There is still much to be learned about the nature of children's physical activity patterns and the mechanisms underlying the widely accepted health benefits of physical activity during childhood (15,18). Furthermore, the relationship between physical activity and its determinants is not fully understood yet. If research into physical activity among youth is to advance, it is vitally important to refine the methods used to assess the physical activity in which they engage.
Children's physical activity has traditionally been measured with self-reports. Self-reports are easily administered low-cost measurements. However, they do not capture the sporadic short-burst nature of children's physical activity very well (2). Furthermore, self-reports are influenced by recall bias or social desirability. Accelerometers have therefore, in recent times, become the method of choice in physical activity research. These lightweight unobtrusive devices provide objective information about the frequency, intensity, and duration of physical activity. In most studies, the raw acceleration signal is converted into activity counts. Total or mean activity counts per day and minutes per day spent above a certain intensity threshold are reported. This does not value the richness of accelerometer data (8). Using more sophisticated approaches to data processing, such as quadratic discriminant analysis, Bayesian classification, decision trees, and artificial neural network (ANN) models can improve the accuracy of accelerometer-derived measures of physical activity (3,12).
Although it is difficult to compare studies on activity classification-given the variation in study population, type of accelerometer, measurement protocol, statistical approach, intention (prediction vs classification), and outcome-their results are promising. Pattern-recognition-based approaches have been shown to be successful in classifying many controlled physical activities among adults and elderly, such as lying, sitting, standing, walking, running, cycling, and falling, on the basis of single-sensor accelerometer data (3,6,9,10,12,14,17). To our knowledge, studies applying these approaches to accelerometer data from children are currently lacking. The physical activity pattern of children is very different from that of adults. Children's physical activity pattern is characterized by frequent spasmodic bursts of short duration (2). They participate in intermittent and unstructured activities, and the type of activities children engage in changes as they develop, going from informal active play during early childhood to activities that begin to mirror those of adults during adolescence (16). Because of these differences, it remains to be seen whether activity classification based on accelerometer data is just as successful in children as in adults and elderly. In addition, in most pattern-recognition-based studies to date, a single sensor was placed on the subjects' hips, either a uniaxial, biaxial, or triaxial accelerometer (1,3,10,11,14,17). A model based on hip accelerometer data may not correctly classify certain physical activity types, such as cycling. Placing the accelerometer on the ankle may be a more efficient alternative. Furthermore, it is unclear whether the classification accuracy of a pattern-recognition-based model based on triaxial accelerometer data is higher than that of a model based on uniaxial accelerometer data. Although triaxial accelerometers have been developed under the assumption that "more is better," evidence supporting this assumption is scarce (5,7).
The purpose of this study was to identify types of physical activity (i.e., sitting, standing, walking, running, rope skipping, playing soccer, and cycling) among school-aged children using ANN models based on uniaxial accelerometer data from the hip or the ankle. Secondarily, it was examined whether the accuracy of the developed ANN models improved by using triaxial rather than uniaxial accelerometer data to identify children's physical activity type.
Subjects and data collection.
Children between the age of 9 and 12 yr were recruited from three elementary schools by sending written information about the purpose and nature of the study to their parents. Finally, 58 healthy children (31 boys, 27 girls) were permitted by one of their parents to participate in the study. The characteristics of these children are shown in Table 1. Each subject was observed by a research assistant during a fixed sequence of 20 min comprising the following activities: sitting during a writing task, standing, walking, running, rope skipping, playing soccer (i.e., kicking the ball back and forth to the research assistant), and cycling. The research assistant recorded the starting and the finishing times of each activity with a stopwatch. To imitate real life, all activities were performed at a self-paced speed. With the exception of sitting, all activities were conducted outdoors in the direct vicinity of the subject's school in similar weather conditions (i.e., no rain, mild wind). For cycling, the subjects used their own bicycle. All subjects wore uniaxial and triaxial ActiGraph accelerometers (GT1M/GT3X; ActiGraph, Pensacola, FL) on both the hip and the ankle. The ActiGraph is the most validated and widely used accelerometer. It has good reproducibility, validity, and feasibility when used to assess physical activity patterns in children (5,7). Accelerometer data (counts) were collected in 1-s epochs. Potential covariates such as body height and weight were measured with a portable stadiometer (seca 22551721009; Vogel & Halke GmbH & Co., Hamburg, Germany) and a digital scale (Soehnle 62882; Leifheit AG, Nassau, Germany).
