PreprintPDF Available

Accelerometry Data in Health Research: Challenges and Opportunities: Review and Examples

March 2018

DOI:10.1101/276154

License
CC BY-NC-ND 4.0

Authors:

Marta Karas

Harvard University

Jiawei Bai

Johns Hopkins Bloomberg School of Public Health

Marcin Strączkiewicz

Harvard University

Show all 9 authorsHide

Wearable accelerometers provide detailed, objective, and continuous measurements of physical activity (PA). Recent advances in technology and the decreasing cost of wearable devices led to an explosion in the popularity of wearable technology in health research. An ever-increasing number of studies collect high-throughput, sub-second level raw acceleration data. In this paper, we discuss problems related to the collection and analysis of raw accelerometry data and refer to published solutions. In particular, we describe the size and complexity of the data, the within- and between-subject variability, and the effects of sensor location on the body. We also discuss challenges related to sampling frequency, device calibration, data labeling, and multiple PA monitors synchronization. We illustrate these points using the Developmental Epidemiological Cohort Study (DECOS), which collected raw accelerometry data on individuals both in a controlled and the free-living environment.

Acceleration values from three orthogonal axes of an accelerometer located on the left wrist. Each axis data is shown in a different color. The top panel displays 24 hours of data collected between 12AM and 12AM. The middle panel displays a one hour interval from 8 : 40AM to 9 : 40AM (indicated in the top panel as a dashed-line rectangle). The bottom panel displays a one minute interval from 8 : 51AM to 8 : 52AM marked as a dashed-line rectangle in the middle panel. The signal was acquired at a sampling frequency of fs = 80Hz.

…

Data recorded by an accelerometer located on the left wrist while walking (left column) and getting dressed (right column), for two individuals (top and bottom row). Each axis is shown in a different color.

…

Acceleration from three orthogonal axes of an accelerometer located on the hip (left column) and left wrist (right column), while dealing cards (top row), getting dressed (middle row) and walking (bottom row). Each axis data are shown in a different color.

…

Acceleration values from three orthogonal axes of an accelerometer located on the left wrist, collected during two walking tasks performed by the same individual. Each axis data is shown in a different color. The upper panel corresponds to walking with both hands moving naturally, whereas the bottom panel corresponds to walking with arms crossed on the chest.

…

Acceleration data from three orthogonal axes at the hip collected around the time when a participant performed a 400-meter-walk activity. The dashed-line red box indicates the portion of the 400-meter-walk period identified by a human observer.

…

Figures - uploaded by Vadim Zipunnikov

Content may be subject to copyright.

Content uploaded by Vadim Zipunnikov

Content may be subject to copyright.

Noname manuscript No.

(will be inserted by the editor)

Accelerometry data in health research: challenges

and opportunities

Review and examples

Marta Karas ·Jiawei Bai ·Marcin

Strączkiewicz ·Jaroslaw Harezlak ·

Nancy W. Glynn ·Tamara Harris ·

Vadim Zipunnikov ·Ciprian

Crainiceanu ·Jacek K. Urbanek

Received: date / Accepted: date

Marta Karas

Department of Biostatistics

Bloomberg School of Public Health

Johns Hopkins University

Tel.: +1-317-665-4551

E-mail: mkaras@jhu.edu

Jiawei Bai

Department of Biostatistics

Bloomberg School of Public Health

Johns Hopkins University

Marcin Strączkiewicz

Department of Epidemiology and Biostatistics

School of Public Health

Indiana University Bloomington

Jaroslaw Harezlak

Department of Epidemiology and Biostatistics

School of Public Health

Indiana University Bloomington

Nancy W. Glynn

Center for Aging and Population Health

Department of Epidemiology

Graduate School of Public Health

University of Pittsburgh

Tamara Harris

Laboratory of Epidemiology Demography, and Biometry

National Institute on Aging

Vadim Zipunnikov

Department of Biostatistics

Bloomberg School of Public Health

Johns Hopkins University

Ciprian Crainiceanu

Department of Biostatistics

Bloomberg School of Public Health

.CC-BY-NC-ND 4.0 International licensenot peer-reviewed) is the author/funder. It is made available under a The copyright holder for this preprint (which was. http://dx.doi.org/10.1101/276154doi: bioRxiv preprint first posted online Mar. 5, 2018;

2 Marta Karas et al.

Abstract Wearable accelerometers provide detailed, objective, and continu-

ous measurements of physical activity (PA). Recent advances in technology

and the decreasing cost of wearable devices led to an explosion in the popula-

rity of wearable technology in health research. An ever increasing number of

studies collect high-throughput, sub-second level raw acceleration data. In this

paper we discuss problems related to the collection and analysis of raw acce-

lerometry data and provide insights into potential solutions. In particular, we

describe the size and complexity of the data, the within- and between-subject

variability and the eﬀects of sensor location on the body. We also provide a

short tutorial for dealing with sampling frequency, device calibration, data

labeling and multiple PA monitors synchronization. We illustrate these po-

ints using the Developmental Epidemiological Cohort Study (DECOS), which

collected raw accelerometry data on individuals both in a controlled and the

free-living environment.

Keywords Wearable computing ·Accelerometry ·Wearable accelerometers ·

Physical activity ·Accelerometers

1 Introduction

Wearable physical activity (PA) monitors provide detailed, continuous, and

objective measurements of individual PA in the free living environment. They

can complement or completely replace current subjective measurements col-

lected via questionnaires. Recent advances in technology and the decreasing

cost of wearable devices led to an explosion in the popularity of wearable tech-

nology in health research. Here we argue that, just like any new measurement

used in health science, there is a need to understand, reproduce, and com-

municate the measurements produced by these new devices. This can lead to

improved design of experiments, higher quality of the acquired data, and more

generalizable results. At the core of all modern PA monitors there is a small

accelerometer, a Microelectromechanical system (MEMS) that measures ac-

celerations relative to the Earth’s gravitational ﬁeld. Hence, PA monitors are

often referred to as wearable accelerometers. The output of these devices is a

three-dimensional time series of accelerations expressed in gravitational units

in the frame of reference of the device. More clearly, the device has its own fra-

me of reference up-down, left-right, backward-forward. This frame is typically

diﬀerent from and can change with the frame of reference of the person who

wears the device or who observes the experiment. These raw data produced by

Johns Hopkins University

Jacek K. Urbanek

Center on Aging and Health

Division of Geriatric Medicine and Gerontology

Department of Medicine

School of Medicine

Johns Hopkins University

E-mail: jurbane2@jhu.edu

Accelerometry data in health research: challenges and opportunities 3

accelerometers are transformed using various algorithms into PA summaries,

which have diﬀerent labels (e.g. steps, calories, activity counts) and can be

aggregated at diﬀerent temporal resolutions (e.g. minutes, hours, or days).

Due to battery and memory limitations, PA monitors used to return only

aggregated minute-level data in the form of proprietary activity counts (AC)

(Chen et al, 2012). The deﬁnition of ACs varies between- and within-device

manufacturers, across time, body location and between studies. Despite the-

se initial problems, they have been used eﬀectively as a relative measure of

PA within the same study, especially when the same devices and software we-

re used and devices were calibrated (Matthews et al, 2008; Healy et al, 2011;

Schrack et al, 2014; Xiao et al, 2015). As battery and memory limitations have

been mitigated, it has become possible to collect and store high-throughput,

three-axial, sub-second level acceleration data, ranging between 10 and 200

observations per second. PA monitors have also become increasingly sophi-

sticated and are now routinely equipped with a selection of supplementary

sensors including gyroscopes, thermometers, inclinometers, pulsometers, light

intensity and skin conductance sensors. Data collected by these supplementary

sensors are beyond the scope of this paper.

The collection of raw accelerometry data opens a spectrum of new scientiﬁc

and analytic problems. For example, researchers do not need to rely on proprie-

tary aggregated measures and can use well-deﬁned, open-source, reproducible

summaries of the data. This allows to compare and combine studies that col-

lect raw accelerometry data at the same location on the body and provides

explicit measures of activity on a recognized measurement scale. For example,

UKBiobank (Doherty et al, 2017) uses the magnitude of the 3-dimensional

vector of acceleration summarized in 5-second intervals, whereas the Women’s

Health Initiative (WHI) study has explored 1-second summaries based on stan-

dard deviations of accelerometry along each axis (Bai et al, 2016). Raw and

summarized data have also been used for recognition of activity types. A com-

mon approach is to derive accelerometry data features in a particular window

and use supervised classiﬁcation approaches to predict activity type (Pober

et al, 2006; Staudenmayer et al, 2009; Attal et al, 2015). Extensive reviews of

classiﬁcation techniques for activity recognition from accelerometry data are

provided by Bao and Intille (2004) and Preece et al (2009). Dictionary learning

based on raw accelerometry Bai et al (2012); He et al (2014); Xiao et al (2016)

has also been proposed in the context of ﬁne-resolution movement prediction.

The increased granularity of the sub-second level data may contain impor-

tant additional information, but it also creates new challenges. Indeed, the

volume and structure of the data are much more complex for the raw da-

ta, especially when it is recorded for weeks at a time in free-living settings.

In this paper we discuss problems related to the collection and analysis of

sub-second level accelerometry data. To illustrate these problems, we use data

collected as a part of Developmental Epidemiological Cohort Study (DECOS)

in a controlled and in the free-living environment.

Some of these problems have been well documented in the literature. For

example, Trost et al (2005) highlighted problems associated with device se-

4 Marta Karas et al.

lection, placement of accelerometers, epoch length, and compliance enhancing

strategies. Schrack et al (2016) discussed the limitations of using uncalibrated,

population-level data extraction algorithms in older adults. Staudenmayer et al

(2012) and Troiano et al (2014) noted that the diversity of measurements and

analytic methods for accelerometer data makes it diﬃcult to compare results

across diﬀerent studies, and advocate for standardization of measurement and

pre-processing pipelines across studies.

Currently, there is no universally accepted and standardized approach for

measuring PA using wearable accelerometers in health research. However, so-

me excellent guidelines and standardized protocols have been published. For

example, Matthews et al (2012) compiled a list of best practices for making

decisions about important choices, such as the number of monitors needed, de-

vice placement, device initialization, device tracking, and data collection. They

also provide guidance on how to report the use of PA monitors in population-

based studies. Freedson et al (2012) provided recommendations for the use

of wearable monitors for assessing PA for researchers, end users, as well as

developers of activity monitors. The authors also provide guidelines for sensor

output calibration and validation and discuss necessary steps for maximizing

the generalizability of the data analysis.

