Content uploaded by Marco Altini
Author content
All content in this area was uploaded by Marco Altini on Jan 06, 2017
Content may be subject to copyright.
Variable-Length Accelerometer Features and Electromyography to
Improve Accuracy of Fetal Kicks Detection During Pregnancy Using a
Single Wearable Device
Marco Altini1, Elisa Rossetti2, Michiel Rooijakkers2, Julien Penders1,
Dorien Lanssens3, Lars Grieten3and Wilfried Gyselaers3
Abstract— In this paper, we propose a method to improve
accuracy of fetal kicks detection during pregnancy using a
single wearable device placed on the abdomen. Monitoring
fetal wellbeing is key in modern obstetrics as it is routinely
used as a proxy to fetal movement. However, accurate, nonin-
vasive, long-term monitoring of fetal movement is challenging,
especially outside hospital environments. A few accelerometer-
based systems have been developed in the past few years, to
tackle common issues in ultrasound measurement and enable
remote, self-administrated monitoring of fetal movement. These
solutions typically consist in multi-accelerometer systems of
limited practical applicability, relying on simple features such
as the signal magnitude. In this paper, we propose two methods
to improve accuracy of fetal kicks detection using a single
wearable device, in particular aiming at reducing false positives
and increasing positive predictive value (PPV) when lacking a
reference accelerometer outside the abdominal area. Firstly,
we propose variable-length accelerometer features. Secondly,
we combine accelerometer data with electromyography (EMG).
Both the proposed techniques aim at providing more contextual
information related to maternal movement while still using a
single wearable device. We compare our method to a system
comprising 6accelerometer sensors over a dataset including
22 recordings and reference maternal annotations, highlighting
how kicks detection PPV can be improved by up to 10%
when including variable-length features and up to 11% when
including EMG features.
I. INTRODUCTION AND RELATED WORKS
Monitoring fetal movement during pregnancy is the most
practical and widespread method to assess fetal wellbeing,
one of the most important and complex tasks of modern
obstetrics. As birth outcomes are strongly linked to the
development of fetal conditions during pregnancy [1], several
techniques have been developed to monitor fetal movement
up to date [2].
Some methods require hospital stays or trained personnel,
for example ultrasound, relying on high frequency sound
waves being used to generate an image of the fetus and
can be used only for a limited amount of time due to
safety concerns [3], [4]. Other methods, such as continu-
ous cardiotocography, require cumbersome infrastructure and
This work was funded by Bloomlife
1M. Altini and J. Penders are with Bloomlife, San Francisco, USA
altini.marco@gmail.com
2M. Rooijakkers and E. Rossetti are with Bloom Technologies, Genk,
BE
3D. Lanssens, L. Grieten and W. Gyselaers are with the Department of
Future Health, Ziekenhuis Oost-Limburg, Genk, BE
hospital visits, also involving trained personnel to set up the
device and process the produced information [5], [6].
Thus the inability of these methods to monitor fetal
movement outside of sporadic spot checks in the hospital
environment is one of the major causes of concern and
motivations behind the development of other passive methods
for home-monitoring, such as accelerometer based solutions.
Wearable devices including on-board accelerometers pro-
vide new opportunities to investigate passively and safely
fetal movement inside [7] or outside [8] the hospital. Most
studies to date involved one single accelerometer placed on
the abdomen and reported rather low sensitivity and speci-
ficity [9]. Other researchers added a reference accelerometer
with the rationale that by monitoring maternal movement
artifacts using an accelerometer placed outside of the ab-
dominal area, fetal movement should be separable from
maternal movement and therefore detected more accurately
[8]. However, most solutions limit practical applicability as
additional sensors are placed far from the abdomen, given the
inability of additional accelerometers placed on the upper
thoracic area to discriminate between maternal and fetal
movement [10]. As a result, systems are often bulky and
cumbersome [11].
Recently, a few authors introduced machine learning tech-
niques to classify a set of features into a binary problem
(movement vs no-movement), obtaining promising results
[12], [13], [11], [14]. In the context of using supervised
learning methods to classify movements and non-movements,
an additional challenge arises. Fetal movement occurs only
for a small percentage of the time during a measurement,
therefore proper methods such as downsampling of the
majority class (i.e. no-movement) need to be employed [11].
