Machine-Learning Estimation of Body Posture and Physical Activity by Wearable Acceleration and Heartbeat Sensors

Signal & Image Processing: An International Journal (SIPIJ) Vol.10, No.3, June 2019
DOI: 10.5121/sipij.2019.10301 1
MACHINE-LEARNING ESTIMATION OF BODY
POSTURE AND PHYSICAL ACTIVITY BY WEARABLE
ACCELERATION AND HEARTBEAT SENSORS
Yutaka Yoshida2
, Emi Yuda3, 1
, Kento Yamamoto4
, Yutaka Miura5
, and
Junichiro Hayano1
1
Department of Medical Education, Nagoya City University Graduate School of Medical
Science, Nagoya, Japan
2
Nagoya City University Graduate School of Design and Architecture, Nagoya, Japan
3
Department of Electrical Engineering, Tohoku University Graduate School of
Engineering, Sendai, Japan
4
University of Tsukuba Graduate School of Comprehensive Human Sciences,
Tsukuba, Japan
5
Department of Nutrition, Shigakkan University, Obu, Japan
ABSTRACT
We aimed to develop the method for estimating body posture and physical activity by acceleration signals
from a Holter electrocardiographic (ECG) recorder with built-in accelerometer. In healthy young subjects,
triaxial-acceleration and ECG signal were recorded with the Holter ECG recorder attached on their chest
wall. During the recording, they randomly took eight postures, including supine, prone, left and right
recumbent, standing, sitting in a reclining chair, sitting in chairs with and without backrest, and performed
slow walking and fast walking. Machine learning (Random Forest) was performed on acceleration and
ECG variables. The best discrimination model was obtained when the maximum values and standard
deviations of accelerations in three axes and mean R-R interval were used as feature values. The overall
discrimination accuracy was 79.2% (62.6-90.9%). Supine, prone, left recumbent, and slow and fast walk
were discriminated with >80% accuracy, although sitting and standing positions were not discriminated by
this method.
KEYWORDS
Accelerometer, Holter ECG, Posture, Activity, Machine learning, Random Forest, R-R interval
1. INTRODUCTION
The importance of physical activity for health maintenance and promotion is gathering increasing
attention. Sedentary behaviour is known to be associated with the incidence of malignant
arrhythmias, and there is accumulating evidence to support that the maintenance and increase of
physical activity decrease the risk of sudden cardiac death [1]. Increased physical activity by
exercise can be expected to be effective in lowering the functional level of daily life and in
preventing life-style related diseases including diabetes, dyslipidaemias, metabolic syndrome,
heart diseases, and malignancies [2]. The World Health Organization's Global Recommendation
for Physical Activity [1] states an increase in total step counts in daily life as a goal in the field of
daily physical activity and exercise. It says that an increase of 1500 steps a day accounts for 2%
reduction in the incidence and mortality by non-communicable diseases and 1.5-mm Hg reduction
in blood pressure.

