Journal Logo


Accurate Prediction of Energy Expenditure Using a Shoe-Based Activity Monitor


Author Information
Medicine & Science in Sports & Exercise: July 2011 - Volume 43 - Issue 7 - p 1312-1321
doi: 10.1249/MSS.0b013e318206f69d


Physical activity (PA) levels and the energy expenditure (EE) associated with PA influence human health (33). As a result, individuals are advised to participate in programs that promote increased EE via exercise, PA, and changes in posture allocation (e.g., less sitting) (16). Accurately quantifying levels of PA and associated EE in adults and children will provide insights into the dose-response relationship between PA/EE and health outcomes, allow evaluation of the effectiveness of interventions that aim to increase PA/EE, and aid in treating metabolic disorders associated with obesity. Monitoring PA patterns objectively (e.g., via accelerometry (ACC)) can improve PA/EE estimates, but devices that can accurately estimate total daily and activity-specific EE are essential. For example, the magnitude of positive energy balance that results in gradual weight gain is on the order of 25-100 kcal·d−1. In addition, instruments that are unobtrusive and easy to use may improve compliance and reduce limitations to PA owing to the device interfering with movement.

ACC has emerged as one of the most popular approaches to EE prediction (6,10,12,15,25,28). Although useful, single accelerometers have one major drawback in that they tend to significantly underestimate the energy cost of static postures such as standing activities (e.g., household tasks) and non-weight bearing activities (e.g., cycling) (18). As a result, they fail to explain a considerable portion of EE variability in daily living tasks. One strategy to improve EE estimation has been to use multiple sensors, either additional accelerometers or other types of sensors (e.g., HR) (7,16,29,30). For example, combining HR and ACC has been shown to substantially improve the accuracy of EE prediction (7,29,30), as has the use of multiple accelerometers (35). Recently, several studies have demonstrated improved EE estimation with a single accelerometer by using more sophisticated modeling approaches including artificial neural networks (28), distributed lag and spline modeling (11), and branched algorithms (12,13). Another way to achieve an improvement in EE accuracy has been to use the ACC data to classify activity, which is used in predictive models based on the type or intensity of the activity (6,12,28).

HR monitoring and multiple ACC are the most common ways explored to supplement single ACC in EE prediction. Exploration of other approaches may lead to an improved prediction accuracy and greater convenience to a weight management participant. Recently, we developed a wearable shoe-based device (27) that has an embedded accelerometer and pressure sensors positioned in the insole. The main appeal of using the device for EE prediction is its potential accuracy, nonintrusiveness, light weight, and ease of use. We have developed a posture and activity recognition model for this device, which is able to achieve 98% accuracy in subject-independent classification of six major postures and activities (sitting, standing, walking, ascending stairs, descending stairs, and cycling) (26). This enables an activity-specific branched approach to EE prediction that may result in relatively good EE estimates for a variety of daily living tasks. The inclusion of insole pressure sensors in the device also allows the exploration of whether the intensity of PA may be correlated with range and frequency of foot pressure changes and whether pressure data can supplement the accelerometer data for further improvement in the accuracy of predicting EE. Thus, we developed and validated a method for using accelerometer and pressure sensors signals to predict EE conditioned on a specific activity group and without need of individual calibration. Several studies reported using shoe-based sensors (3,17,20); however, their research concentrated on detecting gait characteristics rather than posture/activity classification and EE estimation. There are also several commercially distributes shoe-based systems (such as Pedar [31] and F-Scan [14]), which incorporate pressure sensors in the insoles for the dynamic pressure measurements. Although these systems have wide applicability such as kinetic analysis of gait, shoe research and design, orthotic design, podiatry, and sports biomechanics, they are not designed specifically for posture/activity recognition and EE prediction. A study reported in Zhang et al. (34) used an array of 32 plantar pressure sensors to classify locomotion (walking, running, and up/down stairs). Another study (32) estimated daily EE using a foot-contact pedometer but did not attempt to classify postures or specific activities with the device. We introduce a shoe-based device that will be the first in the area of footwear-based systems to be used for accurate posture/activity recognition and EE estimation.

The main purpose of this study was to test the overall feasibility of EE prediction using a novel shoe-based device; in particular, we aimed to perform the following tasks: 1) to compare the accuracy of EE prediction using this device versus existing methods using single ACC or ACC/HR sensors, 2) to compare the accuracy of prediction performance of a model using accelerometer and pressure sensors signals versus a model that uses only accelerometer signal, 3) to validate the branched modeling approach for prediction of EE for each specific posture and activity, and 4) to evaluate the need of sensors to be embedded in both shoes. We hypothesized that the combination of ACC and pressure data would provide more accurate EE estimates compared with single ACC/EE methods and that sensors would only be required in a single shoe.