In our previous work [11] we used Random Forests and
a set of time domain features computed over short windows
(0.5 seconds) and highlighted consistent improvements when
including a reference accelerometer on the back. A reference
accelerometer typically provides advantages in terms of
reducing false positives as maternal movement can be better
discriminated from fetal movement, as shown in Fig. 1.
In this paper, we hypothesized that employing variable-
length accelerometer features and sensor-fusion techniques
we could reduce false positives for prediction models derived
using a single wearable device. Our assumption is based
on exploiting the different patterns and time dynamics of
maternal and fetal movements.
0
100
200
0 5 10 15
Timestamp (minutes)
Motion intensity
label
nothing
kick
Multi − data sample accelerometer 1
0
50
100
150
200
0 5 10 15
Timestamp (minutes)
Motion intensity
Multi − data sample accelerometer 3
0
10
20
30
40
50
0 5 10 15
Timestamp (minutes)
Motion intensity
Reference accelerometer on the back − 6
0.0
0.1
0.2
0.3
0 5 10 15
Timestamp (minutes)
Motion intensity
label
nothing
kick
Single sensor, short time window
0.0
0.1
0.2
0 5 10 15
Timestamp (minutes)
Motion intensity
Single sensor, long time window
0
100
200
300
0 5 10 15
Timestamp (minutes)
EHG intensity
Single sensor, EHG data
Fig. 1. Left side plots: motion intensity (mean of the band passed
accelerometer signal summed over the three axes) for the TMSi device,
comprising 6accelerometers (only 3shown). Fetal kicks as manually
annotated by the subject are highlighted. We can clearly see spikes due
to maternal movement appearing on the reference accelerometer on the
back, number 6. On the other hand, fetal kicks are visible only on the
accelerometers on the abdomen, with similar accelerometer patterns across
sensor locations. Right side plots: motion intensity and EMG data for
the Bloomlife device, computed over variable-length time windows. Fetal
kicks are visible only for short time windows (0.5seconds) while maternal
movement artifacts are propagated on long time windows as well, hence we
hypothesized false positives could be reduced by including variable-length
features.
Thus, practical applicability and ease of use in home
settings could be improved without sacrificing accuracy.
The proposed techniques aim at reducing false positives by
providing more contextual information related to maternal
movement while still using a single wearable device to cope
with the absence of a reference accelerometer or a more
obtrusive system. We compare our method to a system com-
prising 6accelerometer sensors over a dataset including 22
recordings with reference maternal annotations, highlighting
how kicks detection PPV can be improved by up to 10%
when including variable-length features and up to 11% when
including EMG features.
II. DATA ACQUISITION
A. Accelerometers Data and Reference
Twenty-two recordings of about 20 minutes duration were
collected from 22 pregnant women at different gestational
ages during pregnancy, all from week 30 onwards. Fetal
movements per 20 minutes measurements were 34 ±68,
ranging between 0for inactive babies to 315 for hiccups
cases. Measurements were performed using two devices.
Firstly, we used a research version of the Bloomlife wearable
device [15], configured to acquire two channels EMG at
4096 Hz and triaxial accelerometer data at 128 Hz from a
Fig. 2. On-body accelerometers placement for the TMSi device: 5
accelerometers placed on the abdomen, while the sixth accelerometer,
placed on the back, is not visible. Bloomlife sensor placed below TMSi
accelerometer 1, capturing accelerometer and EMG data. Also visible are
TMSi electrodes used to acquire ExG data, not used in this study.
single accelerometer placed on the abdomen (see Fig. 2). The
Bloomlife wearable sensor was attached to the skin using a
medical grade adhesive patch. Secondly, we used the Porti7
device from Twente Medical Systems International (TMSi)
as a multi-accelerometer device for comparison. Accelerom-
eter data were bandpass filtered between 1and 20 Hz with
a second order butterworth IIR filter since fetal movement
is expected to be in this frequency band [12]. EMG data
were bandpassed filtered between 0.1 and 3 Hz to capture
low frequency maternal muscle activity on the abdomen. Five
accelerometer sensors of the TMSi device were positioned
on the abdomen with the navel serving as central marker. The
sixth sensor was placed on the back as reference for maternal
movement (see Fig. 2). All subjects were lying down in a
hospital bed and given a handheld toggle which they were
advised to press when feeling fetal movement. The output of
the button was always used as reference for fetal movements.
