Feasibility of the Energy Expenditure Prediction for Athletes and Non-Athletes from Ankle-Mounted Accelerometer and Heart Rate Monitor

Due to the nature of micro-electromechanical systems, the vector magnitude (VM) activity of accelerometers varies depending on the wearing position and does not identify different levels of physical fitness. Without an appropriate energy expenditure (EE) estimation equation, bias can occur in the estimated values. We aimed to amend the EE estimation equation using heart rate reserve (HRR) parameters as the correction factor, which could be applied to athletes and non-athletes who primarily use ankle-mounted devices. Indirect calorimetry was used as the criterion measure with an accelerometer (ankle-mounted) equipped with a heart rate monitor to synchronously measure the EE of 120 healthy adults on a treadmill in four groups. Compared with ankle-mounted accelerometer outputs, when the traditional equation was modified using linear regression by combining VM with body weight and/or HRR parameters (modified models: Model A, without HRR; Model B, with HRR), both Model A (r: 0.931 to 0.972; ICC: 0.913 to 0.954) and Model B (r: 0.933 to 0.975; ICC: 0.930 to 0.959) showed the valid and reliable predictive ability for the four groups. With respect to the simplest and most reasonable mode, Model A seems to be a good choice for predicting EE when using an ankle-mounted device.


Methods
Study design. We employed indirect calorimetry as a criterion measure (CM) and an ankle-mounted accelerometer monitoring device to measure the EE of four representative groups: athletes (non-endurance group and endurance group) and non-athletes (sedentary group and exercise-habit group). An accelerometer-based HR monitoring device was simultaneously used to measure their HR during the exercise tests (treadmill speed: 4.8, 6.4, 8.0, 9.7, and 11.3 km/h). Because HRR can directly reflect the physical fitness level of the participant, we adopted the HRR parameter as a correction factor for the prediction equation to increase the accuracy of the EE predicted. From these measurement periods, linear regression equations were determined for each physical fitness level based on average accelerometer VM, body weight (BW), and without/with HRR parameters (i.e., Model A and Model B, respectively), and the corresponding EE measured by indirect calorimetry. We explored whether modifying the traditional EE prediction equation (Freedson VM3 combination equation, 2011) 32 would be ideal for ankle-mounted devices with users of different fitness levels. The study procedures were approved by the Institutional Review Board of Fu Jen Catholic University (New Taipei City, Taiwan) and informed consent was provided by all participants before the experiment. This study was conducted in full conformance with the relevant guidelines and regulations, i.e., the principles of the Declaration of Helsinki guidelines. The study was registered at http://irb.rdo.fju.edu.tw/ (reference number: C106056). We recruited volunteers with four kinds of physical fitness characteristics using an open, independent, and random method.

Participants.
A total of 120 adults, non-athletes and athletes, voluntarily participated in this study. The participants were divided into four groups of 30 people each according to their physical fitness. Non-athletes were divided into a sedentary group (SG) (male: 40.0%, female: 60.0%) and an exercise-habit group (EHG) (male: 46.7%, female: 53.3%) according to their daily living and exercise habits. Athletes were divided into a non-endurance group (NEG) (male: 56.7%, female: 43.3%) and an endurance group (EG) (male: 63.3%, female: 36.7%) according to their specialties. The SG was defined as participants without any regular exercise habits, or time spent is sitting or lying down. The EHG was defined as participants who regularly exercised at least three days per week with moderate-intensity physical activity each time. The NEG comprised non-endurance athletes who were competitors in intercollegiate games and engaged in routine training at least 5 times per week in a specific program, with a focus on sprint races and throwing sports. The EG was composed of endurance athletes who were competitors in intercollegiate games and participated over 5 times each week in routine training in a specific program, with a focus on middle-or long-distance races. Only the volunteers with a good health condition were recruited for this study. Due to some conditions could prevent volunteers from completing the test safely, such as exercise contraindications, taking drugs that could affect the metabolic rate or cardiovascular disease. During the treadmill test, participants were equipped with experimental devices, including mask, accelerometer, and heart rate monitor in a laboratory with an average room temperature of 23 °C. The participants' anthropometric measurements and body composition features are listed in Table 1.