2
Even though these required physical activity criteria have been proposed based on scientific
evidence, the information used to assess physical activity is often depends on that obtained by
conventional methods such as a pedometer, in which information such as the type and frequency
of postural changes and the speed of action and movement is ignored. As a result, the adjustment
for the effects of age, gender, weight, and height on the assessment of physical activity, if
implemented, become less rigorous.
To overcome this problem, several studies proposed acceleration and other sensors [3-10] and
algorithms including machine learning for the precise estimation of body posture and the type and
level of physical activity [6,7,10]. The reported performance, however, differed with the type,
number, and placement of sensors, signal sampling frequency, analysis window size, machine
learning technique, and measurement setting, i.e., laboratory or free-living conditions [11]. In this
study, we investigated the posture and activity estimation performance of a built-in accelerometer
of a chest-mounted Holter electrocardiographic (ECG) recorder widely used for clinical purposes
in combination with the Random Forest machine learning technique.
2. EXPERIMENT METHODS
2.1. Subjects
We studied 11 healthy subjects (2 men, 9 women, 22 ± 1 yr) who gave a written informed consent
to participate this study. The protocol of the present study was approved by the Institutional
Review Board of Nagoya City University Graduate School of Medical Sciences and Nagoya City
University Hospital (approval number, 60-18-0211).
2.2. Measurement device
Holter ECG recorders with a built-in 3-axis acceleration sensor (Cardy 303 pico+, Suzuken Co.,
Ltd; size, W28×D42×H9 mm, weight, 13 g) were used for measuring acceleration and ECG
signals. The sampling frequency of the ECG was 128 Hz. The accelerometer recorded
acceleration from left to right, cranio-caudal, and postero-anterior directions as the X, Y, and Z
values, respectively, at a frequency of 31.25 Hz.
2.3. Experimental protocol
Experiments were conducted in an air-conditioned laboratory at 24 ± 2 ºC. The Holter ECG
recorder and electrodes were attached on subjects’ chest wall at 09:30. Subjects were instructed to
take supine, prone, right and left recumbent, and standing positions, to sit in three different chairs
(a recliner and chairs with and without a backrest), and to walk slowly and walk fast, in a
randomized order. Subjects maintained each posture and activity for 110 s, and then moved to the
next posture or activity within 10 s.
2.4. Data analysis
2.4.1. Analysis of 3-axis acceleration time series and RRI time series
The acceleration signals in X, Y, and Z axes at 31.25 Hz were under-sampled at 2 Hz per axis,
and average, median, maximum, minimum, and standard deviation (SD) were calculated for
every 10 s for each axis. Because the acceleration data were unstable during the beginning of each
posture and activity, the first 20 s of each posture was removed, resulting in nine 10-s segments
per posture/activity.

3
From the ECG signal, all R waves (sharp deflections corresponding to the ventricular electrical
excitation) were detected and an R-R interval time series were obtained. R-R intervals consisting
of consecutive sinus rhythms were extracted, resampled at 2 Hz using a step function, and
averaged over every 10 s.
2.4.2. Feature Selection
As the feature datasets of acceleration variables for machine learning, we examined (1) the
averages and SDs of 3 axes, (2) the medians and SDs of 3 axes, (3) the maximums and SDs of 3
axes, (4) the minimums and SDs, and (5) all indices of 3 axes. In addition, the average of the R-R
interval was added to each of the acceleration datasets, yielding 10 feature datasets. The number
of data for one posture was nine (90 seconds) per variable. The data for sitting positions in
different chairs were put together and nine time points were randomly selected from the pooled
dataset.
2.4.3. Machine learning of classification and verification
We used Random Forest for the machine learning classifier. To reduce the bias of the learning
data, learning and verification were performed using k-fold cross validation method (k = 11).
There was no overlap between the training and the test data. These processes were performed
with Python (3.6) obtained from the open data science platform Anaconda (ANACONDA,
https://www.anaconda.com/).
Verification of discrimination was performed by calculating the true positive (TP) and false
negative (FN) for each posture from each subject, and then adding the TP and the FN from all
subjects. The recall rate defined as equation (i) was calculated for each posture. For example, in
the discrimination of the supine position, the TP was the number of successful discriminations of
supine and FN was the number of unsuccessful discriminations of supine.
Fig. 1. Time series of R-R interval and 3-axis acceleration during experiment.
R-Rc = Right recumbent, L-Rc = Left recumbent, Sitting rec1 = sitting in a recliner with knee
extending, Sitting rec2 = sitting in a recliner with knee bending, Sitting back = sitting in a chair
with a backrest.