The experimenter manually annotated fetal movements as a
pre-processing step, by locating accelerometer movements
anticipating button triggers.
III. DATA ANALYS IS
A. Features
We computed the same feature sets for the TMSi (case
Multi) and Bloomlife (case Single) devices. To account for
different dynamics in maternal and fetal movement, we
computed features over two time windows of 0.5and 4
seconds. The rationale is that short fetal movements should
be averaged out over longer time windows but captured over
short ones, while maternal movements should appear over
windows of both durations. A time window of 0.5 seconds
was chosen as fetal kicks result in accelerations of short
duration, typically around 0.5 seconds, as shown in Fig.
1 [11]. On the other hand, the longer window was set to
4seconds as this duration is long enough to average out
accelerations due to fetal kicks (see Fig. 1) while being
short enough to limit processing delays and do not cause
maternal movements to impact algorithm output for longer
periods of time.Thus, variable-length features should reduce
false positives (see Fig. 1). We computed low-complexity
time domain features to possibly enable easy implementation
on an embedded device. Features were: mean, standard
deviation, interquartile range, correlation between axis, sum,
min, max and magnitude. Each feature was computed per
axis, per sensor and per window size. EMG features were
computed for Single and were the same time domain statistics
listed above for accelerometer data, computed over 4seconds
windows only. EMG is representative of muscular activity
on the abdomen and therefore, similarly to accelerometer
data, can capture maternal movement, albeit with slightly
different dynamics as the signal is typically slower than
almost instantaneous accelerometer spikes.
B. Features Selection, Class Imbalance and Classification
Classification was performed using random forests. Fea-
tures were not selected before classification, as random
forests can pick a subset of the available features at each
iteration. In particular, we set the number of features to
select at each iteration to the square root of the total number
of features, therefore potentially maintaining all information
at training phase, with respect to other features selection
methods. Given the small number of kicks with respect to the
total available data, class imbalance needs to be addressed
as well. Similarly to features selection, we let the random
forests classifier pick a subset of samples during training. The
optimal ratio between reference class (kicks) and majority
class (non-kicks) was determined by cross-validating and
optimizing for F-score, i.e. choosing the ratio that showed
optimal F-score. Our optimal balance included all data from
the minority class and one fifth of the majority class data.
C. Performance Metrics and Validation Method
We compared four feature sets associated to the two
systems used in this study in order to highlight the impact
of the novel methods proposed to improve accuracy of a
single wearable device in detecting fetal kicks by reducing
false positives. In particular, we compare; 1) TMSi (6 ac-
celerometer system) and variable-length features (Multi SL)
2) Bloomlife (single wearable sensor) and features computed
over a short (Single S) time window only 3) Bloomlife and
features computed over both short and long time windows
(Single SL) 4) Bloomlife and features computed over both
short and long time windows plus EMG features (Single
SLE). All models were derived and validated using leave one
participant out cross-validation and a binary classification
problem distinguishing fetal kicks and non-fetal kicks (e.g.
non-movement, noise, etc.). Given the binary classification
problem and data imbalance, we chose Sensitivity and PPV
as two metrics representative of the ability of the algorithms
to detect sporadic fetal kicks. Performance metrics were
determined according to the strategy depicted in Fig. 3 and
computed on the entire data stream for all participants during
cross-validation.
Fig. 3. Graphical example of our evaluation strategy. TP = true positive,
FN = false negative, FP = false positive.
Finally, we also provide plots on the relation between
detected and annotated fetal kicks over entire recordings for
each participant.