Unfortunately, harmonization of PA data between existing studies is often

impossible due to the diﬀerent formulation and interpretation of activity co-

unts, body placement, and lack of universally accepted measurement. For

example, step counts produced by Fitbit cannot be compared to activity co-

unts produced by ActiGraph, as they are not even on the same scale. In fact,

is has been shown that step counts might diﬀer substantially when measured

by diﬀerent devices (Storti et al, 2008; Fortune et al, 2014). Activity counts

measured by the same device may also be diﬀerent depending on the sensor

location (Fairclough et al, 2016). For example, a device located on the hip or

ankle is sensitive to ambulation, whereas a device located on the wrist will

detect both ambulation and hand movements. Moreover, some activity counts

may change substantially with the sampling frequency (Brønd and Arvidsson,

2016) even for the same device placed at a particular body location. These

problems are not likely to be resolved as long as summary data obtained via

proprietary manufacturer algorithms continue to be used. The best strategy

is to go back to the raw data and construct open-source data pre-processing

approaches that become increasingly accepted by the community. This, unfor-

tunately, is not a panacea, as pre-processing pipelines need to take into account

problems associated with device calibration, missing data, data size and com-

plexity, and measurement translation and communication.

To provide a concrete illustration of these problems and some potential

solutions, we use the Developmental Epidemiological Cohort Study (DECOS)

data, which is described in Sections 2.1- 2.2. Next we introduce the nota-

tion and statistical methods used to pre-process and summarize the data in

Section 2.3. In Section 3 we illustrate issues related to the analysis and inter-

pretation of raw accelerometry data and review some published approaches

designed to handle the complexity and heterogeneity of accelerometry data

Accelerometry data in health research: challenges and opportunities 5

collected in the free-living environment. In Section 4 we summarize the ideas

and discuss their implication for health studies using raw accelerometry data.

2 Methods

2.1 Study participants

Forty-nine community-dwelling older adults were recruited from the Pitts-

burgh, Pennsylvania area for the Developmental Epidemiologic Cohort Study

(DECOS), part of the National Institute on Aging (NIA) Aging Research Eva-

luating Accelerometry (AREA) project (Lange-Maia et al, 2015). DECOS is

a cross-sectional study designed to examine the impact of accelerometry we-

ar location on PA and sedentary behavior assessment among healthy older

adults. Individuals were excluded from DECOS if they suﬀered from any of

the following conditions: hip fracture, stroke in the past 12 months, cerebral

hemorrhage in the past 6 months, heart attack, angioplasty, heart surgery in

the past 3 months, chest pain during walking in the past 30 days, current treat-

ment for shortness of breath or a lung condition, usual aching, stiﬀness, or pain

in their lower limbs and joints and bilateral diﬃculty bending or straightening

the knees fully.

2.2 Data collection

Participants were equipped with three tri-axial wearable PA monitors (Acti-

Graph GT3x+) that collected raw accelerometry data at a sampling frequency

of 80Hz (80 observations per second for each axis). Monitors were located on

the hip using an elastic belt and on both wrists using watch straps. During the

“in-the-lab” phase of the experiment all participants were asked to perform

a series of physical tasks including: lying still, standing still, washing dishes,

sitting still, dough kneading, dressing, folding towels, vacuuming, shopping,

writing, dealing cards, standing up from a chair, walking for 20 meters, wal-

king for 20 meters with arms crossed on the chest, fast walking for 20 meters,

fast walking for 20 meters with arms crossed on the chest, treadmill walking at

1.5mph for 5 minutes, walking for 40 meters and fast walking for 400 meters.

Before each task participants were given verbal instructions by a supervising

trained professional recording the times of the beginning and end of each task

with a stopwatch. During the free-living portion of the experiment, partici-

pants were equipped with accelerometers for seven consecutive days and were

told to maintain their normal, unsupervised, free-living activities. They were

instructed to take oﬀ the activity monitors only during sleep.

6 Marta Karas et al.

2.3 Open-source summaries of accelerometry data

The raw accelerometry data are collected along three orthogonal axes in the

device-speciﬁc frame of reference. We denote the vector of raw acceleration

data by x(t) = {x1(t), x2(t), x3(t)}, where xm(t) is the acceleration measure-

ment along the m(= 1,2,3) axis at time t(for notational simplicity we drop

the subject index). The acceleration time series, x(t), are sampled at a ﬁxed

frequency f. For example, in our application, the sampling frequency is 80 Hz,

which means that for each second along each axis there are 80 acceleration

recordings. Here we describe four open-source methods that have been used in

the literature to provide summaries of accelerometry data. All methods con-

sider non-overlapping time windows of a given length Hand reduce the 3H

measurements in the window to a single number.

Euclidean Norm Minus One (ENMO) was ﬁrst introduced as a summary

metric for accelerometry data in van Hees et al (2013). It is directly based on

the Euclidean norm of x(t), deﬁned as

r(t) = qx2

1(t) + x2

2(t) + x2

3(t).(1)

The ENMO at time tis deﬁned as r(t)−1 when r(t)−10 and 0 otherwise,

or notationally, max [r(t)−1,0]. Further, the ENMO in a window of size His

deﬁned as the average ENMO across the time points in that window. Formally,

ENMO(t;H) = 1

H−1

h=0

max [r(t+h)−1,0].(2)

ENMO can be quite sensitive to calibration errors when the device-speciﬁc

ENMO at rest is not close to zero. An additional calibration procedure was

introduced to mitigate the eﬀects of calibration (van Hees et al (2014)). This

procedure performs a linear transformation on the raw data before computing

the Euclidean norm, resulting in a new version of ENMO. Calibration para-

meters amand dmare estimated for each axis m= 1,2,3, which are later used

to linearly transformed the original data to x0(t) = {x0

1(t), x0

2(t), x0

3(t)}, such

at x0

m(t) = dm+amxm(t) for m= 1,2,3.

The Vector Magnitude Count (VMC) is an aggregation statistic that is also

known as the Mean Amplitude Deviation (MAD) V¨ah¨a-Ypy¨a et al (2015).

We use the notation VMC to avoid the confusion between the two MAD

acronyms used in accelerometry literature, one for mean amplitude deviation

(V¨ah¨a-Ypy¨a et al, 2015) and one for median amplitude deviation (Mariani

et al, 2013). VMC computes the L1norm in each time window H. Denote the

average Euclidean norm in the window of length Hstarting at tas ¯r(t;H) =

PH−1

h=0 r(t+h)/H. Then VMC is deﬁned as

VMC(t;H) = 1

H−1

h=0

|r(t+h)−¯r(t;H)|.(3)

Accelerometry data in health research: challenges and opportunities 7

The unnormalized Activity Index (AI0) (Bai et al (2014) is a measure ba-

sed on the combination of the three within-axis standard deviations of the raw

accelerometry signal. Because AI0subtracts the local mean of the accelerome-

try signal, calibration is intrinsic and local, which allows it to adapt to cases

when the device is not calibrated, when it gets decalibrated during studies, or

when the device exhibits time-dependent decalibration. Let σm(t;H) be the

standard deviation of the acceleration along axis m= 1,2,3 in the window of

length Hstarting at t. The exact formula is

σ2

m(t;H) = 1

H−1

h=0

[xm(t+h)−µm(t;H)]2,

where µm(t;H) = 1

HPH−1

h=0 xm(t+h). Then formally,

AI0(t;H) = v

tmax (1

3"3

m=1

σ2

m(t;H)−¯σ2#,0),(4)

where ¯σis the systematic noise standard deviation calculated using the data

collected during some non-moving period. The unnormalized Activity Index

(AI0) is expressed in Earth gravitational units.

The normalized Activity Index (AI, Bai et al (2016)) is strongly related to

the unnormalized Activity Index (AI0). The only diﬀerence is that the axis-

speciﬁc variances are divided by the device-speciﬁc systematic noise. More

speciﬁcally, AI is deﬁned as follows

AI(t;H) = v

tmax (1

3"3

m=1

σ2

m(t;H)−¯σ2

¯σ2#,0),(5)

The downside of using AI versus AI0is that its scale is no longer in Earth

gravitational units. Instead it is expressed in multiples of noise standard de-

viation. The advantage of AI could be when the devices are not calibrated in

terms of their noise level at rest, which may induce batch eﬀects. This, howe-

ver, seems to be a smaller problem than the internal calibration obtained by

subtracting the local mean in AI and AI0.

In the remainder of this paper, we use ENMO, VMC and AI0. We do not

use normalized Activity Index (AI) so as to keep presentations of the statistics

measurements on the same scale.

3 Statistical challenges and examples

3.1 Data volume and complexity

Figure 1 displays three time series that represent the acceleration along each

of the three orthogonal axes of an accelerometer located at the wrist. The

8 Marta Karas et al.

top panel of Figure 1 presents 24 hours of data, with each axis data shown

in a diﬀerent color. In this example, the ﬁrst observation was taken at 12AM,

while the last was taken 24 hours later, also at 12AM. The middle panel in

Figure 1 displays a particular one hour interval from 8 : 40AM to 9 : 40AM

(indicated in the top panel as a dashed-line rectangle). The bottom panel di-

splays the one minute interval marked as a dashed-line rectangle in the middle

panel (from 8 : 51AM to 8 : 52AM). The signal was acquired at a sampling

frequency of fs= 80Hz. Therefore, the number of observations per subject qu-

ickly explodes. For example, a week of accelerometry data collected at 80Hz

results in 80 ∗60 ∗60 ∗24 ∗7 = 48,384,000 observations for each of the three

axes. Thus, even for a small multi-subject observational study researchers are

faced with datasets consisting of billions of observations. This enormous vo-

lume of data creates challenges at every level of the scientiﬁc investigation.

Storage and operational memory of modern computers is not unlimited and

well-optimized solutions are needed for data management. Conducting explo-

ratory data analysis, visualization, and modeling requires additional compu-

tational and methodological resources. Therefore, it is essential to implement

carefully planned protocols for collection, management and analysis of the da-

ta. In recent years several protocols and experimental design guidelines have

been proposed (Esliger et al, 2005; Cain, 2014; NHANES, 2011), though an

universally accepted approach is still elusive. This could be due to the con-

stant change of the technological and methodological landscape. For example,

the US NHANES survey protocols have changed between survey cycles 2003-

2004 (NHANES, 2006), 2005-2006 (NHANES, 2008), 2011-2012 and 2013-2014

(NHANES, 2011) both in terms of device type and body location. In the 2003-

2004 and 2005-2006 cycles, participants wore an ActiGraph 7164 on a waist

belt for 7 days during the non-sleeping time. In later cycles, the GT3X+ Ac-

tiGraph waterproof model was used on the non-dominant wrist for 7 days

without taking it oﬀ. The protocol change was reported (Troiano et al, 2014)

and was designed to improve participant compliance.

The current approach to reduce size and complexity of the raw accelerome-

ter data is to create aggregated summaries in ﬁxed time intervals, as described

in Section 2.3. While these summaries reduce the volume of data, the potential

loss of information can be substantial. To recover some of the information, a

few methods have been proposed for recognition of activities using raw acce-

lerometry data. Some approaches focus on prediction of a group of movement

types (Lyden et al, 2014), whereas others focus on prediction of a speciﬁc

movement type (e.g. walking) (Urbanek et al, 2015). Characterizing the kine-

matics of human walking both in the lab and in the free-living environments

using accelerometry data could provide previously unavailable information on

the physical condition of individuals (Studenski et al, 2011).