4
Table 1. Recall ratio for posture with feature dataset of acceleration
Posture/activity
Acceleration
Average SD Median SD Max SD Min SD All indices
Spine 0.869 0.859 0.859 0.859 0.869
Prone 0.798 0.788 0.717 0.778 0.657
R-recumbent 0.727 0.707 0.747 0.747 0.727
L-recumbent 0.909 0.909 0.909 0.919 0.909
Sitting 0.636 0.646 0.616 0.586 0.747
Standing 0.626 0.576 0.616 0.606 0.535
Slow walk 0.677 0.677 0.768 0.727 0.687
Fast walk 0.697 0.747 0.758 0.677 0.758
Accuracy 0.742 0.739 0.749 0.737 0.736
SD 0.106 0.111 0.103 0.116 0.118
Accuracy was calculated as the mean value of recalls for all postures.
SD = standard deviation
Table 2. Recall ratio for posture with feature dataset of acceleration and R-R interval
Posture/activity
Acceleration
Average SD Median SD Max SD Min SD All indices
Spine 0.869 0.838 0.899 0.879 0.859
Prone 0.818 0.859 0.859 0.859 0.636
R-recumbent 0.747 0.727 0.737 0.737 0.747
L-recumbent 0.909 0.859 0.909 0.919 0.909
Sitting 0.758 0.667 0.717 0.667 0.768
Standing 0.525 0.586 0.626 0.535 0.434
Slow walk 0.657 0.606 0.818 0.687 0.828
Fast walk 0.717 0.737 0.768 0.758 0.737
Accuracy 0.750 0.735 0.792 0.755 0.740
SD 0.122 0.110 0.098 0.128 0.149
Accuracy was calculated as the mean value of recalls for all postures.
Recall＝TP／ (TP+FN) ・・・ (i)
The number of the decision tree and the maximum depth of the tree of the Random Forest were 3
and 20, respectively, and the maximum value of the recall was calculated by the grid search
function in this range.

5
3. RESULTS
Fig. 1 shows the examples of R-R interval and 3-axis acceleration time series during the
experiment. The acceleration signals changed according to the changes in the relative direction of
gravitation vector with postures. In addition, the acceleration signals, particularly in Y-axis
reflecting cranio-caudal direction, showed fluctuations during walking and its amplitude was
greater for fast walk than slow walk.
Table 3. Confusion matrix of feature dataset with the best classification performance
Posture/activity
Estimated posture/activity
Segment Recall
Su Pr Rr Lr Si St Sw Fw
Actual
Su 89 0 7 0 0 0 1 2 99 0.899
Pr 1 85 4 7 1 0 1 0 99 0.859
Rr 7 15 73 1 3 0 0 0 99 0.737
Lr 0 2 7 90 0 0 0 0 99 0.909
Si 2 0 0 0 71 26 0 0 99 0.717
St 0 0 0 0 34 62 3 0 99 0.626
Sw 0 0 0 0 0 1 81 17 99 0.818
Fw 0 0 0 0 0 0 23 76 99 0.768
Su = supine, Pr = prone, Rr = right recumbent, Lr = left recumbent, Si = sitting, St = standing, Sw
= slow walk, Fw fast walk.
Table 4. Gross confusion matrix for coarse-grained posture/activity
Posture/activity
Estimated
Segment Recall
Lying Sitting Standing Walking
Actual
Lying 388 4 0 4 396 0.980
Sitting 2 71 26 0 99 0.717
Standing 0 34 62 3 99 0.626
Walking 0 0 1 197 198 0.999
Segment 390 109 89 204 792
Specificity 0.995 0.945 0.961 0.986 Accuracy
Positive predictive value 0.995 0.651 0.697 0.966 0.907
Data are classification results by feature dataset with the maximum and SD of acceleration and
mean R-R interval.
Tables 1 and 2 shows the recall rate for the posture/activity by each feature dataset. The feature
dataset with the highest mean recall rate was a combination of the maximum and SD of
acceleration and mean R-R interval (mean recall rate ± SD, 0.792 ± 0.098). Supine, prone, left
recumbent, and slow walk were discriminated by recall rates >0.80, while the recall rate for
standing was 0.626.
Tables 3 and 4 shows the confusion matrix for the feature dataset with the best performance
(maximum and SD of acceleration and mean R-R interval). When the postures and activities were
coarse-grained into lying, sitting, standing, and walking, the overall accuracy of classification was
0.907. The recall ratio, specificity, and positive predictive value for classifications of lying and
walking were >0.95.