IV. RESULTS
Fig. 4 shows Sensitivity and PPV results for the four
models compared in this study; 1) TMSi (6 accelerometer
system) and variable-length features (Multi SL) 2) Bloomlife
(single wearable sensor) and features computed over short
(Single S) time window only 3) Bloomlife and features
computed over both short and long time windows (Single SL)
4) Bloomlife and features computed over both short and long
time windows plus EMG features (Single SLE). In particular,
Sensitivity was 0.74 for Multi SL,0.65 for Single S and
0.64 for Single SL and Single SLE. As expected Sensitivity
does not change much by introducing variable-length and
EMG features as the aim of these features is to reduce false
positives. On the other hand, PPV was 0.75 for Multi SL,0.65
for Single S,0.75 for Single SL and 0.76 for Single SLE,
highlighting how variable-length and EMG features could
consistently reduce false positives and increase PPV to the
same or higher values with respect to those obtained using
a6accelerometers system (Multi SL).
0
25
50
75
Multi SL Single S Single SLSingle SLE
Model
Percentage (%)
Sensitivity
0
25
50
75
Multi SL Single S Single SLSingle SLE
Model
Percentage (%)
PPV
Fig. 4. Sensitivity and PPV for the different models compared in this
study. Inclusion of features computed over longer time windows and EMG
data (case Single SLE) increases PPV (due to reduced false positives) to the
same levels shown for a multi sensor system consisting of 6accelerometers
(case Multi SL).
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
0
100
200
300
0 100 200 300
Detected kicks
Actual kicks
Actual vs Detected kicks, Multi SL
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
0
100
200
300
0 100 200 300
Detected kicks
Actual kicks
Actual vs Detected kicks, Single S
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●●
●
●
0
100
200
300
0 100 200 300
Detected kicks
Actual kicks
Actual vs Detected kicks, Single SL
●
●
●●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
0
100
200
300
0 100 200 300
Detected kicks
Actual kicks
Actual vs Detected kicks, Single SLE
Fig. 5. Relation between actual and detected kicks over each 20 minutes
recording for the different models compared in this study.
Finally, we report results on the relation between the
actual and detected total number of kicks over each entire
20-minute recording, Fig. 5, showing increased agreement
between the reference method (manual annotations) and the
models estimates as we add features, from Single S to Single
SLE, also for the number of kicks detected over entire
recordings.
V. DISCUSSION AND CONCLUSIONS
In this paper we proposed a method to improve the
accuracy of fetal kicks detection during pregnancy using
a single wearable device placed on the abdomen. The re-
sults obtained confirm our assumptions highlighted in Fig.
1. In particular, by including variable-length accelerometer
features, short fetal movement is averaged out over longer
time windows but captured over short ones, while maternal
movements of greater intensity appear over windows of both
durations. As a result, a single wearable device can be used
to better discriminate fetal and maternal movement without
the need for a reference accelerometer. Additionally, as EMG
is representative of muscular activity on the abdomen, it can
capture maternal movement, albeit with slightly different dy-
namics. We showed that the proposed method was effective
in reducing false positives by increasing PPV for a single
sensor device to the same levels obtained with a cumbersome
6sensors system (11% improvement).
One of the main limitations of this study is the use of
maternal annotations as reference as there is no trustworthy
reference for fetal movement. While ultrasound is the clinical
standard for fetal movement, limitations apply, even during
research studies. For example, with fetal growth it becomes
impossible to fully display the fetus given the limited field
of vision of the ultrasound probe, starting at approximately
week 20. While this is not a problem during hospital check-
ups, moving and re-positioning the probe while trying to
measure small accelerations as reflected on the pregnant
women abdomen is impractical and can easily introduce
noise. By analyzing the algorithms performance and trade
offs with respect to the same reference, we could get a
better understanding of the influence of different techniques
in effectively detecting fetal movement.
Finally, while false positive detection were reduced suc-
cessfully by the techniques employed, Sensitivity did not
improve using variable-length and EMG features. We spec-
ulate that reduced sensitivity for Single is due to lack of
spacial resolution as fetal kicks could be localized in specific
areas of abdomen. A practical solution to this issue could be
placement of the single wearable device where movement is
typically felt by the pregnant woman. Future work will focus
on some of these aspects as well as the possibility to include
different types of fetal movements.