The Central Committee on Research Involving Human Subjects (Dutch CCMO) offers a stepwise procedure to find out whether a study has to be reviewed according to the Dutch law "Medical Research Involving Human Subject Act (WMO)." As permitted by law, the first step in this procedure was taken care of by an internal committee at TNO Quality of Life, Zeist (The Netherlands), which concluded that the study protocol did not need to be reviewed in full by an independent medical ethics review board. Nevertheless, a written informed consent was obtained from one of the subject's parents.
The accelerometer data were downloaded to a personal computer and processed using the ActiLife GT1M 2.2.3 software program (ActiGraph). Next, the data were assigned to one of the seven physical activities according to the starting and the finishing times of each activity. For each activity, the first and the last 4 s of the acceleration signal were deleted to eliminate the transition period between activities. This time delay was determined by visual inspection of the data set. If the activity (e.g., standing) was carried out several times, the signal was cleaned for each period. The cleared accelerometer data were then used for further analyses. First, a random-effects model was used to study the differences between activity types in mean counts per second while accounting for the clustering of the measurements within subjects. Then, four ANN classification models were developed: a model based on uniaxial accelerometer data from the hip (model 1), uniaxial accelerometer data from the ankle (model 2), triaxial accelerometer data from the hip (model 3), and triaxial accelerometer data from the ankle (model 4). ANNs provide a flexible nonlinear extension of multiple logistic regression, consisting of a regression function with a set of predictors or input variables, a single hidden layer with several hidden units, and one output variable with several categories. Figure 1 shows a feed-forward ANN with three hidden units (13). For the ANN models developed in our study, the following accelerometer signal characteristics were used as input variables: 10th, 25th, 75th, and 90th percentiles; absolute deviation; coefficient of variability (i.e., the ratio of the SD and the mean), and lag-one autocorrelation (6). These characteristics were computed for each axis independently over nonoverlapping segments of 10 s. The hidden units in the model represent weighted combinations of the input variables. Weights are represented by Wij. The categories of the output variable were K = 7 types of physical activity to be classified (i.e., sitting, standing, walking, running, rope skipping, playing soccer, and cycling). The accuracy of the models was evaluated by leave-one-subject-out cross-validation (19). In this method, a set of n − 1 subjects was used as a training set, and the subject left out was used as a test set. This process was repeated for all n subjects. Feed-forward ANN models with a single hidden layer, three hidden units, and a weight decay equal to 0.01 showed the highest classification accuracy. Next, contingency tables were built to evaluate the classification errors of these models in more detail.
All statistical analyses were performed using the software package R version 2.8.0 (R Development Core Team, 2008). The classification models were developed with the function nnet (19). Both R and nnet are freely available.
While accounting for the clustering of the measurements within subjects, significant differences were found in the mean counts per second of the triaxial accelerometer worn on the hip between sitting and all other activities (with the exception of standing). The same results were found for the triaxial accelerometer worn on the ankle. These differences are depicted in Figure 2. This figure represents box plots of the mean accelerometer output across subjects for the x, y, and z axes for each activity type. These axes sense acceleration in the vertical, mediolateral, and anterior-posterior direction of the body, respectively. Whereas the upper three charts show the distribution of mean counts per second per axis of the triaxial accelerometer worn on the hip, the bottom charts show the output of the triaxial accelerometer worn on the ankle. In general, the variance within activities was higher for the accelerometer worn on the ankle than for the hip-worn accelerometer. For the latter, the highest dispersion was found for rope skipping across the three axes. For the accelerometer output of the ankle, the highest dispersion was found for running across the three axes and for cycling on the x axis. A detailed description of these results can be obtained from the authors.