6
Fig. 2. Histogram of the maximum and SD of acceleration and R-R interval for sitting and standing
positions.
Blue: sitting, Orange: standing.
However, the discrimination between sitting and standing were less accurate. Although out of 99
segments in sitting 71 segments were correctly discriminated as sitting, 26 segments were
classified as standing, and 2 as prone position, indicating that 26% of sitting was estimated as
standing. Similarly, out of 99 segments in standing 62 segments were correctly discriminated as
standing, but 34 segments were classified as sitting and 3 as slow walking, indicating that
standing was estimated as sitting at 34% probability. Fig. 3 is a histogram of the maximum and
SD of acceleration and RRI in the sitting and standing postures. Both show the similar patterns of
distribution.

7
4. DISCUSSIONS
In the present study, we examined the usefulness of built-in triaxial accelerometer of a chest-
mounted Holter ECG recorder in combination with machine learning for estimating body posture
and activity. All possible combinations of variables derived from triaxial acceleration signals
every 10 s were used as feature dataset and discriminant models were generated by the Random
Forest algorithm. The best classification performance was obtained by the feature dataset
consisting of the maximum and SD of acceleration. For all feature datasets of acceleration
variables, addition of mean R-R interval improved the classification performance. By the model
using the maximum and SD of acceleration and mean R-R interval classified lying, sitting,
standing, and walking with an overall accuracy of 0.907. The model showed, however, lesser
discrimination performances for sitting and standing than those for lying and walking. While the
recall ratios for lying and walking were 0.980 and 0.999, the ratios for sitting and standing were
0.717 and 0.626, respectively. Our observations show both the strengths and weaknesses of
posture/activity estimation by acceleration signals and seem to suggest hints for improving the
estimation.
There are many studies of the sensors and signal processing algorithms to discriminate body
postures [3-10]. They reported different performance depending on factors related to sensors,
signal processing, data analysis including machine learning techniques, and measurement
settings. As expected, using multiple sensors gives better performance than a single sensor. Yeoh
et al. [3] used 3 accelerometers at the waist and both thighs and achieved an overall accuracy of
100% for classification of lying, sitting, standing, and walking and an overall mean-square error
of 1.76 km/h for walking and running speeds. As to the type of sensor, Godfrey et al. [4] used
chest mounted tri-axial accelerometer sensitive to both static and dynamic acceleration and
reported sensitivity and specificity of >0.83 to detect activity types and postural transitions
between sitting and standing using a sophisticated algorithm of Velocity Estimate and Scalar
Product Activity. As to sensor placement, Fulk et al. [5] used acceleration and pressure sensors
built into the shoe and successfully distinguished between sitting and standing with an accuracy
of 82-100% in patients after stroke. Also, Doulah et al. [7] reported the detection of sit-to-stand
posture transitions by orthosis-mounted sensors and Fanchamps et al. [9] used a acceleration
sensor attached to the thigh in post-stroke patients. To discriminate activity types and postures,
machine learning-based approaches are employed in most of recent studies, such as artificial
neural network [7], support vector machines [5], Random Forest [10], Naïve Bayes, and K-
nearest neighbours [11]. Finally, as to measurement setting, the reported performance of activity
and posture discriminations obtained in free-living settings were generally lower than those
obtained in laboratory settings [9,11]. In the present study, we used a simple tri-axial static
acceleration sensor built in a chest-mounted Holter ECG recorder, measured data in a laboratory
setting, and generated the discriminant model for posture and activity by Random Forest
technique. Thus, it is difficult to compare the present results with those of earlier studies, but the
discriminant accuracies were comparable to them except that between sitting and standing.
The low discrimination accuracy between sitting and standing may be attributable to the use of
chest-mounted accelerometer that was less sensitive to dynamic acceleration and insensitive to
the changes in angular velocity. Although the gravity vector is almost vertical in the sitting
position, it may shift backward to some extent depending on the type of chair (more precisely, on
the angle at which the body leans on the backrest). In contrast, the gravity vector in the standing
posture is vertical theoretically; however, since the Holter ECG recorder is worn on the
inclination of the chest, some backward inclination occurs. During standing acceleration may
fluctuate slightly due to body sway, while it may more stable during sitting. In this experiment,
however, the fluctuation of acceleration signal was observed also during sitting (Fig. 3). In