REFERENCES
[1] E. Symonds, “On-line processing of the fetal electrocardiogram. a new
direction for fetal monitoring.” The Journal of reproductive medicine,
vol. 32, no. 7, pp. 509–512, 1987.
[2] H. P. van Geijn, “2 developments in ctg analysis,” Bailli`
ere’s clinical
obstetrics and gynaecology, vol. 10, no. 2, pp. 185–209, 1996.
[3] K. ˚
A. Salvesen, “Efsumb: safety tutorial: epidemiology of diagnos-
tic ultrasound exposure during pregnancy?european committee for
medical ultrasound safety (ecmus),” European journal of ultrasound,
vol. 15, no. 3, pp. 165–171, 2002.
[4] E. Sheiner, I. Shoham-Vardi, and J. S. Abramowicz, “What do clinical
users know regarding safety of ultrasound during pregnancy?” Journal
of ultrasound in medicine, vol. 26, no. 3, pp. 319–325, 2007.
[5] Z. Alfirevic, D. Devane, G. Gyte et al., “Continuous cardiotocography
(ctg) as a form of electronic fetal monitoring (efm) for fetal assessment
during labour,” Cochrane Database Syst Rev, vol. 3, no. 3, 2006.
[6] R. M. Grivell, Z. Alfirevic, G. Gyte, and D. Devane, “Antenatal
cardiotocography for fetal assessment,” Cochrane Database Syst Rev,
vol. 1, 2010.
[7] B. Boashash, M. S. Khlif, T. Ben-Jabeur, C. E. East, and P. B.
Colditz, “Passive detection of accelerometer-recorded fetal movements
using a time–frequency signal processing approach,” Digital Signal
Processing, vol. 25, pp. 134–155, 2014.
[8] K. Nishihara, N. Ohki, H. Kamata, E. Ryo, and S. Horiuchi, “Auto-
mated software analysis of fetal movement recorded during a pregnant
woman?s sleep at home,” PloS one, vol. 10, no. 6, p. e0130503, 2015.
[9] G. Thomas, O. T. John, M. Mostefa, B. Boualem, C. Ian, W. Stephen,
F. Miguel, C. Susan, and C. Paul, “Detecting fetal movements using
non-invasive accelerometers: A preliminary analysis,” in Information
Sciences Signal Processing and their Applications (ISSPA), 2010 10th
International Conference on. IEEE, 2010, pp. 508–511.
[10] M. S. H. Khlif, B. Boashash, S. Layeghy, T. Ben-Jabeur, P. B. Colditz,
and C. East, “A passive dsp approach to fetal movement detection for
monitoring fetal health,” in Information Science, Signal Processing
and their Applications (ISSPA), 2012 11th International Conference
on. IEEE, 2012, pp. 71–76.
[11] M. Altini, P. Mullan, M. Rooijakkers, S. Gradl, J. Penders, N. Geusens,
L. Grieten, and B. Eskofier, “Detection of fetal kicks using body-worn
accelerometers during pregnancy: Trade-offs between sensors number
and positioning,” pp. 5319–5322, Aug 2016.
[12] S. Layeghy, G. Azemi, P. Colditz, and B. Boashash, “Non-
invasivemonitoring of fetal movements using time-frequency fea-
tures of accelerometry,” in Acoustics, Speech and Signal Processing
(ICASSP), 2014 IEEE International Conference on. IEEE, 2014, pp.
4379–4383.
[13] ——, “Classification of fetal movement accelerometry through time-
frequency features,” in Signal Processing and Communication Systems
(ICSPCS), 2014 8th International Conference on. IEEE, 2014, pp.
1–6.
[14] L. Minjie, T. Yongkang, L. Yunfeng, Y. Song, and D. Jingxin, “Fetal
movement detection based on mems accelerometer,” une, vol. 13,
p. 15, 2016.
[15] “Bloomlife company website,” https://bloomlife.com/, accessed: 2016-
11-22.