Accelerometry data in health research: challenges and opportunities 9

Figure 1 Acceleration values from three orthogonal axes of an accelerometer located on

the left wrist. Each axis data is shown in a diﬀerent color. The top panel displays 24 hours

of data collected between 12AM and 12AM. The middle panel displays a one hour interval

from 8 : 40AM to 9 : 40AM (indicated in the top panel as a dashed-line rectangle). The

bottom panel displays a one minute interval from 8 : 51AM to 8 : 52AM marked as a

dashed-line rectangle in the middle panel. The signal was acquired at a sampling frequency

of fs= 80Hz.

3.2 Data heterogeneity

Interpretation of raw accelerometry data is an open and challenging problem

due to the high heterogeneity of data, both within- and between-subjects.

Within-subject variability is observed when one person performs the activity,

but the characteristics of that activity change. For example, when walking,

consecutive strides diﬀer slightly in duration and shape due to natural stride-

to-stride variability (IJmker and Lamoth, 2012; Urbanek et al, 2017). They

can also diﬀer substantially during the day, depending on the level fatigue of

the individual, context of walking (e.g. hiking versus shopping), and local con-

straints (e.g. running to a meeting when late versus slow walking to the kitchen

in the morning). Between-subject variability contains additional factors due

to diﬀerences in body size, musculature, will, and ability to perform certain

tasks. To further illustrate these points, the left column of Figure 2 displays

acceleration data collected on the left wrist during walking for two individuals.

The periodic character of time series is striking. However, both the duration

and amplitude of accelerometry data can vary from cycle to cycle. A more

extreme example is displayed in the right column of Figure 2, where the same

two individuals perform the getting dressed activity. Indeed, the more random

10 Marta Karas et al.

Figure 2 Data recorded by an accelerometer located on the left wrist while walking (left

column) and getting dressed (right column), for two individuals (top and bottom row). Each

axis is shown in a diﬀerent color.

character of the observations and lack of synchronization between- and within

individuals is remarkable.

These data indicate that it is important to better deﬁne the types of phy-

sical activity. Indeed, the bottom panels of Figure 2 show that data can be

very diﬀerent even for what is deﬁned as the same type of activity (e.g. get-

ting dressed). In retrospect, this should not be surprising as “getting dressed”

is a complex task that is only vaguely deﬁned, may involve diﬀerent types of

clothes, body sizes and shapes, movements that individuals use to getting dres-

sed, and order of various tasks. It should be apparent that if we want to make

any progress in this area, we need to identify well-deﬁned sub-movements that

then translate them into research language. Clearly, in this case, the activity

“getting dressed” does not have a sharp deﬁnition, especially from the point

of view of an accelerometer. Things may be diﬀerent if we had a video camera

instead of an accelerometer, but this raises other problems that exceed the

scope of the current paper.

Classifying activity types while accounting for between- and within-subject

variability is under intense methodological development. The approach discus-

sed in Bai et al (2012) and He et al (2014) uses short segments of training

accelerometry data, called “movelets”, to construct dictionaries used as re-

ference to predict activity types on new data. Dictionaries are activity- and

subject-speciﬁc to account for the individual variations in movement patterns

across subjects. Xiao et al (2016) proposed a related movelet-based method to

use labeled activity data from some subjects to predict the activity labels of

other subjects.

Accelerometry data in health research: challenges and opportunities 11

3.3 Sensor location

The location of the accelerometer sensor on the body can also have substan-

tial eﬀects on predicting activity types, PA volume, and PA distribution as a

function of time of the day. Indeed, the sensor collects only its own accelera-

tion, which is a proxy of the acceleration of the particular body location where

the sensor is attached. Therefore, a wrist sensor will produce diﬀerent signals

from a hip or thigh sensor. To illustrate this point, Figure 3 displays the raw

accelerometry data collected simultaneously by sensors located on the hip (left

column) and left wrist (right column) while dealing cards (top panel), getting

dressed (middle panel) and walking (bottom panel). For dealing cards, an acti-

vity that requires mostly hand-movements, the signal amplitude for the wrist

sensor is much higher than for the hip sensor. For more complex, whole-body

activities, such as getting dressed, the signals corresponding to both locations

have higher amplitudes than for dealing cards. However, the amplitude of the

signal at the wrist is higher and there are no clear correlations between the hip

and wrist signals. In the case of walking, the amplitudes of the data collected

on the wrist and the hip look periodic and highly correlated. This is likely

due to the fact that both hands and legs are involved in walking, with roughly

the same frequency of movement. The slightly lower amplitudes observed at

the wrist are probably due to the more intense PA in the lower body during

walking. In our experience, an accelerometer placed at the ankle would display

even higher amplitudes of the acceleration signal.

Figure 4 provides the summary metrics introduced in Section 2.3 using

data collected during various activities performed in the controlled lab envi-

ronment. The boxplots for ENMO, VMC and AI0are displayed in the top,

middle and bottom panels, respectively, while the data for the hip and left

wrist are displayed in the left and right columns, respectively. All summary

statistics are calculated for a window size of 5 seconds during writing, washing

dishes, vacuuming, getting dressed and walking for each of 5-second intervals

and all 49 individuals. The diﬀerence between the wrist and hip data can be

observed for all summaries and activities requiring dynamic upper body mo-

vements (washing dishes, vacuuming). These diﬀerences are much lower for

low intensity hand movements (writing), though AI0seems to better capture

these diﬀerences. Also, substantial diﬀerences can be observed for moderate-

intensity whole-body activities (getting dressed), where, again, AI0seems to

be more sensitive. For walking, the diﬀerence in the distribution of all sum-

maries between the wrist and the hip is relatively small. When comparing AI0

versus ENMO versus VMC at the wrist there is a more clear diﬀerentiation

between activities. This is likely due to the fact that AI0automatically cor-

rects for possible device miss-calibration, whereas the other methods do not. It

is an open problem whether the methods perform more similarly after device

and/or signal calibration.

Summaries have been used extensively in the literature. For example, Ko-

ster et al (2016) showed that diﬀerent cut-points of the vector magnitude can

be used for the classiﬁcation of sedentary time in older adults using both

12 Marta Karas et al.

Figure 3 Acceleration from three orthogonal axes of an accelerometer located on the hip

(left column) and left wrist (right column), while dealing cards (top row), getting dressed

(middle row) and walking (bottom row). Each axis data are shown in a diﬀerent color.

hip-worn and wrist-worn ActiGraph accelerometers. It has also been shown

that activity recognition algorithms perform diﬀerently across body locations.

For example, Rosenberger et al (2013) reported greater sensitivity and speci-

ﬁcity of both sedentary and moderate- to vigorous-intensity PA when using

accelerometry data collected at the hip compared to wrist. Trost et al (2014)

reported higher accuracy for hip-derived than for wrist-derived data in the

classiﬁcation of speciﬁc, whole-body engaging activities. Higher accuracy of

classifying sitting was obtained with data collected from the wrist compared

to data collected from the hip. Del Din et al (2016) showed that estimated ga-

it characteristics, such as step time and length, can depend on body location

and suggested that the chest location is more appropriate than the wrist. We

conclude that the body location may lead to diﬀerent results, that body loca-

tion will favor movements that directly engage that particular location, and

that translating various summaries into well deﬁned activity categories (e.g.

walking, sedentary) requires location-speciﬁc calibration of the accelerometry

summaries.

Accelerometry data in health research: challenges and opportunities 13

Figure 4 Boxplots of ENMO, (top panels) VMC (middle panels) and AI0(bottom panels)

statistics derived for τ= 5 seconds-length intervals of data collected from the hip (left

column) and left wrist (right column) during writing, washing dishes, vacuuming, getting

dressed and walking (x-axis), for all 49 individuals.

3.4 Device rotation

Because wearable PA monitors collect acceleration data relative to earth gra-

vity, they are sensitive not only to movement but also to their own orientation

with respect to Earth’s gravity. To better understand this, the top panels in Fi-

gure 5 display accelerometry data collected by a wrist-worn device during two

walking tasks performed by the same individual. The upper panel corresponds

to walking with both hands moving naturally, whereas the bottom panel cor-

responds to walking with arms crossed on the chest. The change in the device

orientation is manifested in the change of mean values, most clearly seen in

the signals shown in green. In the free living environment changes in device

rotation are quite common and can be due to multiple sources. For example,

when walking individuals could sway their hands normally, hold them in the-

ir pockets or perform an activity (e.g. holding a smart phone), walking with

hands hanging loose or with hands in the pockets. Additionally, the device can

rotate around the wrist or move higher or lower on the hand, resulting in an

altered distribution of the observed signal.

To prevent a device from rotating, it can be directly attached to the skin

with adhesive pads or aﬃxed indirectly with a waistband clip or elastic belt

(Matthews et al, 2012). Indeed, it has been recommended that devices should

be ﬁtted as tightly to the body as possible (Boerema et al, 2014). However,

even with these precautions, belts can loosen up, resulting in device orien-

tation changes (O’Neill et al, 2017). Devices that adhere to the skin might

14 Marta Karas et al.

Figure 5 Acceleration values from three orthogonal axes of an accelerometer located on the

left wrist, collected during two walking tasks performed by the same individual. Each axis

data is shown in a diﬀerent color. The upper panel corresponds to walking with both hands

moving naturally, whereas the bottom panel corresponds to walking with arms crossed on

the chest.

be attached upside-down or placed in a slightly diﬀerent position when deta-

ched and reattached. Moreover, for devices that have it, the orientation chart

can be obscured, which can result in improper device orientation (Edwardson

et al, 2016). We consider that all these precautions should be carefully im-

plemented and adapted to the speciﬁc problem that one tries to address. In

addition, it is important to understand the extent of the problem in speciﬁc

applications and, when possible, attempt to correct it. Several methods were

proposed to address this problem by rotating the observed three-dimensional

vector to the common, reference, orientation (Xiao et al, 2016; Yurtman and

Barshan, 2017). When one uses summary metrics that are robust to device

orientation, this problem is less important. For example, ENMO, VMC and

AI are all rotation invariant.

3.5 Sampling frequency

Sampling frequency fs(expressed in Hz) is the parameter describing how often

accelerometry data are collected by the device. In modern wearable accelero-

meters sampling frequency usually ranges between 10 to 200 Hz, though it

could be set as high as 1000 Hz for specialized applications, such as preci-

se human movement tracking during sport activities (Dominguez-Vega et al,

Accelerometry data in health research: challenges and opportunities 15

Figure 6 Boxplots of ENMO, (top panels) VMC (middle panels) and AI0(bottom panels)

for 5 second time windows. Data is shown for all 49 individuals in the study and were

collected from the left wrist during writing, washing dishes, vacuuming, getting dressed and

walking (x-axis). Data were collected with the original sampling frequency fs= 80 Hz and

then decimated to simulate sampling frequencies of 40, 20 and 10 Hz.