Anthropometric Measurements and Body Composition. Anthropometric measurements and body composition features were assessed through standard procedures. Each participant's height, body weight, and body composition were measured with the InBody 570 Body Composition Analyzer (Biospace, Inc. Seoul, Korea). The InBody 570 is a multi-frequency bioelectrical impedance analyser (MFBIA). This is a well-established method to determine body composition, such as body weight, body mass index (BMI), skeletal muscle mass, percentage of body fat, and resting metabolic rate (RMR). All participants wore light sportswear, removed metal jewelry, shoes and socks, and stood on the machine barefoot to minimize possible errors.

Measurement of EE by Indirect Calorimeter and Accelerometer Method.
For the indirect calorimeter method, the Cardiopulmonary Exercise Testing System (Vmax Encore 29 System, VIASYS Healthcare Inc., Yorba Linda, CA, USA) was employed to conduct the metabolic CM. Warming up of the Vmax system for at least 15 minutes and the calibration should be done before each test. Participants were equipped with a well-fitted, noses and mouths covering mask (Hans Rudolph Inc., Kansas City, MO, USA) which connected to flow sensor and sampling line was to collect the amount of oxygen consumption (VO 2 ) and carbon dioxide output (VCO 2 ) they breathed (breath by breath).
For the accelerometer method, all the test data was collected by an ActiGraph GT9X-Link (Actigraph Corporation, Pensacola, FL, USA), a tri-axial accelerometer (size: 3.5 × 3.5 × 1 cm, weight: ~14 g). The initialization of the ActiGraph accelerometer and chest mount Polar H10 HR monitor (Polar Electro Oy, Kempele, Finland) was done by ActiLife6 software (version 6.12.1, ActiGraph, Cary, NC, USA). In this study, the sampling frequency of the accelerometer and heart rate monitor was set to 30 Hz, and the physical activity parameters (i.e., calories, and VM), and HR data were collected in 10-second intervals set manually via ActiLife6 software. The ActiGraph accelerometer was attached to their right ankle, and the participants took one test at a time.
Treadmill Test. The setting in the laboratory, the participants were requested to complete treadmill (h/p cosmos mercury 4.0, Nussdorf-Traunstein, Germany) exercise tests (from walking to running) at speeds of 4.8, 6.4, 8.0, 9.7, and 11.3 km/h. Before the experiment started, the resting heart rate (HR rest ) in the sitting position was first measured. The participant sat in a resting position for 20 min, and the lowest HR recorded during the last 5 min was set as the resting value 30 . Each test took at least 3 minutes, and tests at different speeds were divided by two-minute breaks in an approach modified from Tudor-Locke et al. 33 . The indirect calorimetry, HR monitor, and accelerometer outputs for all tests were synchronously and continuously recorded during the process. If a participant failed to keep pace with the treadmill speed, heart rate exceeded the safety range (220 bpm -age) or gave up midway, the experiment data of this test and any personal information would be erased. Data Analysis. All 120 participants completed the exercise test, and all indirect calorimetry (Vmax system), HR monitor, and accelerometer (ActiGraph GT9X-Link) data were exported into Microsoft Excel (Excel version in Microsoft Office 2013 for Windows). The Vmax system, HR monitor data, and the ActiGraph GT9X-Link accelerometer data were used to calculate the time-series parameters of the 10 × 10 s block. According to Lyden et al. 20 , in the treadmill exercise test, the participant performs in unsteady and steady states during each specific speed interval, which means that the acquired raw data consist of unsteady and steady data, of which, steady data is considered valid. Therefore, the first 120 seconds of each activity were deleted to ensure that the participant's exercise intensity level had reached the steady state and the last 10 seconds were eliminated to minimize any researcher error in timing synchronization between the accelerometer and the metabolic measurements. After the elimination of the first 120 seconds and the last 10 seconds, we required at least 30 s of the remaining data for inclusion in the analyses. The data process is shown in Fig. 1.