Sixteen adult subjects participated in the study. The university's institutional review board approved the study, and each subject provided informed consent. To test the device on a diverse population, we recruited participants who were lean to obese. On the basis of self-report, participants weight was stable (<2-kg weight fluctuation) during the previous 6 months. Individuals who were healthy, nonsmokers, and sedentary to moderately active (less than two to three bouts of exercise per week or participation in any sporting activities <3 h·wk−1) were invited to participate in the study. Pregnant women and those who had impairments that prevented PA were excluded. The physical characteristics of participants are shown in Table 1.

Subject characteristics.

Study design.

Participants reported to the laboratory in a fasted state (>4 h) for a single 3-h visit. Each participant was asked to perform a variety of postures/activities while wearing a portable metabolic cart system and the appropriately sized shoe device with embedded sensors. The postures included sitting and standing, and the activities included walking, jogging, stair ascent/descent, and cycling (Table 2). Each posture/activity trial was 6 min in duration, and subjects were allowed 5 min of rest between trials. Trial order was not randomized. Metabolic data were not collected during stair ascent/descent because this activity was performed in a two-storey stairwell that did not allow establishment of metabolic steady state. As a result, we estimated EE as each participant performed 13 different activities from four posture/activity groups (Sit, Stand, Walk/Jog, and Cycle).

Study protocol.

Participants were not restricted in the way they assumed postures and or performed activities. Standing did not require any specialized equipment; a chair with a rigid back was used for sitting; walking/jogging was performed on a motorized treadmill (Gait Trainer 1; Biodex, Shirley, NY); cycling used a bicycle ergometer (Ergomedic 828E; Monark, Uppsala, Sweden). During the fidgeting trials, subjects were allowed to make small, normal leg movements (e.g., crossing legs or shifting weight).

EE measurement.

To determine metabolic rate and associated EE during each trial, we measured the rates of oxygen consumption (V˙O2) and carbon dioxide production (V˙CO2) using a portable open-circuit respirometry system (Oxycon Mobile; Viasys, Yorba Linda, CA). Before the experimental trials, we calibrated the system with known gas concentrations and volumes. For each trial, we allowed 4 min for subjects to reach steady state (no significant increase in V˙O2 during the final 2 min and a RER <1.0) and calculated the average V˙O2 and V˙CO2 (mL·s−1) during minutes 4-6 of each trial. We calculated gross metabolic rate (W·kg−1) from V˙O2 and V˙CO2 using a standard equation (6). EE was then calculated from V˙O2 and RER.

Movement and foot pressure measurement.

The sensor data for this study were collected by a wearable sensor system embedded into shoes (Fig. 1). Each shoe incorporated five force-sensitive resistors embedded in a flexible insole and positioned under the critical points of contact: heel, metatarsal bones, and the great toe (hallux). The acceleration data were collected from a three-dimensional MEMS accelerometer positioned on the back of the shoe. The goal of the accelerometer was to detect orientation of the shoe with respect to gravity, to characterize the motion trajectory and to help characterize the amount of movement in a specific posture or activity. Pressure and acceleration data were sampled at 25 Hz and sent over a wireless link to the base computer.

Shoe device: A, Overall view of the shoe device. B, The rear view of a shoe including the accelerometer, battery, and power switch. C,Pressure-sensitive insole with five pressure sensors: 1, heel; 2, third metatarsal head; 3, first metatarsal head; 4, fifth metatarsal head; and 5, hallux.

The wireless system used for data acquisition was based on Wireless Intelligent Sensor and Actuator Network (WISAN) (21). The battery, power switch, and the WISAN board were installed at the back of the shoe as shown on Fig. 1B. The sensor system was lightweight (<40 g) and created no visible interference with the motion patterns in subjects.


For the model construction, we used a group rather than an individual approach: the data used for training were pooled from several subjects, and such model was then tested on the validation set, which included data from subject(s) who were not in the training set. For each posture and activity, the sensor data were collected during a 1-min interval in which subjects were in metabolic steady state (minutes 4-6 of each trial). Each 1-min recording resulted in approximately 1500 (25 Hz·60 s) points of pressure and acceleration data per channel. For the 16 subjects who participated in the study, there were a total of 208 such recordings.