2015). One of the hidden problems is that the summaries produced by vario-

us devices can depend on the sampling frequency of the device. This could

have substantial implications if, for example, in a study the sampling frequ-

ency is varied between- and within-individuals and/or devices. For example,

according to ActiGraph, LLC, the manufacturer of ActiGraph PA monitors,

the observed activity counts depend on the sampling frequency (ActiGraph,

2016).

Figure 6 illustrates the boxplots of ENMO (top panel), VMC (middle pa-

nel) and AI0(bottom panel) calculated for all 49 participants during writing,

washing dishes, vacuuming, getting dressed and walking in 5-second time in-

tervals. Data were collected by the device located on the left wrist with the

original sampling frequency fs= 80 Hz. Additionally, data has been decimated

to simulate sampling frequencies of 40, 20 and 10 Hz. The median summary

values are reported in Table 1. Interestingly, all open source summaries are

relatively stable as a function of frequency, with stronger decreases at 10 Hz.

For the AI0statistic, the change is most pronounced for walking, where the

median AI0decreases by 19.0% from 80 to 10 Hz. For the other activities the

reduction is much smaller in the 1 to 7% range. These smaller diﬀerences in

open source measures are encouraging. Indeed, Brønd and Arvidsson (2016)

investigated the eﬀects of sampling frequency on ActiGraph activity counts.

They compared results obtained during walking and running with the default

30 Hz sampling frequency with 40 and 100 Hz sampling frequencies. For fast

16 Marta Karas et al.

Table 1 Median values of ENMO, VMC and AI0statistics derived for τ= 5 seconds-length

intervals of data collected from the left wrist during writing, washing dishes, vacuuming,

getting dressed and walking, for all 49 individuals. Results are obtained for original sampling

frequency fs= 80Hz (3rd column) and simulated sampling frequencies fs= 40,20,10 H z

(4-6th columns).

Statistic Activity 80Hz 40Hz 20Hz 10Hz

ENMO Writing 0.015 0.015 0.016 0.016

ENMO Washing Dishes 0.014 0.014 0.014 0.014

ENMO Vacuuming 0.031 0.031 0.031 0.028

ENMO Getting Dressed 0.040 0.040 0.040 0.036

ENMO Walking 0.195 0.195 0.191 0.176

VMC Writing 0.003 0.002 0.002 0.002

VMC Washing Dishes 0.007 0.006 0.006 0.004

VMC Vacuuming 0.023 0.022 0.022 0.019

VMC Getting Dressed 0.034 0.033 0.033 0.029

VMC Walking 0.198 0.197 0.193 0.175

AI0Writing 0.007 0.007 0.007 0.007

AI0Washing Dishes 0.026 0.027 0.027 0.026

AI0Vacuuming 0.071 0.071 0.070 0.066

AI0Getting Dressed 0.089 0.089 0.088 0.083

AI0Walking 0.193 0.192 0.183 0.156

run activity, they reported approx. 6,800 counts per minute (cpm) mean for

default 30 Hz frequency, and an increase of mean cpm as high as 24% for 40

Hz and 18% for 100 Hz frequency.

The Table 3 in Appendix A) provides the medians, 25-th, and 75-th per-

centile for the same ﬁve activities, but broken down by subject for 5 diﬀerent

subjects. Results are only shown for the hip accelerometer, though similar re-

sults are available for wrist accelerometers. The ranking of activity intensities

by medians across subjects is the same, but the medians for individual sub-

jects are quite variable even for the same activity and summary metric. For

reference, at the hip, AI0is roughly around 0 milli g for writing, between 1

and 2 milli g for washing dishes, 3 and 14 milli g for vacuuming, 5 and 10

milli g for getting dressed, and 20 to 100 milli g for walking. Here we used

only the minimum of the ﬁrst and maximum of the third quartiles across the

ﬁve subjects to create these ranges. These measures are averages per second

during the 5 second intervals and not totals. If one would like to transform

these numbers into totals per minute then the values need to be multiplied by

60; similar for other intervals.

3.6 Measurement bias and batch eﬀects

Measurement bias is the diﬀerence between the measured accelerations and

their true values. Estimation of both bias and measurement error of acce-

lerometry data requires a dedicated experimental setup utilizing calibrated

vibration exciters. The device is exposed to a known acceleration and the

measured acceleration time series is compared to this known acceleration, as

Accelerometry data in health research: challenges and opportunities 17

described in Bassett et al (2012). Authors noted that newer devices, such Ac-

tiGraph GT1M, undergo initial unit calibration during production and are

supposed to be calibrated for as long as they are used. However, the calibra-

tion standards of the diﬀerent manufacturers may be diﬀerent. Therefore, we

recommend doing some basic calibration checks before utilizing the device in

a study.

Even when dynamic calibration is conducted, bias in static measurements

can still exist. Ideally, an accelerometer resting on a ﬂat surface, with one me-

asurement axis oriented perpendicularly to the ground, should measure con-

stant acceleration of 1gfor that axis and 0gfor the other two orthogonal axes.

In practice, measurements may deviate slightly due to their imprecision or to

the quality of assembly, which may, for example, misalign the accelerometer

axes with the monitors’ casing. In theory, the vector magnitude (Eq. (1)) for

resting state (no movement) is equal to 1g, as the earth’s gravity is the only

force acting on the accelerometer. In practice, however, the vector magnitude

at rest can be slightly diﬀerent from 1g. To quantify this calibration bias, we

calculated a vector magnitude r(t) averages from 5-second time windows, de-

noted r(t)τ=5, for the acceleration signal collected from the hip during sitting

still activity, for all 49 subjects. We report the percentiles of this distribution

in Table 2. Ideally, if no bias was present in the data, r(t)τ=5 values should all

be equal to 1. The median is 1.007 indicating very close agreement with what

we expect. However, the 5 and 95 percentiles were 0.964 and 1.047, indicating

that 10% of the devices have a deviation of 5% or more from 1g at rest.

Table 2 Percentiles of r(t)τ=5, vector magnitude r(t) averages from 5-second time windows,

for the acceleration signal collected from the hip during sitting still activity, for all 49

subjects.

Percentile 5 25 50 75 95

r(t)τ=5 [g] 0.964 0.991 1.007 1.016 1.047

The Activity Index (Bai et al, 2016) was designed to be robust to bias by

subtracting the local mean around each axis and to measurement error by con-

structing the mean relative to the variability at rest. An additional calibration

procedure was also introduced for ENMO to mitigate the eﬀects of calibration

bias (van Hees et al (2014)). This procedure performs a linear transformation

on the raw data before computing the Euclidean norm, resulting in a calibrated

version of ENMO.

3.7 Data labeling

Classiﬁcation of physical activity types requires gold standard labels for activi-

ty, which can be obtained through direct observation or by inspection of video

recordings. Obtaining gold standard labels is a labor intensive process, which

18 Marta Karas et al.

Figure 7 Acceleration data from three orthogonal axes at the hip collected around the time

when a participant performed a 400-meter-walk activity. The dashed-line red box indicates

the portion of the 400-meter-walk period identiﬁed by a human observer.

often restricts the process to “in-the-lab” experiments or to a few subjects.

Even gold standard labels are of diﬀerent quality, depending on the accura-

cy and resolution at which the activity type is predicted. For example, when

one is interested in predicting standing up from a chair, the duration of the

activity is in the 1 to 4 second range. This makes accurate labeling even for a

human observer extremely diﬃcult. If one is interested in predicting whether

the person walked or not in a particular 1 minute interval, the labels will be

less accurate and will simply indicate whether the person has walked in a par-

ticular interval and approximately around what time stamp and for how long.

However, even these labels can be substantially mis-aligned. This can be due

to multiple factors including imperfect synchronization of clocks, time elapsed

between the beginning and recording of the task, basic observer or data en-

try error. As an example, Figure 7 displays a portion of the acceleration data

recorded at the hip for one subject. The dashed-line box is the portion of 400-

meter-walk period labeled by the human observer. There is a clear 45 second

shift in the label relative to the actual activity. In such cases using the original

label without inspection would lead to inferior prediction algorithms and wa-

ste of resources during the modeling phase. We chose to manually inspect all

“in-the-lab” data for each subject and re-label walking as periods that closely

correspond to the sustained harmonic walking (SHW). The overlap between

the labels provided by the human observer and the labels improved by human

inspection of the data was below 80% in 18 out of 49 subjects.

The most eﬀective approach to proper labeling is synchronization of accele-

rometry data with video-recordings of the experiment (Bussmann et al, 1998).

This method has been successfully used in many “in-the-lab” experiments

(Godfrey et al, 2015; Del Din et al, 2016). However, using video-recordings

for labeling free-living data is harder and subject to privacy considerations.

Indeed, it is possible to equip participants with body-worn cameras, but vi-

deo data de-identiﬁcation can be quite challenging as it might contain family

Accelerometry data in health research: challenges and opportunities 19

members, car plates, and addresses. In spite of these limitations, body-worn

cameras have been used to label PA collected in the free-living environment.

For example, Ellis et al (2016) used video data to train a PA classiﬁer and

Hickey et al (2017) used video data to train a walking prediction algorithm.

In Hickey’s experiment, body-worn cameras were facing down to record only

the feet movement.

Another, less labor intensive, approach for precise labeling of raw accele-

rometry data is to use landmarks introduced by the individual wearing the

device. For example, in the case of wrist-worn devices, participants can clap

their hands before and after each tasks. Claps result in high-amplitude, short-

time spikes in the observed data that can be used to estimate timestamps for

each task. For other body locations, participants can vigorously tap the device

to generate proper landmarks in the data. This approach was successfully used

in Straczkiewicz et al (2016). Another approach is to use the device-speciﬁc

own event-markers to place labeling landmarks. Some modern PA monitors

have built-in event marker buttons (e.g. Actiwatch Spectrum Plus and Pro).

These event-markers generate binary indicators at the same granularity as

that of the observed accelerations. When pressed, they return a value of 1

until pressed again. Initially, they were intended to mark major everyday ac-

tivities (e.g. sleep) and critical events (e.g. falls), but they can be used for

labeling data, as well. Unfortunately, the utility of event markers is limited by

the compliance of participants (Chen et al, 2014; Boudebesse et al, 2015).

3.8 Synchronization of multiple PA monitors

Many studies collect data using multiple PA sensors. Synchronizing data across

diﬀerent monitors allows to combine information about speciﬁc human move-

ments at multiple locations on the body (Bao and Intille, 2004; Cleland et al,

2013; He et al, 2014; Altini et al, 2015). Most devices can be set up to initialize

their measurement collection at a given time (e.g. at midnight) and/or can be

initialized manually. In practice, even if such approaches are used, measure-

ments might still be desynchronized between devices. We identify two main

reasons for device desynchronization.