Indirect Calorimetry. The VO 2 and VCO 2 values for each walking/running trial were used to calculate EE as a continuous summed value in calories (Kcal/min) to determine criterion measure EE (CM-EE) using Weir's equation 34 , which considers the energy derived from different fuel sources. The Eq. (1) is as follows: where VO 2 is oxygen consumption in liters per minute, and VCO 2 is the rate of carbon dioxide production in liters per minute. Body weight, height, and age were used to calculate resting metabolic rate (RMR) for each participant 35 . This RMR value was then subtracted from their total EE value for each walking/running trial, providing a final value for physical activity-related EE 36 . Accelerometer. The ActiGraph GT9X-Link data (GT9X-EE) were analyzed in ActiLife6 software, and the GT9X-EE was calculated for each walking/running trial using the prediction equation which developed by Sasaki et al. 37 , labeled "Freedson VM3 Combination equation (2011)" 32 in the Actilife software. This equation uses the three-dimensional (3D) vector magnitude (VM) (i.e., a combination of the tri-axial accelerometer data) to predict calories (kcal/min). The Eq. (2) is as follows: where the 3D = + + VM (axis 1) (axis 2) (axis 3) 2 2 2 according to 10 s epoch lengths analyzed in ActiLife6 software, and BW is body weight (kg).
All the EE values were divided by the body weight for standardization and are presented in kcal/kg/min. HRR = HR max − HR rest refers to the difference in HR max (each test stage) and HR rest , where HR max is the maximum heart rate at each test stage, and HR rest is the pre-test measure of resting HR.

Statistical Analysis.
All statistical data of this study that be adopted was analyzed via SPSS Statistics version 20 statistical software (IBM Corp., New York, NY, USA). The significance level was set to α = 0.05. All data were represented in the form of mean ± SD. The comparisons of the fundamental data and body compositions of the four groups of participants were conducted via One-way analyses of variance (ANOVAs) with post-hoc Bonferroni tests, as well as their differences in EE measurements in each speed test. To discuss the differences in measurements of the two measurement systems used in this study, CM-EE and GT9X-EE, paired t-tests, the Cohen's d effect size (ES), and mean absolute percentage error (MAPE) were determined for analysis. Linear regression with cross-validation (70% of samples for modeling and 30% of samples for validation) was used to modify the EE prediction model with the variables of VM and body weight (BW), and without/with HRR (Model A and Model B, respectively). The validity and reliability of the EE measurements were further evaluated using the Pearson coefficient of determination (r) and intraclass correlation coefficient (ICC) respectively.

Results
The ANOVA statistical data derived from the CM-EE and GT9X-EE accelerometer outputs of the four groups of different physical fitness levels (SG, EHG, NEG, and EG) are presented in Fig. 2. The results indicated a significant difference in CM-EE measurements among these four groups (p < 0.001). Apart from the speed of 4.8 km/h, when the CM-EE measurements showed minor differences (p > 0.05), the CM-EE measurements of the NEG and EG were higher than those of the SG and EHG (p < 0.001, t-test with Bonferroni correction) at the other treadmill speeds (6.4, 8.0, 9.7, 11.3 km/h). The statistical data of the GT9X-EE accelerometer output indicated that the four groups differed beyond the level of significance (p > 0.05). Table 2 lists the CM-EE and GT9X-EE values of the SG, EHG, NEG, and EG in exercise, and the differences in these two measurement systems (paired t-test), ES, MAPE, and ICC. We found that the CM-EE and GT9X-EE results of these four groups had significant differences in all speed tests (p < 0.001), rather high degrees of difference in ES (SG: 5.11-6.38; EHG: 5.96-6.59; NEG  Table 3 shows the linear regression results of two prediction models formed by VM activity counts, weight, and without/with HRR (Models A and B). The linear regression results revealed that both Models A and B had high R 2 values and low standard errors of estimate (SEE). The adjusted EE and CMEE results of Models A and B are listed in Table 4. Table 5

Discussion
We employed indirect calorimetry and ankle-mounted accelerometer outputs to examine the EE estimation of groups with different physical fitness and to establish an appropriate prediction equation. The cross-validation results showed that the differences in correlation in four groups between 70% and 30% were from 0.000 to 0.008, and all correlations were higher than 0.913. This means that the new EE prediction equations developed by this research have a high degree of utility. The EE measurement results of ankle-mounted accelerometer outputs were unlikely to identify different groups of participants, and, compared with the CMEE, over-estimation was found (p < 0.001). We employed the Freedson VM3 Combination equation (2011) 32 , and the measurement results    30 , we found that including HRR parameters in the prediction equation could calibrate for individuals' physical fitness and standardize an individual's physical fitness level, helping to increase the accuracy of EE estimates via a wearable device. This phenomenon was also found in this study. Similarly, after the inclusion of the HRR parameters in the prediction equation (Model B), the validity and reliability of the four groups' EE estimate considerably improved, but such improvement was not significant compared with that of Model A (where the coefficient of the Freedson VM3 Combination equation (2011) 32 was modified).