8
addition, the backward movement of the vertical acceleration vector may also occur due to
walking and other forward movements, further complicating the determination. To overcome this
problem, not only static features of acceleration such as mean, SD but also dynamic features of
acceleration such as the pattern of changes seem necessary.
To our knowledge, there was no study to report posture discrimination using both acceleration
and R-R interval. Although the discrimination accuracy differed depending on the selection of the
feature dataset, the discrimination accuracy was consistently improved by combining 3-axis
acceleration and R-R interval. Since R-R interval is expected to shorten with standing [12], mean
R-R interval may be useful feature to discriminate sitting and standing. This suggests possibility
of using the indices of heart rate variability as the additional feature to discriminate
posture/activity. Particularly, low frequency to high frequency power ratio may be useful for the
feature to detect standing [13,14].
5. CONCLUSIONS
We examined the posture and activity discrimination performance of signals from a built-in tri-
axial acceleration sensor in a chest-mounted Holter ECG recorder using Random Forest machine
learning technique. We observed acceptable discriminate accuracy of activities and postures
except for that between sitting and standing. Additionally, we found that the inclusion of R-R
interval data as a feature improves the discriminatory accuracy. Because Holter ECG recorders
with built-in accelerometer are widely used clinically, our findings seem to provide useful insight
into the benefits and limitations of the clinical use of the accelerometer signal.
REFERENCES
[1] World Health Organization, Global recommendations on Physical Activity for Health. Geneva: World
Health Organization; 2010.
[2] Sofi, F., Valecchi, D., Bacci, D., Abbate, R., Gensini, G. F., Casini, A., Macchi, C. (2011) "Physical
activity and risk of cognitive decline: a meta-analysis of prospective studies", J. Intern. Med., Vol.
269, No. 1, 107-117.
[3] Yeoh, W. S., Pek, I., Yong, Y. H., Chen, X., Waluyo, A. B. (2008) "Ambulatory monitoring of human
posture and walking speed using wearable accelerometer sensors", Conf Proc IEEE Eng Med Biol
Soc, Vol. 2008, No., 5184-5187.
[4] Godfrey, A., Bourke, A. K., Olaighin, G. M., van de Ven, P., Nelson, J. (2011) "Activity
classification using a single chest mounted tri-axial accelerometer", Med. Eng. Phys., Vol. 33, No. 9,
1127-1135.
[5] Fulk, G. D., Sazonov, E. (2011) "Using sensors to measure activity in people with stroke", Top Stroke
Rehabil, Vol. 18, No. 6, 746-757.
[6] Palmerini, L., Rocchi, L., Mellone, S., Valzania, F., Chiari, L. (2011) "Feature selection for
accelerometer-based posture analysis in Parkinson's disease", IEEE Trans Inf Technol Biomed, Vol.
15, No. 3, 481-490.
[7] Doulah, A., Shen, X., Sazonov, E. (2017) "Early Detection of the Initiation of Sit-to-Stand Posture
Transitions Using Orthosis-Mounted Sensors", Sensors, Vol. 17, No. 12.
[8] Vaha-Ypya, H., Husu, P., Suni, J., Vasankari, T., Sievanen, H. (2018) "Reliable recognition of lying,
sitting, and standing with a hip-worn accelerometer", Scand. J. Med. Sci. Sports, Vol. 28, No. 3,
1092-1102.