The following data were available for each recording:

  • response variable: EE (kcal·min−1);
  • anthropometric measurements (weight, height, body mass index (BMI), age, gender, shoe size);
  • triaxial accelerometer signals: superior-inferior acceleration (Acc1), medial-lateral acceleration (Acc2), anterior-posterior acceleration (Acc3); and
  • pressure sensors signals: heel (Sens1), third meta (Sens2), first meta (Sens3), fifth meta (Sens 4), and hallux (Sens5).

To validate the branching approach, EE prediction was performed as a two-step process, with step 1 being classification of postures/activities into one of the four groups, namely, "Sit," "Stand," "Walk," and "Cycle," and step 2 being prediction of EE using one of the four regression models built for a given posture/activity group. Each 1-min interval of sensor data was first classified as belonging to one of the four activity groups using our earlier developed algorithm for posture/activity recognition (26). The same sensor data from each 1-min interval were consequently used for training and validation of one of the four regression models for predicting EE. Thus, the branching approach involved constructing four branch models, namely, "Sit," "Stand," "Walk," and "Cycle," contingent on previous classification of every 1-min recording into one of these groups for training or validation. Another major goal was to justify the use of the pressure sensors (in addition to accelerometer) in EE prediction. This led to the development of the following four models to predict EE (kcal·min−1) using predictors described above:

  1. BACC-PS: This model was branched by activity and consisted of four separate branch models ("Sit," "Stand," "Walk," and "Cycle"). The predictors included anthropometric measurements, accelerometer, and pressure sensors as predictors.
  2. BACC: This branched model also consisted of four separate branch models ("Sit," "Stand," "Walk," and "Cycle") and included anthropometric measurements and accelerometer data as predictors, but the pressure data were not used.
  3. ACC-PS: This was a nonbranched model (no activity classification) that used the same predictors as BACC-PS.
  4. ACC: This was a nonbranched model using the same predictors as BACC.

The purpose of constructing different models was to investigate whether the performance is improved by branching the model (i.e., classifying the activity) and also by including predictors derived from pressure signals.

Accelerometer and pressure sensors signals expressed in ADC units (the signals were digitized by a 12-bit analog-to-digital converter) were preprocessed to extract meaningful metrics to be used as predictors for the model. For both pressure and acceleration sensors, all of the following metrics were extracted and tested for the inclusion into each model as predictors: coefficient of variation (cv); standard deviation (std); number of "zero crossings" (zc), i.e., number of times the signal crosses its median normalized by the signal's length; entropy H of the distribution X of signal values (ent) computed as: H(X) = −∑pklogpk, where pk is the relative frequency of values fallen into the kth interval (of 20 equally sized intervals) in the sample distribution of signal values. These metrics were selected for the following reasons. The coefficient of variation (CV) and SD of a signal should indicate the amount of motion produced during recording, with the difference that CV are affected by the signals' mean value (e.g., the gravitational component of acceleration), although SD is not. The number of median crossings is an indicator of the frequency of changes in the signal, which is important to identify the intensity of motion (like speed of walking). Entropy reflects the distribution of the signal across the range of its values and is a valuable predictor for walking because, as speed of walking increases, the time of feet ground contact decreases relative to the swing time, and thus, signal values become more uniformly distributed across the range, leading to an increased entropy.

For each model, we used the derived metrics as possible predictors for the ordinary least squares linear regression. The transformed predictors (log, inverse, and square root) and interactions (as products of two or more candidate predictors) were also considered as separate linear terms within regression.

In branched models, a separate model was constructed for each type posture/activity: "Sit," "Stand," "Walk," and "Cycle." For all branched (as collections of the four separate branch models) and nonbranched models, selection of the most significant set of predictors was performed using the forward selection procedure. We used the "leave-one-out" approach for cross-validation when training and predicting the EE for each type of activity for every subject. For every left-out subject, all of the data related to this subject were removed from the training set. Model (coefficients) computed using the rest of the subjects sample was then used to predict the EE for all trials of the left out subject. The best set of predictors had to provide the best fit (by producing the maximum adjusted coefficient of determination (R2adj) and the minimum Akaike Information Criterion (AIC)) in the training step and the best predictive performance (the minimum mean squared error (MSE) and the minimum mean absolute error (MAE)) in the validation step.