First, most operating systems used in personal computers are not real-time

operational systems. Therefore, time of execution of any command can not be

precisely determined (see details in Stallings (2008)). That may result in sub-

second level diﬀerences in measurements start times on multiple PA monitors.

Second, the internal drift of device clocks can lead to inaccurate stamping

of the time interval (Bennett et al, 2015). Such drifts are usually small (a

few seconds per day) and can typically be ignored. However, when combining

sub-second level data from multiple sensors in the free-living environment,

the eﬀects of the drift can have substantial side eﬀects. To illustrate these

problems, Figure 8 displays 20 seconds of data collected in the free-living

environment by two monitors located on the left (top panels) and the right

wrist (bottom panels). The two sensors were synchronized at the beginning of

20 Marta Karas et al.

Figure 8 Acceleration values representing 20 seconds of data collected in the free-living

environment using two monitors located on the left wrist (top panels) and the right wrist

(bottom panels). The two sensors were synchronized at the beginning of the experiment.

The left and right column provide data collected on the ﬁrst and seventh day of observation,

respectively.

the experiment. The left and right column provide data collected on the ﬁrst

and seventh day of observation, respectively. The dotted vertical lines mark

the end of a high-amplitude activity for each device. Although visually, data

from the two devices appears to be correlated during both days, a shift in the

two recordings is apparent on both days. On day 1 the time-shift is around

2.5 seconds, which is probably due to imperfect timing of device initialization.

On day 7 the time shift is about 4 seconds, with the additional 1.5 seconds

probably due to drift in device clocks. This drift need not be in the same

direction or of the same magnitude for all devices. Such de-synchronization

would have minor implications for PA summaries at the minute level collected

over a 7 to 14 days period, but they can lead to substantial diﬀerences when

one is interested in analyzing sub-second level data.

The time shift introduced during device synchronization and initialization

can be addressed using video recordings or landmarks. Lab experiments ra-

rely last more than a few hours, and the eﬀects of time drift can often be

ignored. However, for data collected ‘in the free-living environment, the eﬀects

of time drift are cumulative, which can raise substantial challenges for data

analysis. A possible solution could be to use devices designed speciﬁcally for

parallel measurements. For example, GaitUp (GaitUp, 2017) is a system for

synchronous measurement of feet movement. Alternatively, one can perform

Accelerometry data in health research: challenges and opportunities 21

landmark-based synchronization for every time-interval in the data (e.g. every

day).

4 Discussion

We have presented challenges related to the collection and analysis of raw,

sub-second level accelerometry data. The increased granularity of observa-

tions when moving from the minute to the sub-second resolution leads to a

large increase both in the volume and complexity of the data. This makes raw

data harder to use than summarized data, but also holds promise of unlocking

additional information glossed over by taking minute-, hour-, or day-level sum-

maries. For example, describing gait parameters during the course of the day

in the free-living environment and characterizing their potential association

with health outcomes cannot be done without using sub-second level data.

Raw accelerometry data requires specialized visualization and analytic me-

thods. This is fertile ground both for scientiﬁc researchers, as additional in-

formation is likely embedded in the raw signals, and for data scientists, as

new methods and insights are becoming increasingly necessary. To start ad-

dressing this complexity we make a few points that are worth remembering:

1) activity counts are summaries of the raw data, which can depend on the

device manufacturer, software version, and body location; 2) open source sum-

mary statistics are increasingly available, though more research is needed to

understand their relative performance; 3) raw and summarized PA data can

vary substantially with the device location, between- and within-individuals;

4) proper location choice might yield data signatures tailored to a particular

study purpose; 5) device orientation can change over the course of an experi-

ment and needs to be standardized both within- and between-individuals and

devices; 6) sampling frequency can aﬀect both raw accelerometry and sum-

marized measurements; 7) device calibration, bias removal, and measurement

error quantiﬁcation can lead to higher quality data; 8) proper labeling of data

is very important for training activity classiﬁers at the sub-second level, espe-

cially for short activities; and 9) synchronizing multiple devices must be done

carefully and needs to be accounted for during the design of the experiment

phase.

In spite of these challenges, the number of publications focused on raw ac-

celerometry is continuously increasing, especially in the area of activity type

classiﬁcation. This is due in part to the increased popularity of these devices,

their convenient design, and reduced cost. The application of raw accelerome-

try data in epidemiological studies is still in its infancy, though some important

steps forward have been made. We anticipate that, as the interest changes to

understanding the details of human movement kinematics in the free-living

environment, the focus on raw data will become stronger. The number of stu-

dies that both collect and disseminate raw activity data will probably provide

a huge boost to raw accelerometry data research. For example, the UK Bio-

22 Marta Karas et al.

bank PA dataset is currently the largest of its kind. It was collected using the

open-hardware AX3 acceleration sensor (Doherty et al, 2017).

In closing, we oﬀer a few practical suggestions for the scientists who would

like to conduct their own activity studies: 1) discuss your plans with a team

that has expertise in activity research; 2) avoid the pitfalls of accelerometry

research by incorporating robust, fault-tolerant designs of experiments; 3) if

possible, use established protocols for data collection and pre-processing; 4)

record and store the raw accelerometry data in addition to summaries, such as

activity counts; 5) conduct a lab study and record the activity summaries for a

well-deﬁned group of activities in 10 to 100 individuals who are representative

of the population to be studied.

5 Acknowledgements

The authors would like to acknowledge Annemarie Koster, PhD and Paolo

Caserotti, PhD for designing the DECOS experiments.

6 Funding

This research was supported by Pittsburgh Claude D. Pepper Older Ameri-

cans Independence Center, Research Registry, and Developmental Pilot Grant

(PI: Glynn) NIH P30 AG024826 and NIH P30 AG024827. National Institu-

te on Aging Professional Services Contract HHSN271201100605P. NIA Aging

Training Grant (PI: AB Newman) T32-AG-000181. The project was suppor-

ted, in part, by the Intramural Research Program of the National Institute on

Aging.

References

ActiGraph (2016) What are counts? URL https://actigraph.desk.

com/customer/en/portal/articles/2515580-what-are-counts- , acces-

sed: 2017-11-23

Altini M, Penders J, Vullers R, Amft O (2015) Estimating energy expenditure

using body-worn accelerometers: A comparison of methods, sensors num-

ber and positioning. IEEE Journal of Biomedical and Health Informatics

19(1):219–226, DOI 10.1109/JBHI.2014.2313039

Attal F, Mohammed S, Dedabrishvili M, Chamroukhi F, Oukhellou L, Amirat

Y (2015) Physical Human Activity Recognition Using Wearable Sensors.

Sensors (Basel) 15(12):31,314–38, DOI 10.3390/s151229858

Bai J, Goldsmith J, Caﬀo B, Glass TA, Crainiceanu CM (2012) Movelets: A

dictionary of movement. Electronic Journal of Statistics 6:559–578, DOI

10.1214/12-EJS684

Accelerometry data in health research: challenges and opportunities 23

Bai J, He B, Shou H, Zipunnikov V, Glass TA, Crainiceanu CM (2014) Nor-

malization and extraction of interpretable metrics from raw accelerometry

data. Biostatistics 15(1):102–116, DOI 10.1093/biostatistics/kxt029

Bai J, Di C, Xiao L, Evenson KR, LaCroix AZ, Crainiceanu CM, Buchner DM

(2016) An activity index for raw accelerometry data and its comparison with

other activity metrics. PLoS ONE 11(8), DOI 10.1371/journal.pone.0160644

Bao L, Intille SS (2004) Activity Recognition from User-Annotated Accele-

ration Data Pervasive Computing. Pervasive Computing 3001:1–17, DOI

10.1007/978-3-540-24646-6˙1

Bassett DR, Rowlands A, Trost SG (2012) Calibration and validation of we-

arable monitors. Medicine and Science in Sports and Exercise 44(SUPPL.

1), DOI 10.1249/MSS.0b013e3182399cf7

Bennett TR, Gans N, Jafari R (2015) Multi-sensor data-driven: Synchro-

nization using wearable sensors. In: Proceedings of the 2015 ACM In-

ternational Symposium on Wearable Computers, ACM, New York, NY,

USA, ISWC ’15, pp 113–116, DOI 10.1145/2802083.2808393, URL http:

//doi.acm.org/10.1145/2802083.2808393

Boerema S, van Velsen L, Schaake L, T¨onis T, Hermens H (2014) Optimal

Sensor Placement for Measuring Physical Activity with a 3D Accelerometer.

Sensors 14(2):3188–3206, DOI 10.3390/s140203188

Boudebesse C, Geoﬀroy PA, Henry C, Germain A, Scott J, Lajnef M, Leboyer

M, Bellivier F, Etain B (2015) Links between sleep and body mass index in

bipolar disorders: An exploratory study. European Psychiatry 30(1):89–93,

DOI 10.1016/j.eurpsy.2014.04.006

Brønd JC, Arvidsson D (2016) Sampling frequency aﬀects the processing of

Actigraph raw acceleration data to activity counts. Journal of Applied Phy-

siology 120(3):362–369, DOI 10.1152/japplphysiol.00628.2015

Bussmann J, Tulen J, van Herel E, Stam H (1998) Quantiﬁcation of physi-

cal activities by means of ambulatory accelerometry: A validation study.

Psychophysiology 35(5):488–96, DOI 10.1017/S0048577298971153

Cain KL (2014) Accelerometer scoring protocol for the ipen-adult stu-

dy. URL http://www.ipenproject.org/documents/methods_docs/IPEN_

Protocol.pdf, accessed: 2017-11-19

Chen JH, Waite L, Kurina LM, Thisted Ra, McClintock M, Lauderdale DS

(2014) Insomnia Symptoms and Actigraph-Estimated Sleep Characteristics

in a Nationally Representative Sample of Older Adults. The journals of

gerontology Series A, Biological sciences and medical sciences (8):1–8, DOI

10.1093/gerona/glu144

Chen KY, Janz KF, Zhu W, Brychta RJ (2012) Redeﬁning the roles of sensors

in objective physical activity monitoring. Medicine and Science in Sports

and Exercise 44(SUPPL. 1), DOI 10.1249/MSS.0b013e3182399bc8

Cleland I, Kikhia B, Nugent C, Boytsov A, Hallberg J, Synnes K, McClean

S, Finlay D (2013) Optimal placement of accelerometers for the detection

of everyday activities. Sensors (Basel, Switzerland) 13(7):9183–9200, DOI

10.3390/s130709183

24 Marta Karas et al.

Del Din S, Hickey A, Hurwitz N, Mathers JC, Rochester L, Godfrey A (2016)

Measuring gait with an accelerometer-based wearable: inﬂuence of device

location, testing protocol and age. Physiological measurement 37(10):1785–

1797, DOI 10.1088/0967-3334/37/10/1785

Doherty A, Jackson D, Hammerla N, Pl¨otz T, Olivier P, Granat MH, White

T, Van Hees VT, Trenell MI, Owen CG, Preece SJ, Gillions R, Sheard S,

Peakman T, Brage S, Wareham NJ (2017) Large scale population assessment

of physical activity using wrist worn accelerometers: The UK biobank study.