Physical exercise is accomplished by the contraction and relaxation of skeletal muscles. However, the contraction and relaxation of muscles function normally only when the energy supplied to them is sufficient. The larger the muscle mass, the more energy is consumed. We also found that an athlete's muscle mass (NEG: 32.1 ± 7.1 kg, EG: 29.1 ± 4.5 kg) was greater than that of a non-athlete (SG: 26.6 ± 5.7 kg, EHG: 28.4 ± 6.0 kg) (p < 0.05), the EE of the athlete group during exercise was also higher than that of the non-athlete group (p < 0.05), and a similar pattern was noted in the RMRs of the different groups. As a result, when the treadmill speed (or exercise intensity) was the same, the CMEE indicated that when the EEs of two athlete groups were higher than those of non-athlete groups, the EE of the NEG was the highest and could change according to the treadmill speed (or exercise intensity) following the same pattern. This result is identical to that of a previous study indicating that the RMR and TEE of athletes are higher than those of non-athlete groups 3 Table 4. Measured EE by Vmax system (CM-EE) and estimated EE (Models A and B) by ankle-mounted GT9X in five treadmill walking and running tests (mean ± SD).
www.nature.com/scientificreports www.nature.com/scientificreports/ the TEE of 14 male endurance athletes and non-athletes and found that the TEEs of athletes and non-athletes were 3895 ± 600 kcal/day and 2722 ± 475 kcal/day (p < 0.05) respectively, and the REE of athletes was also higher than that of non-athletes (1407.3 ± 170.2 kcal/day, and 1259.1 ± 105.6 kcal/day, respectively) 40 . Ndahimana et al. analyzed the TEE of college-level tennis players and non-athletes (aged 19 to 24 years) and reported that the TEEs of tennis players and non-athletes were 2780.3 ± 429.5 kcal/day and 2012.3 ± 160.5 kcal/day (p = 0.001), respectively 3 . Matsushita et al. discussed the TEE of 34 female rhythmic gymnasts athletes and 23 female students without exercise habits (non-athletes) and found that the TEE of athletes and non-athletes were 2317 ± 261 kcal/ day and 1907 ± 204 kcal/day (p < 0.001), respectively 39 .
Accelerometer-based physical activity monitoring devices estimate the exercise intensity, EE, and other physical activity parameters based on VM activity counts produced during the exercise. Due to the limits of such devices, the generated VM activity counts can vary when the wearing position changes. In other words, with the same exercise intensity, VM activity counts can increase or decrease due to different device-wearing positions and result in bias in EE estimates compared with standard measurements. In this study, the GT9X EE measurements at all speed levels in all four groups were overestimated when an ankle-mounted device was worn (p < 0.001), and a large difference in effect size (4.34 to 7.22) was found. The EE prediction equation currently applicable to the accelerometer device is designed for healthy adult participants, and the device is worn on the waist to collect activity parameters to develop an estimation equation using a linear regression method 19,20 . Our results indicate that although the GT9X EE and standard measurements of ankle-mounted devices demonstrated poor reliability at all speed levels (SG: ICC = 0.108; EHG: ICC = 0.108; NEG: ICC = 0.131; EG: ICC = 0.115), the GT9X EE and CMEE measurements of each group had high validity (SG: r = 0.870; EHG: r = 0.869; NEG: r = 0.871; EG: r = 0.845). Therefore, although the EE and CMEE estimated by ankle-mounted devices have high validity, they have poor reliability due to the characteristics of VM activity counts. To avoid bias, an appropriate EE estimation equation is required when the different vibration counts caused by wearing location are not considered.