Table 2 reports the accuracy of the cross-validated results of the four developed ANN models in terms of correctly classifying the seven activity types. To give an indication of the variability at each step of the cross-validation process, the 10th and 90th percentiles are also included. In general, ANN models based on hip accelerometer data performed better than models based on ankle accelerometer data. The hip models correctly classified the activities 72.4% and 76.8% of the time using uniaxial and triaxial accelerometer data, respectively, whereas the ankle models achieved a percentage of 57.2% and 68.2%. The hip models were better able to correctly classify the activities walking, rope skipping, and running, whereas the ankle models performed better when classifying sitting. A comparison of models based on data from triaxial accelerometers with models based on data from uniaxial accelerometers shows us that the percentage of correctly classified activities was higher for the triaxial models (i.e., 76.8% and 68.2%) than for the uniaxial models (i.e., 72.4% and 57.2%). In general, the models based on triaxial accelerometer data produced a better classification of the activities standing, running, rope skipping, playing soccer, and cycling than the models based on uniaxial accelerometer data. When the hip models are examined in more detail, the model based on triaxial accelerometer data especially produced a better classification of the activities playing soccer and cycling in comparison with its uniaxial counterpart (i.e., playing soccer = 82.1% vs 71.4% correctly classified, cycling = 81.1% vs 70.5%). For the ankle models, the largest difference between the triaxial model and the uniaxial model was shown for running (i.e., 67.8% vs 29.2% correctly classified), cycling (i.e., 81.6% vs 50.6%), and rope skipping (i.e., 48.9% vs 30.4%).
To evaluate the classification errors of the four models in more detail, a contingency table was built for each model representing the relationship between the observed and the predicted physical activities. Tables 3 and 4 show that the highest percentage of misclassification errors occurred for allocation to the activities standing and sitting. None of the four models discriminated between these activities; they often misclassified standing as sitting. Another physical activity that was difficult to classify was cycling, especially with models based on uniaxial accelerometer data. Whereas the model based on uniaxial accelerometer data from the hip misclassified cycling 12.3% of the time as sitting and 8.6% of the time as playing soccer, the model based on uniaxial accelerometer data from the ankle misclassified cycling 18.8% of the time as walking, 13.8% of the time as rope skipping, and 9.8% of the time as playing soccer. A high percentage of misclassification was also found for the ankle models when classifying the activities running and rope skipping. The model based on uniaxial accelerometer data from the ankle often misclassified running as rope skipping, playing soccer, or cycling. Rope skipping, on its part, was often misclassified as playing soccer or running by both the uniaxial and the triaxial ankle models.
In this study, the accuracy of four relatively simple ANN models was examined to identify children's physical activity type on the basis of uniaxial or triaxial accelerometer data from the hip or the ankle. In general, ANN models based on hip accelerometer data performed better than those based on ankle accelerometer data. This is in line with our previous work in adults evaluating two similar ANN models based on uniaxial accelerometer data from the hip and the ankle (6). In adults, the hip model correctly classified the activities sitting, standing, using the stairs, walking, and cycling 80.4% of the time, whereas the ankle model achieved a percentage of 77.7%. In both studies, the hip models performed better than the ankle models. This may be explained by the large individual differences in ankle accelerometer data. In the present study, the variation in ankle accelerometer data was much higher than the variation in hip accelerometer data. Second, our results showed that the models based on data from triaxial accelerometers performed better than their uniaxial counterparts. Uniaxial accelerometers have the limitation that different activities can produce very similar accelerometer output (17). All in all, the highest percentage of correctly classified activities in children was achieved when using triaxial accelerometer data from the hip (i.e., 76.8% correctly classified). This model performed well (>80% correctly classified) when classifying walking, running, rope skipping, playing soccer, and cycling. However, it performed worse when classifying standing (i.e., 28.3% correctly classified). Standing was often misclassified as sitting. This misclassification error may have been caused by the short duration of standing in our protocol (i.e., 1 min). Adding other sensor data, such as HR data, may also solve this problem.
Additional analyses showed that combining the triaxial accelerometer data from both the hip and the ankle did not lead to a higher percentage of correctly classified activities than when only using hip-worn triaxial accelerometer data (i.e., 74.5% vs 76.8% correctly classified). However, adding demographic input variables (i.e., body height and weight, sex, and age) into the models as suggested by Pober et al. (12) did lead to a higher accuracy. The percentage of correctly classified activities increased from 72.4% to 74.4% for model 1, from 57.2% to 60.1% for model 2, from 76.8% to 80.0% for model 3, and from 68.2% to 70.2% for model 4.