PLoS ONE 12(2), DOI 10.1371/journal.pone.0169649

Dominguez-Vega ZT, Martinez-Mendez R, Lorias-Espinoza D (2015) High

sampling rate datalogger for the characterization of acceleration signals on

the human body running. In: 25th International Conference on Electronics,

Communications and Computers, CONIELECOMP 2015, pp 173–177, DOI

10.1109/CONIELECOMP.2015.7086947

Edwardson CL, Winkler EA, Bodicoat DH, Yates T, Davies MJ, Dunstan DW,

Healy GN (2016) Considerations when using the activPAL monitor in ﬁeld

based research with adult populations. Journal of Sport and Health Science

(May):13–24, DOI 10.1016/j.jshs.2016.02.002

Ellis K, Kerr J, Godbole S, Staudenmayer J, Lanckriet G (2016) Hip and

wrist accelerometer algorithms for free-living behavior classiﬁcation. Medi-

cine and Science in Sports and Exercise 48(5):933–940, DOI 10.1249/MSS.

0000000000000840

Esliger DW, Copeland JL, Barnes JD, Tremblay MS (2005) Standardizing and

optimizing the use of accelerometer data for free-living physical activity

monitoring. Journal of Physical Activity and Health 2(3):366–383, DOI

http://dx.doi.org/10.1123/jpah.2.3.366

Fairclough SJ, Noonan R, Rowlands AV, van Hees V, Knowles Z, Boddy

LM (2016) Wear compliance and activity in children wearing wrist- and

hip-mounted accelerometers. Medicine and Science in Sports and Exercise

48(2):245–253, DOI 10.1249/MSS.0000000000000771

Fortune E, Lugade V, Morrow M, Kaufman K (2014) Validity of using tri-

axial accelerometers to measure human movement - Part II: Step counts at

a wide range of gait velocities. Medical Engineering and Physics 36(6):659–

669, DOI 10.1016/j.medengphy.2014.02.006

Freedson P, Bowles HR, Troiano R, Haskell W (2012) Assessment of physical

activity using wearable monitors: Recommendations for monitor calibration

and use in the ﬁeld. Medicine and Science in Sports and Exercise 44(SUPPL.

1):S1–S4, DOI 10.1249/MSS.0b013e3182399b7e

GaitUp (2017) Products. URL https://www.gaitup.com/products/, acces-

sed: 2017-11-25

Godfrey A, Del Din S, Barry G, Mathers JC, Rochester L (2015) Instrumenting

gait with an accelerometer: A system and algorithm examination. Medical

Engineering and Physics 37(4):400–407, DOI 10.1016/j.medengphy.2015.02.

003

He B, Bai J, Zipunnikov VV, Koster A, Caserotti P, Lange-Maia B, Glynn

NW, Harris TB, Crainiceanu CM (2014) Predicting human movement with

Accelerometry data in health research: challenges and opportunities 25

multiple accelerometers using movelets. Medicine and Science in Sports and

Exercise 46(9):1859–1866, DOI 10.1249/MSS.0000000000000285

Healy GN, Matthews CE, Dunstan DW, Winkler EAH, Owen N (2011) Seden-

tary time and cardio-metabolic biomarkers in US adults: NHANES 200306.

European Heart Journal 32(5):590–597, DOI 10.1093/eurheartj/ehq451

van Hees VT, Gorzelniak L, Dean León EC, Eder M, Pias M, Taherian S,

Ekelund U, Renstr¨om F, Franks PW, Horsch A, Brage S (2013) Separating

Movement and Gravity Components in an Acceleration Signal and Impli-

cations for the Assessment of Human Daily Physical Activity. PLoS ONE

8(4), DOI 10.1371/journal.pone.0061691

van Hees VT, Fang Z, Langford J, Assah F, Mohammad A, da Silva ICM,

Trenell MI, White T, Wareham NJ, Brage S (2014) Autocalibration of acce-

lerometer data for free-living physical activity assessment using local gravity

and temperature: an evaluation on four continents. Journal of Applied Phy-

siology 117(7):738–744, DOI 10.1152/japplphysiol.00421.2014

Hickey A, Del Din S, Rochester L, Godfrey A (2017) Detecting free-living

steps and walking bouts: Validating an algorithm for macro gait analysis.

Physiological Measurement 38(1):N1–N15, DOI 10.1088/1361-6579/38/1/

IJmker T, Lamoth CJ (2012) Gait and cognition: The relationship between

gait stability and variability with executive function in persons with and

without dementia. Gait and Posture 35(1):126–130, DOI 10.1016/j.gaitpost.

2011.08.022

Koster A, Shiroma EJ, Caserotti P, Matthews CE, Chen KY, Glynn NW,

Harris TB (2016) Comparison of Sedentary Estimates between activPAL

and Hip- and Wrist-Worn ActiGraph. Medicine and Science in Sports and

Exercise 48(8):1514–1522, DOI 10.1249/MSS.0000000000000924

Lange-Maia BS, Newman AB, Strotmeyer ES, Harris TB, Caserotti P, Glynn

NW (2015) Performance on fast- and usual-paced 400-m walk tests in older

adults: are they comparable? Aging Clinical and Experimental Research

27(3):309–314, DOI 10.1007/s40520-014-0287-y

Lyden K, Keadle SK, Staudenmayer J, Freedson PS (2014) A method to

estimate free-living active and sedentary behavior from an accelerome-

ter. Medicine and Science in Sports and Exercise 46(2):386–397, DOI

10.1249/MSS.0b013e3182a42a2d

Mariani B, Rouhani H, Crevoisier X, Aminian K (2013) Quantitative estima-

tion of foot-ﬂat and stance phase of gait using foot-worn inertial sensors.

Gait and Posture 37(2):229–234, DOI 10.1016/j.gaitpost.2012.07.012

Matthews CE, Chen KY, Freedson PS, Buchowski MS, Beech BM, Pate RR,

Troiano RP (2008) Amount of time spent in sedentary behaviors in the

United States, 2003-2004. American Journal of Epidemiology 167(7):875–

881, DOI 10.1093/aje/kwm390

Matthews CE, Hagstr¨omer M, Pober DM, Bowles HR (2012) Best practices

for using physical activity monitors in population-based research. Medicine

and Science in Sports and Exercise 44(SUPPL. 1):68–76, DOI 10.1249/MSS.

0b013e3182399e5b

26 Marta Karas et al.

NHANES (2006) National Health and Nutrition Examination Survey. 2003-

2004 Data Documentation, Codebook, and Frequencies. Physical Activi-

ty Monitor. Centers for Disease Control and Prevention (CDC). Natio-

nal Center for Health Statistics (NCHS), Hyattsville, MD: U.S. Depart-

ment of Health and Human Services, Centers for Disease Control and Pre-

vention, URL https://wwwn.cdc.gov/Nchs/Nhanes/2003-2004/PAXRAW_

C.htm, accessed: 2017-11-19.

NHANES (2008) National Health and Nutrition Examination Survey. 2005-

2006 Data Documentation, Codebook, and Frequencies. Physical Activi-

ty Monitor. Centers for Disease Control and Prevention (CDC). Natio-

nal Center for Health Statistics (NCHS), Hyattsville, MD: U.S. Depart-

ment of Health and Human Services, Centers for Disease Control and Pre-

vention, URL https://wwwn.cdc.gov/Nchs/Nhanes/2005-2006/PAXRAW_

D.htm, accessed: 2017-11-19.

NHANES (2011) National Health and Nutrition Examination Survey.

2011-2012 Physical Activity Monitor (PAM) Procedures Manual. Cen-

ters for Disease Control and Prevention (CDC). National Center for

Health Statistics (NCHS), Hyattsville, MD: U.S. Department of He-

alth and Human Services, Centers for Disease Control and Pre-

vention, URL https://wwwn.cdc.gov/nchs/data/nhanes/2011-2012/

manuals/Physical_Activity_Monitor_Manual.pdf, accessed: 2017-11-19.

O’Neill B, McDonough SM, Wilson JJ, Bradbury I, Hayes K, Kirk A, Kent

L, Cosgrove D, Bradley JM, Tully MA (2017) Comparing accelerometer,

pedometer and a questionnaire for measuring physical activity in bronchiec-

tasis: a validity and feasibility study. Respiratory Research 18(1):16, DOI

10.1186/s12931-016-0497-2

Pober DM, Staudenmayer J, Raphael C, Freedson PS (2006) Development of

Novel Techniques to Classify Physical Activity Mode Using Accelerometers.

Medicine and Science in Sports and Exercise 38(9):1626–1634, DOI 10.1249/

01.mss.0000227542.43669.45

Preece SJ, Goulermas JY, Kenney LPJ, Howard D, Meijer K, Crompton R

(2009) Activity Identiﬁcation using Body-Mounted Sensors - A Review of

Classiﬁcation Techniques. Physiological Measurement 30(4):R1–R33, DOI

10.1088/0967-3334/30/4/R01

Rosenberger ME, Haskell WL, Albinali F, Mota S, Nawyn J, Intille S (2013)

Estimating activity and sedentary behavior from an accelerometer on the

hip or wrist. DOI 10.1249/MSS.0b013e31827f0d9c

Schrack JA, Zipunnikov V, Goldsmith J, Bai J, Simonsick EM, Crainiceanu

C, Ferrucci L (2014) Assessing the physical cliﬀ: Detailed quantiﬁcation

of age-related diﬀerences in daily patterns of physical activity. Journals of

Gerontology - Series A Biological Sciences and Medical Sciences 69(8):973–

979, DOI 10.1093/gerona/glt199

Schrack JA, Cooper R, Koster A, Shiroma EJ, Murabito JM, Rejeski WJ,

Ferrucci L, Harris TB (2016) Assessing Daily Physical Activity in Older

Adults: Unraveling the Complexity of Monitors, Measures, and Methods.

DOI 10.1093/gerona/glw026

Accelerometry data in health research: challenges and opportunities 27

Stallings W (2008) Operating Systems: Internals and Design Principles, 6th

edn. Prentice Hall Press, Upper Saddle River, NJ, USA

Staudenmayer J, Pober D, Crouter S, Bassett D, Freedson P (2009) An arti-

ﬁcial neural network to estimate physical activity energy expenditure and

identify physical activity type from an accelerometer. Journal of Applied

Physiology (Bethesda, Md : 1985) 107(4):1300–7, DOI 10.1152/japplphysiol.