The input for the models was the data from 16 subjects who had complete metabolic and sensor data for all 13 trials. In the "walk" activity group, some subjects did not have EE record (unable to achieve metabolic steady state while jogging) or had no sensor signals recorded for some trials within this group; these 10 trials were dropped from each model's input. An additional 1-min recording for cycling activity for a particular subject contained more than 50% of corrupted data due to sensor failure. This recording was also dropped from the analysis. Thus, the sample size of the input data for each model was (16 × 13) - 11 = 197 trials.

Measured and predicted EE values in kilocalories per minute for each experiment were then converted to METs for both branched ACC-PS and ACC models and their nonbranched versions. The conversion from kilocalories per minute to METs was done by representing the EE for any given epoch as a multiple of resting EE. We used EE during quiet sitting as a valid estimate of resting metabolic rate for each subject because of established convention (1,2). This conversion was performed to enable direct comparison of our results with those that have been recently published (8,12,28).

One of the goals of the analysis was to establish the need of using sensors on both shoes. Several versions of the branched ACC-PS model (as a representative model) were constructed using accelerometer and pressure sensors data separately from each shoe and both shoes together.


The following performance assessment measures were computed for each EE prediction model:

  • RMSEMET, the root mean squared error for EE prediction expressed in METs. This error is computed as the difference between model-predicted EE and the measured EE for each trial.
  • 95% confidence intervals (CI) for RMSEMET, computed as bootstrapping estimates by generating 5000 samples of absolute errors (predicted minus actual EE) drawn from the original sample, calculating RMSEMET for each such sample and computing bounds for the middle 95% of the created population of RMSEMET values.
  • ARD, the average relative difference (signed):
  • Bias, the mean difference between predicted and measured EE in METs:
  • Interval of agreement for prediction of EE in METs, calculated as given in Ainsworth et al. (1):

Bland-Altman (4) plot analysis was conducted to reveal any systematic pattern of the error (calculated as the difference between predicted and measured EE) across the range of measurements and to assess the bias and interval of agreement for the prediction of EE.

Passing-Bablok (24) regressions (a robust alternative to least squares regression) for all four models and for two units of prediction (kcal·min−1 and METs) were constructed as described. Passing-Bablok regression is best suited for method comparison because it allows measurement error in both variables, does not require normality of errors, and is robust against outliers. In addition, Passing-Bablok regression procedure estimates systematic errors in form of fixed (by testing if 95% CI includes 0) and proportional bias (by testing if 95% CI includes 1).


Each raw signal of the accelerometer (three sensors) and five pressure sensors was represented by a vector of approximately 1500 measurements (25 measurements per second). Sample raw signal for all eight sensors is given in Figure, Supplemental Digital Content 1, for walking 2.5 mph activity. Using the raw signal data, predictors for each model were computed by the following approach. Metrics for the accelerometer and pressure sensors signals (cv, std, zc, and ent) were computed separately for the left and the right shoes. Then, the corresponding predictor values were formed as the average of the left and right shoe metrics for the accelerometer signals and as the maximum value of the left or right shoe metrics for the same pressure sensor signals. The reason for the maximum (rather than the average) in combining the left and right shoe metrics is that some of the pressure sensors experienced occasional failure in three subjects. The signal from a failed sensor would register as a constant zero value (no pressure), thus using the maximum pressure ensured that no data from failed sensor were used in training or validation. In particular, use of the maximum value resulted in the reduction of the corrupted data from 5% to around 1.7%.

To facilitate the branching approach, our automatic classification model (27) for posture/activity recognition was applied to each of the 197 1-min recordings to assign it into four activity groups ("Sit," "Stand," "Walk," and "Cycle") for further construction and validation of the corresponding branch models. For these data, there was 100% rate of correct classification among all 1-min recordings with respect to the four activity groups.

Final linear regression coefficients for the branched ACC-PS and branched ACC models after selection of the best set of predictors are reported in Table, Supplemental Digital Content 2, and Table, Supplemental Digital Content 3, respectively. The final nonbranched ACC-PS and nonbranched ACC model regression coefficients are given in Table, Supplemental Digital Content 4, Among the anthropometric characteristics of subjects used as possible predictors, only weight and BMI showed significance for EE prediction for all models. In particular, gender-stratified models did not show any improvement in the prediction performance. A similar effect was reported by previous studies where gender has not been shown to improve EE estimates from ACC data (5,6,19). The coefficients for all models were obtained by averaging the coefficients of the 11 runs (one for each left out subject) of the ordinary least squares regression on the training sets. Most of the CV for coefficients of all of these models were within (0.07, 0.3), which suggests that the regression coefficients were highly stable.

