):1455-60.
10. Devroye L. The uniform convergence of nearest neighbor regression function estimators and their application in optimization. IEEE Trans Inform Theory. 1978;4(2):142-51.
11. Foster RC, Lanningham-Foster LM, Manohar C, et al. Precision and accuracy of an ankle worn accelerometer-based pedometer in step counting and energy expenditure. Prev Med. 2005;41(3-4):778-83.
12. Freedson P, Kozey S, Lyden K, Staudenmayer J. Neural networks to predict METs from accelerometer counts. Med Sci Sports Exerc. 2010;42(5):117.
13. Go AS, Hylek EM, Philips KA, et al. Prevalence of diagnosed atrial fibrillation in adults: national implications for rhythm management and stroke prevention: the AnTicoagulation and Risk Factors in Atrial Fibrillation (ATRIA) study. JAMA. 2001;285:2370-5.
14. Guyon I. An introduction to variable and feature selection. J Mach Learn Res. 2003;3:1157-82.
15. Haskell WL, Blair SN, Hill JO. Physical activity: health outcomes and importance for public health policy. Prev Med. 2009;49(4):280-2.
16. Jolly K, Taylor R, Lip GY, et al. The Birmingham Rehabilitation Uptake Maximisation Study (BRUM). Home-based compared with hospital-based cardiac rehabilitation in a multi-ethnic population: cost-effectiveness and patient adherence. Health Technol Assess. 2007;11(35):1-118.
17. Kavanagh J, Barrett R, Morrison S. The role of the neck and trunk in facilitating head stability during walking. Exp Brain Res. 2006;172(4):454-63.
18. Kavanagh JJ, Morrison S, Barrett RS. Coordination of head and trunk accelerations during walking. Eur J Appl Physiol. 2005;94(4):468-75.
19. Kim D, Kim HC. Estimation of activity energy expenditure based on activity classification using multi-site triaxial accelerometry. Electronics Lett. 2008;44(4):266-7.
20. Kumuhara H, Tanaka H, Schutz Y. Are pedometers adequate instruments for assessing energy expenditure? Eur J Clin Nutr. 2009;63:1425-32.
21. Langer D, Gosselink R, Sena R, et al. Validation of two activity monitors in patients with COPD. Thorax. 2009;64:641-2.
22. Lo B, Atallah L, Aziz O, ElHelw M, Darzi A, Yang GZ. Real-time pervasive monitoring for postoperative care. In: Proceedings of Body Sensor Networks 2007, Volume 1 of IFMBE. 2007. p. 122-7.
23. Lo B, Pansiot J, Yang GZ. Bayesian analysis of sub-plantar ground reaction force with BSN. In: Proceedings of the 2009 Sixth international Workshop on Wearable and Implantable Body Sensor Networks; 2009 June 3-5. 2009. p. 133-7.
24. Mathie MJ, Coster AC, Lovell NH, Celler BG. Accelerometry: providing an integrated, practical method for long-term, ambulatory monitoring of human movement. Physiol Meas. 2004;25(2):R1-20.
25. McLaughlin JE, King GA, Howley ET, Bassett DR Jr, Ainsworth BE. Validation of the Cosmed K4b2 portable metabolic system. IntJ Sports Med. 2001;22:280-4.
26. Navot A, Shpigelman L, Tishby N, Vaadia E. Nearest neighbor based feature selection for regression and its application to neural activity. In: Neural Information Processing Systems 18. Cambridge (MA): MIT Press; 2006. p. 995-1002.
27. Pansiot J, Lo B, King RC, McIlwraith D, Yang GZ. Climbsn: climber performance monitoring with BSN. In: IEEE Proceedings of the 5th International Workshop on Wearable and Implantable Body Sensor Networks 2008 (BSN). Hong Kong, China; 2008. p. 33-6.
28. Physical activity and cardiovascular health. NIH Consensus Development Panel on Physical Activity and Cardiovascular Health. JAMA. 1996;276(3):241-6.
29. Rothney MP, Neumann M, Beziat A, Chen KY. An artificial neural network model of energy expenditure using non-integrated acceleration signals. J Appl Physiol. 2007;103:1419-27.
30. Schapire RE, Singer Y. Improved boosting algorithms using confidence-rated predictions. Mach Learn. 1999;37(3):297-336.
31. Shrikrishna D, Atallah L, Lo B, et al. Comparison of a novel ear worn sensor (e-AR) with an armband sensor for physical activity monitoring in COPD. In: Annual Congress of the European Respiratory Society ERS 2010 in Barcelona, Spain; 2010 Sep 18-22. 2010. p. 511-2.
32. Staudenmayer J, Pober D, Crouter S, Bassett D, Freedson P. An artificial neural network to estimate physical activity energy expenditure and identify physical activity type from an accelerometer. J Appl Physiol. 2009;107:1300-7.
33. Swartz AM, Strath SJ, Bassett DR Jr, Obrien WL, King GA, Ainsworth BE. Estimation of energy expenditure using CSA accelerometers at hip and wrist sites. Med Sci Sports Exerc. 2000;32(9 suppl):S450-6.
34. Wannamethee SG, Shaper AG. Physical activity in the prevention of cardiovascular disease. Sports Med. 2001;31(2):101-14.
35. Wixted AJ, Thiel DV, Hahn AG, Gore CJ, Pyne DB, James DA. Measurement of energy expenditure in elite athletes using MEMS-based triaxial accelerometers. IEEE Sensors J. 2007;7(4):481-7.
36. Yang GZ. Body Sensor Networks. New York (NY): Springer-Verlag; 2006. p. 423-80.
37. Zakeri I, Adolph AL, Puyau MR, Vohra FA, Butte NF. Application of cross-sectional time series modeling for the prediction of energy expenditure from heart rate and accelerometry. J Appl Physiol. 2008;104:1665-73.


Supplemental Digital Content

© 2011 American College of Sports Medicine