To improve the limits of accelerometer sensing devices, some researchers have proposed combining the activity parameters of accelerometers with HR parameters to increase the accuracy of EE estimates 29,41-43 . Romero-Ugalde et al. highlighted the importance of activity EE in health care and the necessity of having accurate and efficient EE measures. The EE estimation equation developed through the combination of heart rate and waist-mounted accelerometer signals (R 2 = 0.8385, root mean square error (RMSE) = 59.3617 J/kg/min) can be used accurately estimate the EE of light to moderate physical activity (binary classification accuracy of 0.8155) 44 . Chang et al. increased the EE estimate accuracy of an accelerometer-based monitoring device during uphill walking and running by modifying the traditional empirical equation of accelerometers using HR and HRR as compensation factors. The results indicated that compared with HR, the combination of HRR and accelerometer parameters provides better predictive ability 30 . As already reported in the literature, heart rate measurements from wearable devices can also be affected by estimation errors. This could affect the reliability of the improved estimation equations 45 . We found that both Models A and B managed to increase the validity and reliability of the ankle-mounted device's EE measurements in all four groups, with the estimates close to the real values. The MAPE of Model A and standard measurements were divided into SG: 6.95%, EHG: 8.24%, NEG: 8.46% and EG: 2.91%; the MAPE of Model B and standard measurements were divided into SG: 6.40%, EHG: 2.62%, NEG: 4.95% and EG: 2.99%. However, we observed no significant differences between Model A (r = 0.931 to 0.972, ICC = 0.913 to 0.954) and Model B (r = 0.933 to 0.975, ICC = 0.930 to 0.960). This means that the accuracy of the two models is approximately equal, either by modifying the coefficients of traditional estimation equations (Model A) or adding HRR parameters to traditional estimation equations (Model B).
During walking and running, the ankle joint is the first body part that suffers from ground reaction force and impact force 46 . This is the same when wearable devices are used; the ankle-mounted device suffers from maximum impact force, which results in the highest VM activity counts. Hibbing et al. reported that, similarly, when 10 activities ranging from rest to running were performed when ActiGraph GT9X accelerometers were mounted on the right hip, both wrists, and both ankles simultaneously, the VM activity counts of the ankle-mounted device were higher than those of devices worn in other positions 47 . If the traditional estimation equation (Freedson VM3 Combination equation, 2011) 32 , which is mainly used for the waist, is used with the ankle-mounted device, overestimation can also result. It can be seen that it is essential to modify the coefficient of the estimation equation, or modified/developed new estimation equations, which can correct the bias caused by different wearing positions. In summary, in the treadmill walking/running tests conducted in this study, the Actigraph worn on the ankle led to overestimated EE measurements. However, the estimation accuracy can be improved by modifying the coefficient of the estimation equation (Model A) or combining HRR parameters with the estimation equation (Model B). Model A (without HRR) seems to be the best choice for EE estimation as it is the simplest and most reasonable method.  Table 5. Validity and reliability analysis of the predicted EE in models and CMEE in different groups. r, Pearson coefficient of determination. ICC, intraclass correlation coefficient.
www.nature.com/scientificreports www.nature.com/scientificreports/ In conclusion, as ankle-mounted activity monitoring devices provide relatively higher VM activity counts, and the EE estimation accuracy is therefore affected. By modifying the coefficient of the Freedson VM3 Combination equation (2011) (Model A) and including HRR parameters in the estimation equation (Model B), we enhanced the EE estimation accuracy. The rapid progress of technology has simplified data collection by using the simplest method to obtain the most accurate estimates. We found that Model A is the better model for measuring the EE estimates of ankle-mounted devices. Since the participants in this study were mainly healthy adult university students and members of university sports teams, the equation that we developed in this study may not be applicable to other groups, such as pre-school children, adolescents, the elderly, or people with specific diseases or physical difficulties.