Almost all coefficients for all models were highly stable over all runs as given by low absolute values of CV. As can be expected, weight and BMI always explain part of the variability of each model, whereas other physical characteristics were highly correlated to weight variable and did not add to the fit or the prediction performance of either model. Results shown in Table 3 include performance comparison of the proposed BACC-PS model, BACC model, nonbranched ACC-PS, and nonbranched ACC.

EE prediction by minute.

Bland-Altman plots (constructed for both EE (kcal·min−1) and EE (METs prediction)) for all four models are given in Figure 2. Panels A and B are Bland-Altman plots for branched models, and panels C and D are Bland-Altman plots for nonbranched models. The common characteristic for all these plots (models) is that the accuracy of prediction is slightly better for small than for large EE values (i.e., better accuracy for sitting and standing).

Bland-Altman plots for shoe-based models: A, BACC-PS model (kcal·min−1). B, BACC model (kcal·min−1). C, Nonbranched ACC-PS model (kcal·min−1). D, Nonbranched ACC model (kcal·min−1).

Passing-Bablok regression analysis was conducted using MATLAB implementation (23) of the method described in Passing and Bablok (24). Examination of the presence of fixed (intercept ≠ 0 if 95% CI does not contain 0) and proportional (slope ≠ 1 if 95% CI does not contain 1) bias of the models showed that except for one case (nonbranched ACC model that showed fixed bias) none of the four models exhibited either kind of bias (see Table, Supplemental Digital Content 5,, examination of the presence of fixed and proportional bias and linearity). All ACC-PS models (branched and nonbranched) provided better prediction over the ACC models as indicated by slope values closer to the unity than those of the ACC model (see Passing-Bablok regression analysis in the supplemental materials). In addition, the branched model regression coefficients seem to be more precise because they provided the narrower CI for both slope and intercept than those for the nonbranched models.

Linearity test indicated absence of linearity for all nonbranched models, whereas for branched models, linearity was always very strong (see Table, Supplemental Digital Content 5, examination of the presence of fixed and proportional bias and linearity). Additional proof of the strength of linear relationship between predicted and measured EE values is given by correlation and concordance coefficients. There is clear tendency of both coefficients to increase from nonbranched to branched models and from ACC to ACC-PS models. Lack of linearity of the nonbranched models is also noticeable in their Passing-Bablok regression plots (Fig. 3), which show clear curvature in the scatter plots unlike in those of the branched models.

Passing-Bablok regression plots for shoe-based models: A, BACC-PS model (kcal·min−1). B, BACC model (kcal·min−1). C, Nonbranched ACC-PS model (kcal·min−1). D, Nonbranched ACC model (kcal·min−1).

As a last step, we investigated the effect of inclusion of predictors from both shoes versus a single shoe into the model using the BACC-PS model. We compared performance of the BACC-PS models that used the difference metrics derived from the difference between signal from left and right shoes and/or the best selected set of predictors (as metrics cv, std, zc, and ent) computed separately for each shoe. Overall, models based on metrics derived for both shoes perform slightly better (RMSE was within 0.68-0.70 METs) than single-shoe models (RMSE was 0.78 METs for left shoe-based model and 0.72 METs for right shoe-based model), see Table, Supplemental Digital Content 6,, comparison of BACC-PS model performance using predictors from single shoe and both shoes. However, the RMSE values were still below those found for BACC and the rest of the models. Also, the improvement of both-shoe models over single-shoe models can be attributed mostly to the lost of data owing to sensors failure: both-shoe models were able to mitigate the effect of the corrupted data by using the fact that simultaneous sensors failure on both shoes was rare and by applying either averaging or maximization to left and right shoe sensor metrics. Thus, for all practical purposes, single-shoe models can be successfully used.


Our results suggest that a shoe-based device with embedded accelerometer and pressure sensors can be used to accurately predict EE during typical postures/physical activities. Such a device may be a useful tool for individuals interested in weight management. The combination of posture allocation/PA data (e.g., minutes sitting and walking) with accurate estimates of EE can be used to help individuals modify or maintain energy balance. A shoe-based device may also be "invisible" and unobtrusive and lead to increased use, further facilitating weight management success.