00465.2009

Staudenmayer J, Zhu W, Catellier DJ (2012) Statistical considerations in the

analysis of accelerometry-based activity monitor data. Medicine and Science

in Sports and Exercise 44(SUPPL. 1), DOI 10.1249/MSS.0b013e3182399e0f

Storti KL, Pettee KK, Brach JS, Talkowski JB, Richardson CR, Kriska AM

(2008) Gait speed and step-count monitor accuracy in community-dwelling

older adults. Medicine and Science in Sports and Exercise 40(1):59–64, DOI

10.1249/mss.0b013e318158b504

Straczkiewicz M, Urbanek JK, Fadel WF, Crainiceanu CM, Harezlak J (2016)

Automatic car driving detection using raw accelerometry data. Physiological

Measurement 37(10):1757–1769, DOI 10.1088/0967-3334/37/10/1757

Studenski S, Perera S, Patel K, Rosano C, Faulkner K, Inzitari M, Brach J,

Chandler J, Cawthon P, Connor EB, Nevitt M, Visser M, Kritchevsky S,

Badinelli S, Harris T, Newman AB, Cauley J, Ferrucci L, Guralnik J (2011)

Gait speed and survival in older adults. JAMA: the Journal of the American

Medical Association 305(1):50–58, DOI 10.1001/jama.2010.1923

Troiano RP, McClain JJ, Brychta RJ, Chen KY (2014) Evolution of accele-

rometer methods for physical activity research. British Journal of Sports

Medicine 48(13):1019–1023, DOI 10.1136/bjsports-2014-093546

Trost SG, Mciver KL, Pate RR (2005) Conducting Accelerometer - Based Ac-

tivity Assessments in Field - Based Research. Medicine and Science in Sports

and Exercise 37(11):531–543, DOI 10.1249/01.mss.0000185657.86065.98

Trost SG, Zheng Y, Wong WK (2014) Machine learning for activity recogni-

tion: hip versus wrist data. Physiological Measurement 35(11):2183–2189,

DOI 10.1088/0967-3334/35/11/2183

Urbanek JK, Vadim Zipunnikov TH, Fadel W, Glynn N, Koster A, Caserotti P,

Crainiceanu C, Harezlak J (2015) Prediction of sustained harmonic walking

in the free-living environment using raw accelerometry data. URL https:

//arxiv.org/abs/1505.04066, accessed: 2017-11-19.

Urbanek JK, Harezlak J, Glynn NW, Harris T, Crainiceanu C, Zipunnikov V

(2017) Stride variability measures derived from wrist- and hip-worn accele-

rometers. Gait and Posture 52:217–223, DOI 10.1016/j.gaitpost.2016.11.045

V¨ah¨a-Ypy¨a H, Vasankari T, Husu P, Suni J, Siev¨anen H (2015) A universal, ac-

curate intensity-based classiﬁcation of diﬀerent physical activities using raw

data of accelerometer. Clinical Physiology and Functional Imaging 35(1):64–

70, DOI 10.1111/cpf.12127

Xiao L, Huang L, Schrack JA, Ferrucci L, Zipunnikov V, Crainiceanu CM

(2015) Quantifying the lifetime circadian rhythm of physical activity: A

covariate-dependent functional approach. Biostatistics 16(2):352–367, DOI

10.1093/biostatistics/kxu045

28 Marta Karas et al.

Xiao L, He B, Koster A, Caserotti P, Lange-Maia B, Glynn NW, Harris TB,

Crainiceanu CM (2016) Movement prediction using accelerometers in a hu-

man population. Biometrics 72(2):513–524, DOI 10.1111/biom.12382

Yurtman A, Barshan B (2017) Activity recognition invariant to sensor orien-

tation with wearable motion sensors. Sensors (Switzerland) 17(8), DOI

10.3390/s17081838

Accelerometry data in health research: challenges and opportunities 29

A Appendix

30 Marta Karas et al.

Table 3 Summary of the four statistics: ENMO, VMC, AI0and AI for ﬁve selected subjects

and all subjects: median, 25-th percentile and 75-th percentile (percentiles are reported

in brackets), obtained from accelerometry data collected at the hip during ﬁve activities:

writing, washing dishes, vacuuming, getting dressed and walking.

id Writing Washing Dishes Vacuuming Getting Dressed Walking

ENMO

Sub. 1 0.021 0.033 0.037 0.035 0.189

(0.019,0.023) (0.032,0.033) (0.035,0.039) (0.032,0.038) (0.161,0.204)

Sub. 2 0.084 0.029 0.031 0.055 0.147

(0.083,0.085) (0.028,0.030) (0.027,0.047) (0.052,0.058) (0.122,0.158)

Sub. 3 0.024 0.037 0.039 0.064 0.175

(0.022,0.024) (0.037,0.038) (0.037,0.043) (0.059,0.071) (0.157,0.198)

Sub. 4 0.005 0.008 0.043 0.041 0.181

(0.003,0.006) (0.006,0.009) (0.028,0.089) (0.024,0.051) (0.154,0.192)

Sub. 5 0.017 0.019 0.038 0.035 0.297

(0.014,0.023) (0.018,0.019) (0.030,0.066) (0.031,0.044) (0.263,0.318)

All 0.015 0.014 0.031 0.040 0.195

Sub. (0.007,0.025) (0.009,0.025) (0.020,0.047) (0.031,0.055) (0.143,0.257)

VMC

Sub. 1 0.004 0.006 0.020 0.021 0.194

(0.002,0.005) (0.006,0.007) (0.015,0.023) (0.017,0.025) (0.166,0.206)

Sub. 2 0.003 0.005 0.018 0.022 0.154

(0.003,0.003) (0.004,0.005) (0.014,0.034) (0.021,0.025) (0.13,0.166)

Sub. 3 0.002 0.006 0.015 0.058 0.182

(0.000,0.003) (0.005,0.007) (0.012,0.026) (0.049,0.068) (0.164,0.205)

Sub. 4 0.003 0.008 0.043 0.041 0.186

(0.000,0.003) (0.006,0.008) (0.028,0.092) (0.024,0.051) (0.158,0.196)

Sub. 5 0.002 0.009 0.030 0.032 0.303

(0.000,0.003) (0.008,0.010) (0.024,0.065) (0.027,0.037) (0.271,0.323)

All 0.003 0.007 0.022 0.034 0.198

Sub. (0.000,0.004) (0.005,0.009) (0.014,0.041) (0.025,0.050) (0.147,0.258)

AI0

Sub. 1 0.000 0.001 0.005 0.006 0.042

(0.000,0.002) (0.001,0.002) (0.003,0.007) (0.005,0.007) (0.033,0.047)

Sub. 2 0.000 0.000 0.005 0.017 0.024

(0.000,0.000) (0.000,0.001) (0.003,0.009) (0.013,0.019) (0.021,0.027)

Sub. 3 0.000 0.001 0.003 0.009 0.034

(0.000,0.001) (0.001,0.002) (0.002,0.006) (0.008,0.009) (0.028,0.040)

Sub. 4 0.000 0.001 0.018 0.014 0.037

(0.000,0.000) (0.001,0.003) (0.007,0.024) (0.010,0.017) (0.033,0.041)

Sub. 5 0.000 0.001 0.009 0.006 0.093

(0.000,0.000) (0.001,0.002) (0.005,0.014) (0.005,0.010) (0.081,0.103)

All 0.000 0.001 0.006 0.009 0.040

Sub. (0.000,0.000) (0.000,0.002) (0.003,0.010) (0.006,0.013) (0.024,0.062)

Sub. 1 5.745 14.505 37.066 28.311 128.847

(2.504,7.369) (12.166,16.730) (33.375,41.246) (24.978,37.332) (113.873,135.39)

Sub. 2 1.124 8.421 44.291 24.230 97.084

(0.536,2.210) (7.398,9.537) (39.960,49.576) (21.487,39.287) (88.253,103.289)

Sub. 3 0.000 10.437 55.963 27.359 111.914

(0.000,2.442) (7.720,15.131) (51.028,59.763) (21.271,33.922) (102.208,121.594)

Sub. 4 0.000 11.002 51.713 59.428 117.736

(0.000,2.598) (9.077,13.150) (40.562,61.589) (34.030,83.187) (109.104,122.844)

Sub. 5 0.000 12.948 39.288 41.695 177.989

(0.000,0.544) (11.625,15.230) (34.278,45.534) (30.223,61.612) (161.927,184.464)

All 0.000 11.396 44.566 32.03 122.077

Sub. (0.000,2.294) (7.407,15.713) (35.171,56.015) (20.833,46.736) (95.299,149.542)

Wearable Accelerometers in Cancer Patients

Chapter

Nov 2021

Cancer is the leading cause of death worldwide accounting for nearly 10 million deaths in 2020 alone. The complexity of cancer treatment, recovery and survivorship in a co-morbid, aging population creates a need for novel monitoring and surveillance approaches. Research over the past few decades has linked higher physical activity (PA) with reduced risk of developing cancer, and a lower mortality in those with a cancer diagnosis. Accurate assessment of physical function is also a vital aspect of the cancer patient clinical evaluation, required to inform treatment decisions and eligibility for cancer clinical trials. Existing methods of evaluating PA are subject to recall bias and limitations. Given the growing body of evidence highlighting the limitations of self-reporting PA measures and as cancer care becomes increasingly patient centered, digital wearable tools could help us address unmet needs by offering unique advantages such as scale, cost, data volume, and continuous real-world objective data capture. In this chapter we discuss current methods of evaluating PA in the cancer patient, how wearable accelerometers have been used in cancer clinical trials and give details of a wearable accelerometer digital health trial collecting longitudinal PA data including successes and challenges.

Quantifying exposure to running for meaningful insights into running-related injuries

Article

Full-text available

Oct 2019

The very term ‘running-related overuse injury’ implies the importance of ‘use’, or exposure, to running. Risk factors for running-related injury can be better understood when exposure to running is quantified using either external or internal training loads. The advent of objective methods for quantifying exposure to running, such as global positioning system watches, smartphones, commercial activity monitors and research-grade wearable sensors, make it possible for researchers, coaches and clinicians to track exposure to running with unprecedented detail. This viewpoint discusses practical issues surrounding the use and analysis of data from such devices, including how wearable devices can be used to assess both internal and external training loads. We advocate for an integrative approach where data from multiple sources are used in combination to directly measure exposure to running in diverse settings.

On Placement, Location and Orientation of Wrist-Worn Tri-Axial Accelerometers during Free-Living Measurements

Article

Full-text available

May 2019
SENSORS-BASEL

Wearable accelerometers have recently become a standalone tool for the objective assessment of physical activity (PA). In free-living studies, accelerometers are placed by protocol on a pre-defined body location (e.g., non-dominant wrist). However, the protocol is not always followed, e.g., the sensor can be moved between wrists or reattached in a different orientation. Such protocol violations often result in PA miscalculation. We propose an approach, PLOE (“Placement, Location and Orientation Evaluation method”), to determine the sensor position using statistical features from the raw accelerometer measurements. We compare the estimated position with the study protocol and identify discrepancies. We apply PLOE to the measurements collected from 45 older adults who wore ActiGraph GT3X+ accelerometers on the left and right wrist for seven days. We found that 15.6% of participants who wore accelerometers violated the protocol for one or more days. The sensors were worn on the wrong hand during 6.9% of the days of simultaneous wearing of devices. During the periods of discrepancies, the daily PA was miscalculated by more than 20%. Our findings show that correct placement of the device has a significant effect on the PA estimates. These results demonstrate a need for the evaluation of sensor position.