9
[9] Fanchamps, M. H. J., Horemans, H. L. D., Ribbers, G. M., Stam, H. J., Bussmann, J. B. J. (2018)
"The Accuracy of the Detection of Body Postures and Movements Using a Physical Activity Monitor
in People after a Stroke", Sensors, Vol. 18, No. 7.
[10] Kerr, J., Carlson, J., Godbole, S., Cadmus-Bertram, L., Bellettiere, J., Hartman, S. (2018) "Improving
Hip-Worn Accelerometer Estimates of Sitting Using Machine Learning Methods", Med. Sci. Sports
Exerc., Vol. 50, No. 7, 1518-1524.
[11] Farrahi, V., Niemela, M., Kangas, M., Korpelainen, R., Jamsa, T. (2019) "Calibration and validation
of accelerometer-based activity monitors: A systematic review of machine-learning approaches", Gait
Posture, Vol. 68, No., 285-299.
[12] Olufsen, M. S., Tran, H. T., Ottesen, J. T., Research Experiences for Undergraduates, P., Lipsitz, L.
A., Novak, V. (2006) "Modeling baroreflex regulation of heart rate during orthostatic stress", Am J
Physiol Regul Integr Comp Physiol, Vol. 291, No. 5, R1355-1368.
[13] Hayano, J., Mukai, S., Fukuta, H., Sakata, S., Ohte, N., Kimura, G. (2001) "Postural response of low-
frequency component of heart rate variability is an increased risk for mortality in patients with
coronary artery disease", Chest, Vol. 120, No., 1942-1952.
[14] Yoshida, Y., Furukawa, Y., Ogasawara, H., Yuda, E., Hayano, J. Longer lying position causes lower
LF/HF of heart rate variability during ambulatory monitoring. Paper presented at: 2016 IEEE 5th
Global Conference on Consumer Electronics (GCCE); 11-14 Oct 2016, 2016; Kyoto, Japan.
AUTHORS
Yutaka Yoshida studied the business administration and computer science at Aichi
Institute of Technology and received Ph.D. degree in 2008. He was a project researcher
at Knowledge Hub of Aichi from 2011 to 2015 and was a researcher at Nagoya City
University Graduate School of Medical Sciences from 2016 to 2017. Since 2018, he has
been a researcher at Nagoya City University Graduate School of Design and
Architecture. His specialized field is biological information engineering, signal
processing and ergonomics. He received the paper award at the Japan Society of
Neurovegetative Research in 2005 and 2007.
Emi Yuda was born in Tokyo, Japan in 1980. She studied informatics at M.V.
Lomonosov Moscow State University until 2003 and then received M.S. degree from
Tsukuba University, Japan. She received Ph.D. from Nihon University in 2019. From
2013 to 2014 she was a research assistant at Santa Monica College in California, USA.
From 2015 to 2019, she was a NEDO project researcher in Nagoya City University
Graduate School of Medical Sciences. Since 2019, she has been an assistant professor in
Tohoku University Graduate School of Engineering. Her currently research is Medical
Informatics and Data Science. She has many achievements in field of Informatics and Big Data.
Junichiro Hayano graduated Nagoya City University Medical School, Nagoya, Japan and
received M.D. degree in 1980. From 1981 to 1983, he received residency trainings of
psychosomatic medicine in Kyushu University School of Medicine, Fukuoka, Japan. He
obtained Ph.D. degree (Dr. of Medical Science) in 1988 from Nagoya City University
Graduate School of Medical Sciences. From 1990 to 1991, he was working as a visiting
associate at the Behavioral Medicine Research Center, Duke University Medical Center,
Durham, NC, USA. In 1984, he got a faculty position at Nagoya City University Medical
School and has been a Professor of Medicine at Nagoya City University Graduate School of Medical
Sciences since 2003. His current interests are applications of dynamic electrocardiography and bio-signal
processing to health sciences.

Machine-Learning Estimation of Body Posture and Physical Activity by Wearable Acceleration and Heartbeat Sensors

More Related Content

What's hot (19)

Similar to Machine-Learning Estimation of Body Posture and Physical Activity by Wearable Acceleration and Heartbeat Sensors (20)

Recently uploaded (20)

Machine-Learning Estimation of Body Posture and Physical Activity by Wearable Acceleration and Heartbeat Sensors