The EE prediction accuracy of our device and branched model is similar to recent studies that have used single accelerometers, multiple accelerometers, and HR/accelerometer combinations. Choi et al. (11) used ActiGraph accelerometers placed at the hip, wrist, and/or ankle and distributed lag and spline modeling to predict EE and reported RMSE of ∼0.6 kcal·min−1 (0.5 METs) across a range of activities with the accelerometer mounted at the ankle. Staudenmayer et al. (28) used a single hip-mounted accelerometer (ActiGraph) and an artificial neural network to estimate EE of a variety of activities and reported an RMSE of 0.75 and 1.22 METs using activity and minute-by-minute estimates of EE, respectively. A study by Crouter et al. (12) that compared EE estimation using hip-mounted accelerometry versus indirect calorimetry reported systematic bias of 0.1 METs with 95% limits of agreement of (−1.4, 1.5) METs. Although our results are not directly comparable to those from Staudenmayer et al. and Crouter et al. because our subjects did not wear a hip-mounted accelerometer and we did not compare actual and predicted EE during the entire period of each trial, the similar RMSE values suggest good agreement. Brage et al. (8) used a device that measured HR and ACC (Actiheart) to estimate EE and found that the RMSE was within (0.87, 1.11) METs during walking/running activities. Thus, our results suggest the use of pressure and acceleration to identify activity and predict EE are at least as accurate or better compared to other, recently proposed methodologies.

Our results support the measurement of plantar pressure as a way to improve EE prediction compared with a single accelerometer. As shown in Table 3, the inclusion of pressure sensor metrics improved all prediction performance measures. In particular, RMSE was reduced ∼7% for branched models (0.77-0.69 METs) and nonbranched models (0.99-0.94 METs). There was also clear reduction in bias and the width of the interval of agreement when comparing ACC to corresponding ACC-PS models. Because there are clear differences in the magnitude and distribution of insole pressure across postures and activities, insole pressure measurement allows for accurate classification of activity, which can then be used to develop activity-specific models that improve estimates of EE. The inclusion of insole pressure also improves EE estimation within an activity classification. In particular, there was a significant decrease in error rate in estimating cycling EE. This likely due to the changes in plantar pressure that are associated with changes in the intensity of cycling, something difficult to detect using an accelerometer. It is interesting to note that we were able to achieve accurate activity classification using only the 1st metatarsal pressure sensor and three-dimensional acceleration (26) and the EE models also used the sensors under the metatarsals. This suggests that, although multiple sensors may be required to achieve a high level of classification and EE estimation accuracy, it may be possible to use fewer pressure sensors without a decrease in performance. In general, these improvements in accuracy add support to the literature demonstrating that devices that use multiple sensors improve EE estimation (7,22,35).

The use of activity-specific branched models significantly improved estimates of EE. In particular, RMSE decreased ∼25% (0.94-0.99 to 0.69-0.77 METs) and the width of the interval of prediction is reduced by almost 1 MET when branching was used. This improvement is likely sensitive to classification accuracy as estimating EE on the wrong activity could lead to substantial errors. Our classification algorithm accuracy was 100%, in part because of the combination of sensors. We elected to classify 13 activities performed by the subjects in this study into four general activity groups based on common postures and activities rather than include more specific categories. This classification attempted to address the most common issues encountered in EE estimation using accelerometers (underestimation of energy cost of standing and non-weight bearing activities such as cycling) by recognizing similar activities as one class and using a branched model for each activity class. For example, inclusion of level, incline/decline, and loaded walking as well as running data in the "walk" class resulted in a wide range of acceleration, pressure, and metabolic values that were used to develop the walk model. This likely improved the models ability to estimate EE during a locomotor task. Recently, Bonomi et al. (5) used a single hip-mounted accelerometer combined with a branched model to classify activities and the intensity of locomotor tasks (i.e., walk and run) and reported improved estimates of activity EE versus a nonclassified approach. Other studies have used branching algorithms based on ACC variability (12) or HR and accelerometer thresholds (8,13) but without activity classification and have also reported improved EE accuracy. Collectively, these results support the use of branching models (with and without activity classification) to improve EE estimation.

Developing EE prediction models based on activity classification, although intuitively appealing, also raises important questions. Chief among them is how many different classification groups are necessary. As noted above, we used a very general activity classification. Although this may have improved the prediction accuracy with our modeling approach, a narrower classification category (e.g., level walking) may have allowed for the development of less complex models to predict activity-specific EE. For example, Bonomi et al. (5) classified activity and estimated EE using a standard value from a compendium. Recent investigations have classified common activities based on a single accelerometer and have elected to use more classes of activities. Bonomi et al. (5) identified six activities (lie, sit/stand, active standing, walk, run and cycle), whereas Staudenmayer et al. (28) identified 18 activities ranging from washing dishes to running. If the focus of a device is to identify the time spent in various activities, a large number of potential activities would seem important. However, a large number of activity designations may make the combination of activity classification/EE estimation more complex, without a marked improvement in EE prediction accuracy given the similarity in EE from many activities. Clearly, additional research is needed to determine the relationship between activity classes and EE prediction accuracy.