Assessing Daily Physical Activity in Older Adults: Unraveling the Complexity of Monitors, Measures, and Methods

Article

Full-text available

Mar 2016

At the 67th Gerontological Society of America Annual Meeting, a preconference workshop was convened to discuss the challenges of accurately assessing physical activity in older populations. The advent of wearable technology (eg, accelerometers) to monitor physical activity has created unprecedented opportunities to observe, quantify, and define physical activity in the real-world setting. These devices enable researchers to better understand the associations of physical activity with aging, and subsequent health outcomes. However, a consensus on proper methodological use of these devices in older populations has not been established. To date, much of the validation research regarding device type, placement, and data interpretation has been performed in younger, healthier populations, and translation of these methods to older populations remains problematic. A better understanding of these devices, their measurement properties, and the data generated is imperative to furthering our understanding of daily physical activity, its effects on the aging process, and vice versa. The purpose of this article is to provide an overview of the highlights of the preconference workshop, including properties of the different types of accelerometers, the methodological challenges of employing accelerometers in older study populations, a brief summary of ongoing aging-related research projects that utilize different types of accelerometers, and recommendations for future research directions.

Prediction of sustained harmonic walking in the free-living environment using raw accelerometry data

Article

Full-text available

May 2015

Accelerometers are now widely used to measure activity in large observational studies because they are easy to wear, measure acceleration at a resolution that humans cannot report, and measure objectively the acceleration of the particular body part they are attached to. Using raw accelerometry data we propose and validate methods for identifying and characterizing walking in the free-living environment. Although existing methods can identify walking in laboratory settings, recognizing it in the free-living environment remains a major methodological challenge. We focus on sustained harmonic walking (SHW), which we define as walking for at least 10 seconds with low variability of step frequency. Our method is robust to within- and between-subject variability, measurement configuration and device rotation. The resulting approach is fast and allows processing of a week of raw accelerometry data (~150M measurements) in half an hour on an average computer. Methods are motivated by and applied to the free-living data obtained from the Developmental Cohort Study (DECOS) where 49 elderly subjects were monitored for one week (~300 GB of data)

Quantifying the lifetime circadian rhythm of physical activity: A covariate-dependent functional approach

Article

Full-text available

Oct 2014

Objective measurement of physical activity using wearable devices such as accelerometers may provide tantalizing new insights into the association between activity and health outcomes. Accelerometers can record quasi-continuous activity information for many days and for hundreds of individuals. For example, in the Baltimore Longitudinal Study on Aging physical activity was recorded every minute for [Formula: see text] adults for an average of [Formula: see text] days per adult. An important scientific problem is to separate and quantify the systematic and random circadian patterns of physical activity as functions of time of day, age, and gender. To capture the systematic circadian pattern, we introduce a practical bivariate smoother and two crucial innovations: (i) estimating the smoothing parameter using leave-one-subject-out cross validation to account for within-subject correlation and (ii) introducing fast computational techniques that overcome problems both with the size of the data and with the cross-validation approach to smoothing. The age-dependent random patterns are analyzed by a new functional principal component analysis that incorporates both covariate dependence and multilevel structure. For the analysis, we propose a practical and very fast trivariate spline smoother to estimate covariate-dependent covariances and their spectra. Results reveal several interesting, previously unknown, circadian patterns associated with human aging and gender.

Insomnia Symptoms and Actigraph-Estimated Sleep Characteristics in a Nationally Representative Sample of Older Adults

Article

Full-text available

Sep 2014
J GERONTOL A-BIOL

Background: Reports of insomnia symptoms are common among the elderly. However, little is known about the relationship between insomnia symptoms and objective assessments of sleep in the general population of older adults. We assessed concordance between insomnia symptoms and actigraphic sleep characteristics in a nationally representative sample of older Americans. Methods: In a national probability sample of 727 adults aged 62-91 years in 2010-2011 from the National Social Life, Health, and Aging Project, respondents were asked how often they (a) feel rested when they wake up, (b) have trouble falling asleep, (c) have trouble with waking up during the night, and (d) have trouble waking up too early and not being able to fall asleep again. Responses to these questions were compared to sleep characteristics estimated from three nights of actigraphy for the same individuals. Statistical analyses were adjusted for age, gender, race and ethnicity, income, assets, and education. Results: Feeling rested (Question (a), above) was not correlated with any actigraphy-estimated sleep characteristics. Questions (b)-(d) each had several significant correlations with the actigraphy metrics, but generally not with the specific objective sleep characteristics that each question intended to reference. In some cases, the associations were not in the expected direction. Conclusions: Although three of four questions about insomnia symptoms were significantly associated with objectively estimated sleep characteristics, responses seem to be general indicators of sleep quality rather than reports of specific sleep characteristics.

Assessing the "Physical Cliff": Detailed Quantification of Age-Related Differences in Daily Patterns of Physical Activity

Article

Full-text available

Dec 2013
J GERONTOL A-BIOL

In spite of evidence that physical activity has beneficial effects on health and age-related functional decline, there is a scarcity of detailed and accurate information on objectively measured daily activity and patterns of such activity in older adults. Participants in the Baltimore Longitudinal Study of Aging (n = 611, 50% male, mean age 67, range 32-93) wore the Actiheart portable activity monitor for 7 days in the free-living environment. The association between activity and age was modeled using a continuous log-linear regression of activity counts on age with sex, body mass index, employment status, functional performance, and comorbid conditions as covariates. In the fully adjusted model, continuous analyses demonstrated that overall physical activity counts were 1.3% lower for each year increase in age. Although there were no differences among morning levels of activity, there was significantly lower afternoon and evening activity in older individuals (p < .01). After adjusting for age, poor functional performance, nonworking status, and higher body mass index were independently associated with less physical activity (p < .001). The use of accelerometers to characterize minute-by-minute intensity, cumulative physical activity counts, and daily activity patterns provides detailed data not gathered by traditional subjective methods, particularly at low levels of activity. The findings of a 1.3% decrease per year in activity from mid-to-late life, and the corresponding drop in afternoon and evening activity, provide new information that may be useful when targeting future interventions. Further, this methodology addresses essential gaps in understanding activity patterns and trends in more sedentary sectors of the population.

Normalization and extraction of interpretable metrics from raw accelerometry data

Article

Full-text available

Sep 2013
BIOSTATISTICS

We introduce an explicit set of metrics for human activity based on high-density acceleration recordings from a hip-worn tri-axial accelerometer. These metrics are based on two concepts: (i) Time Active, a measure of the length of time when activity is distinguishable from rest and (ii) AI, a measure of the relative amplitude of activity relative to rest. All measurements are normalized (have the same interpretation across subjects and days), easy to explain and implement, and reproducible across platforms and software implementations. Metrics were validated by visual inspection of results and quantitative in-lab replication studies, and by an association study with health outcomes.

Intille: Activity recognition from user-annotated acceleration data

Article

Operating Systems: Internals and Design Principles

Article

Jan 2008

W. Stallings

An artificial neural network to estimate physical activity energy expenditure and identify physical activity type from an accelerometer

Article

Aug 2009
J APPL PHYSIOL

The aim of this investigation was to develop and test two artificial neural networks (ANN) to apply to physical activity data collected with a commonly used uniaxial accelerometer. The first ANN model estimated physical activity metabolic equivalents (METs), and the second ANN identified activity type. Subjects (n = 24 men and 24 women, mean age = 35 yr) completed a menu of activities that included sedentary, light, moderate, and vigorous intensities, and each activity was performed for 10 min. There were three different activity menus, and 20 participants completed each menu. Oxygen consumption (in ml x kg(-1) x min(-1)) was measured continuously, and the average of minutes 4-9 was used to represent the oxygen cost of each activity. To calculate METs, activity oxygen consumption was divided by 3.5 ml x kg(-1) x min(-1) (1 MET). Accelerometer data were collected second by second using the Actigraph model 7164. For the analysis, we used the distribution of counts (10th, 25th, 50th, 75th, and 90th percentiles of a minute's second-by-second counts) and temporal dynamics of counts (lag, one autocorrelation) as the accelerometer feature inputs to the ANN. To examine model performance, we used the leave-one-out cross-validation technique. The ANN prediction of METs root-mean-squared error was 1.22 METs (confidence interval: 1.14-1.30). For the prediction of activity type, the ANN correctly classified activity type 88.8% of the time (confidence interval: 86.4-91.2%). Activity types were low-level activities, locomotion, vigorous sports, and household activities/other activities. This novel approach of applying ANNs for processing Actigraph accelerometer data is promising and shows that we can successfully estimate activity METs and identify activity type using ANN analytic procedures.

Development of Novel Techniques to Classify Physical Activity Mode Using Accelerometers

Article

Oct 2006

Use of accelerometers to assess physical activity (PA) is widespread in public health research, but their utility is often limited by the accuracy of data-processing techniques. We hypothesized that more sophisticated approaches to data processing could distinguish between activity types based on accelerometer data, providing a more accurate picture of PA. Using data from MTI Actigraphs worn by six subjects during four activities (walking, walking uphill, vacuuming, working at a computer), quadratic discriminant analysis (QDA) was performed, and a hidden Markov model (HMM) was trained to recognize the activities. The ability of the new analytic techniques to accurately classify PA was assessed. The mean (SE) percentage of time points for which the QDA correctly identified activity mode was 70.9% (1.2%). Computer work was correctly recognized most frequently (mean (SE) percent correct = 100% (0.01%)), followed by vacuuming (67.5% (1.5%)), uphill walking (58.2% (3.5%)), and walking (53.6% (3.3%)). The mean (SE) percentage of time points for which the HMM correctly identified activity mode was 80.8% (0.9%). Vacuuming was correctly recognized most frequently (mean (SE) percent correct = 98.8% (0.05%)), followed by computer work (97.3% (0.7%)), walking (62.6% (2.3%)), and uphill walking (62.5% (2.3%)). In contrast to a traditional method of data processing that misidentified the intensity level of 100% of the time spent vacuuming and walking uphill, the QDA and HMM approaches correctly estimated the intensity of activity 99% of the time. The novel approach of estimating activity mode, rather than activity level, may allow for more accurate field-based estimates of physical activity using accelerometer data, and this approach warrants more study in a larger and more diverse population of subjects and activities.

Accelerometry Data in Health Research: Challenges and Opportunities: Review and Examples

Abstract and Figures

Recommended publications

Accelerometry Data in Health Research: Challenges and Opportunities: Review and Examples

Validation of Gait Characteristics Extracted From Raw Accelerometry During Walking Against Measures...

Adaptive empirical pattern transformation (ADEPT) with application to walking stride segmentation

Automatic car driving detection using raw accelerometry data