To our knowledge, this is the first study that used ANN models to identify children's physical activity type. Therefore, it is difficult to compare our results with those of previous studies. Besides different age groups, also much more complex ANN models have been used in previous studies (9,10,12,14,17). An advantage of the models proposed by Liu and Chang (10) and Khan et al. (9) is that they have used combined characteristics of the triaxial accelerometer signal as input variables in the models. These characteristics may be more suitable for distinguishing between static (e.g., sitting, standing) and dynamic activities than the isolated characteristics from each of the three axes we have used.
Including other accelerometer signal characteristics in the ANN models, such as characteristics that mark the transition between activities or characteristics representing the cyclic nature of certain types of activity (e.g., cycling), may also improve the accuracy of classification models. For these analyses, raw accelerometer data (>20 Hz) rather than filtered accelerometer data (1 Hz) may be needed. Future studies should also determine whether the accuracy of the ANN models can be further improved by including data from other sensors such as HR monitors, global positioning systems, and inclinometers. Next, accelerometer data of periods of sedentary activities, of sleeping, and when the accelerometer is not worn should be examined in more detail. At last, it is recommended to examine the accuracy of ANN models in estimating children's energy expenditure (4).
In conclusion, relatively simple ANN models can correctly classify the type of physical activities in school-aged children on the basis of single-sensor accelerometer data. The highest percentage of correctly classified activities can be achieved when using triaxial accelerometer data from the hip.
This study was funded by the Dutch Ministry of Health, Welfare, and Sport.
The authors thank Hannah Hofman for giving assistance during data collection.
There are no conflicts of interest.
The results of the present study do not constitute endorsement by the American College of Sports Medicine.
1. Bang S, Kim M, Song SK, Park SJ. Toward real time detection of the basic living activity in home using a wearable sensor and smart home sensors. Conf Proc IEEE Eng Med Biol Soc
2. Baquet G, Stratton G, Van Praagh E, Berthoin S. Improving physical activity assessment in prepubertal children with high-frequency accelerometry monitoring: a methodological issue. Prev Med
3. Bonomi AG, Goris AH, Yin B, Westerterp KR. Detection of type, duration, and intensity of physical activity using an accelerometer. Med Sci Sports Exerc
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
5. de Vries SI, Bakker I, Hopman-Rock M, Hirasing RA, van Mechelen W. Clinimetric review of motion sensors in children and adolescents. J Clin Epidemiol
6. de Vries SI, Garre FG, Engbers LH, Hildebrandt VH, van Buuren S. Evaluation of neural networks to identify types of activity using accelerometers. Med Sci Sports Exerc
7. de Vries SI, van Hirtum HW, Bakker I, Hopman-Rock M, Hirasing RA, Van Mechelen W. Validity
and reproducibility of motion sensors in youth: a systematic update. Med Sci Sports Exerc
8. Esliger DW, Copeland JL, Barnes JD, Tremblay MS. Standardizing and optimizing the use of accelerometer data for free-living physical activity monitoring. J Phys Act Health
9. Khan AM, Lee YK, Kim TS. Accelerometer signal-based human activity recognition using augmented autoregressive model coefficients and artificial neural nets. Conf Proc IEEE Eng Med Biol Soc
10. Liu SH, Chang YJ. Using accelerometers for physical actions recognition by a neural fuzzy network. Telemed J E Health
11. Long X, Yin B, Aarts RM. Single-accelerometer-based daily physical activity classification. Conf Proc IEEE Eng Med Biol Soc
12. Pober DM, Staudenmayer J, Raphael C, Freedson PS. Development of novel techniques to classify physical activity mode using accelerometers. Med Sci Sports Exerc
13. Ripley BD. Pattern Recognition and Neural Networks
. Cambridge (UK): Cambridge University Press; 1996. p. 143-5.
14. Rothney MP, Neumann M, Béziat A, Chen KY. An artificial neural network model of energy expenditure using nonintegrated acceleration signals. J Appl Physiol
15. Rowland TW. Promoting physical activity for children's health: rationale and strategies. Sports Med
16. Salmon J, Timperio A. Prevalence, trends and environmental influences on child and youth physical activity. Med Sport Sci
17. 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
18. Twisk JW. Physical activity guidelines for children and adolescents: a critical review. Sports Med
19. Venables WN, Ripley BD. Modern Applied Statistics with S
. New York (NY): Springer-Verlag; 2002. p. 342-4.