Despite very good performance of the proposed model for EE prediction, a limitation of the stated results is a relatively small sample size (16 subjects). However, we introduced a wearable shoe-based system and aimed to test the overall feasibility of EE prediction using this new device on a relatively small pilot sample. Despite its small size, the sample covers a wide range of weight/height/BMI characteristics of subjects. As seen from the CI for RMSE for the proposed BACC-PS model even the upper limit of the interval (0.86 METs) fits within the currently reported results from existing studies on EE prediction (9,28), which allows us to conclude that the current results are reliable with the current sample size. Future work will include collection of a significantly larger data set (with respect to number of subjects and the variety and length of activities) to support results provided in this article.

Although our results are encouraging and suggest that a footwear-based system can provide accurate estimates of EE, such a system is not without limitations. For example, pressure and acceleration data outside the laboratory may be different for a given activity (e.g., cycling) and thus present a challenge for accurate EE estimation. In addition, footwear-based systems are only viable when the individual is wearing shoes. Although this is a strength for worksite or school or other daytime monitoring, it may present a challenge if the objective is estimation of EE during all waking hours. Future studies that explore optimal classification categories and test devices in a free-living (including outdoor) setting are clearly needed. Footwear-based systems will also have to be extremely rugged to withstand the environmental and physical challenges associated with this location.

In summary, our results suggest that measuring the acceleration and insole pressure in the shoe of a single foot can be used to classify activity and when combined with a branched model can accurately estimate the EE associated with common daily postures and activities. The accuracy and unobtrusiveness of a footwear-based device may become an effective weight management tool.

This work was supported in part by the University of Colorado Technology Transfer Office Proof of Concept Grant, and National Institutes of Health grant 1R43DK083229.

Dr. Sazonov and Dr. Browning have equity interest in Physical Activity Innovations LLC.

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


1. Ainsworth BE, Haskell WL, Leon AS, et al. Compendium of physical activities: classification of energy costs of human physical activities. Med Sci Sports Exerc. 1993;25(1):71-80.
2. Ainsworth BE, Haskell WL, Whitt MC, et al. Compendium of physical activities: an update of activity codes and MET intensities. Med Sci Sports Exerc. 2000;32(9 suppl):S498-516.
3. Bamberg SJM, Benbasat AY, Scarborough DM, Krebs DE, Paradiso JA. Gait analysis using a shoe-integrated wireless sensor system. IEEE Trans Inf Technol Biomed. 2008;12(4):413-23.
4. Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet. 1986;1(8476):307-10.
5. Bonomi AG, Plasqui G, Goris AHC, Westerterp KR. Improving assessment of daily energy expenditure by identifying types of physical activity with a single accelerometer. J Appl Physiol. 2009;107(3):655-61.
6. Bonomi AG, Plasqui G, Goris AH, Westerterp KR. Estimation of free-living energy expenditure using a novel activity monitor designed to minimize obtrusiveness. Obesity (Silver Spring). 2010;18(9):1845-51.
7. Brage S, Brage N, Franks PW, Ekelund U, Wareham NJ. Reliability and validity of the combined heart rate and movement sensor Actiheart. Eur J Clin Nutr. 2005;59(4):561-70.
8. Brage S, Brage N, Franks PW, et al. Branched equation modeling of simultaneous accelerometry and heart rate monitoring improves estimate of directly measured physical activity energy expenditure. J Appl Physiol. 2004;96(1):343-51.
9. Brage S, Ekelund U, Brage N, et al. Hierarchy of individual calibration levels for heart rate and accelerometry to measure physical activity. J Appl Physiol. 2007;103(2):682-92.
10. Chen KY, Sun M. Improving energy expenditure estimation by using a triaxial accelerometer. J Appl Physiol. 1997;83(6):2112-22.
11. Choi L, Chen KY, Acra SA, Buchowski MS. Distributed lag and spline modeling for predicting energy expenditure from accelerometry in youth. J Appl Physiol. 2010;108(2):314-27.
12. Crouter SE, Clowers KG, Bassett DR. A novel method for using accelerometer data to predict energy expenditure. J Appl Physiol. 2006;100(4):1324-31.
13. Edwards AG, Hill JO, Byrnes WC, Browning RC. Accuracy of optimized branched algorithms to assess activity-specific physical activity energy expenditure. Med Sci Sports Exerc. 2010;42(4):672-82.
14. F-Scan. Tekscan Products for Pressure Mapping and Force Measurement. [cited Nov 15, 2010]. Available from:
15. Garcia AW, Langenthal CR, Angulo-Barroso RM, Gross MM. A comparison of accelerometers for predicting energy expenditure and vertical ground reaction force in school-age children. Meas Phys Educ Exerc Sci. 2004;8(3):119.
16. Haskell WL, Lee I, Pate RR, et al. Physical activity and public health: updated recommendation for adults from the American College of Sports Medicine and the American Heart Association. Med Sci Sports Exerc. 2007;39(8):1423-34.
17. Havinga PJM, Marin-Perianu M, Thalen JP. SensorShoe: Mobile Gait Analysis for Parkinson's Disease Patients. 2007. Technical Report TR-CTIT-07-63. Enschede (The Netherlands): Centre for Telematics and Information Technology, University of Twente. p. 58.
18. Hendelman D, Miller K, Baggett C, Debold E, Freedson P. Validity of accelerometry for the assessment of moderate intensity physical activity in the field. Med Sci Sports Exerc. 2000;32(9 suppl):S442-9.
19. Hustvedt BE, Svendsen M, Lovo A, et al. Validation of ActiReg to measure physical activity and energy expenditure against doubly labelled water in obese persons. Br J Nutr. 2008;100(1):219-26.
20. Jagos H, Oberzaucher J. Development of a wearable measurement system to identify characteristics in human gait-eSHOE. In: Computers Helping People with Special Needs. 2008 [cited 2009 Dec 16]:1301-1304. Available from:
21. Krishnamurthy V, Fowler K, Sazonov E. The effect of time synchronization of wireless sensors on the modal analysis of structures. Smart Mater Struct. 2008;17(5):055018.
22. Levine J, Melanson EL, Westerterp KR, Hill JO. Measurement of the components of nonexercise activity thermogenesis. Am J Physiol Endocrinol Metab. 2001;281(4):E670-5.
23. Padoan A. MATLAB Central-File detail-Passing and Bablok Regression. [cited March 2, 2010]. Available from:
24. Passing H, Bablok. A new biometrical procedure for testing the equality of measurements from two different analytical methods: application of linear regression procedures for method comparison studies in clinical chemistry. Part I. J Clin Chem Clin Biochem. 1983;21(11):709-20.
25. Plasqui G, Westerterp KR. Physical activity assessment with accelerometers: an evaluation against doubly labeled water. Obesity (Silver Spring). 2007;15(10):2371-9.
26. Sazonov ES, Fulk G, Sazonova N, Schuckers S. Automatic recognition of postures and activities in stroke patients. Conf Proc IEEE Eng Med Biol Soc. 2009;1:2200-3.
27. Sazonov E, Fulk G, Yves S, Hill J, Browning R. Monitoring of posture allocations and activities by a shoe-based wearable sensor. IEEE Trans Biomed Eng. 2011:58(4):983-90.
28. Staudenmayer J, Pober D, Crouter SE, Bassett DR, 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.
29. Strath SJ, Bassett DR, Swartz AM, Thompson DL. Simultaneous heart rate-motion sensor technique to estimate energy expenditure. Med Sci Sports Exerc. 2001;33(12):2118-23.
30. Strath SJ, Bassett DR, Thompson DL, Swartz AM. Validity of the simultaneous heart rate-motion sensor technique for measuring energy expenditure. Med Sci Sports Exerc. 2002;34(5):888-94.
31. Systems Pedar. [cited Nov 15, 2010]. Available from:
32. Tharion WJ, Yokota M, Buller MJ, DeLany JP, Hoyt RW. Total energy expenditure estimated using a foot-contact pedometer. Med Sci Monit. 2004;10(9):CR504-9.
33. US Department of Health and Human Services. The Surgeon General's Call to Action to Prevent and Decrease Overweight and Obesity. Rockville (MD): Public Health Service, Office of the Surgeon General. 2001. 39 p. Available from: US GPO, Washington.
34. Zhang K, Pi-Sunyer FX, Boozer CN. Improving energy expenditure estimation for physical activity. Med Sci Sports Exerc. 2004;36(5):883-9.
35. Zhang K, Werner P, Sun M, Pi-Sunyer FX, Boozer CN. Measurement of human daily physical activity. Obesity. 2003;11(1):33-40.


Supplemental Digital Content

© 2011 American College of Sports Medicine