Assessment_of_Fetal_and_Maternal_Well-Being_During_Pregnancy_Using_Passive_Wearable_Inertial_Sensor.pdf

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 71, 2022 4005111
Assessment of Fetal and Maternal Well-Being
During Pregnancy Using Passive Wearable
Inertial Sensor
Eranda Somathilake , Upekha Hansanie Delay , Janith Bandara Senanayaka , Samitha Lakmal Gunarathne ,
Roshan Indika Godaliyadda , Senior Member, IEEE, Mervyn Parakrama Ekanayake , Senior Member, IEEE,
Janaka Wijayakulasooriya , Member, IEEE, and Chathura Rathnayake
Abstract—Assessing the health of both the fetus and mother
is vital in preventing and identifying possible complications in
pregnancy. This article focuses on a device that can be used
effectively by the mother herself with minimal supervision and
provide a reasonable estimation of fetal and maternal health
while being safe, comfortable, and easy to use. The device
proposed uses a belt with a single accelerometer over the mother’s
uterus to record the required information. The device is expected
to monitor both the mother and the fetus constantly over a long
period and provide medical professionals with useful information,
which they would otherwise overlook due to the low frequency
that health monitoring is carried out at the present. The article
shows that simultaneous measurement of respiratory information
of the mother and fetal movement is in fact possible even in the
presence of mild interferences, which needs to be accounted for
if the device is expected to be worn for extended times.
Index Terms—Accelerometers, breathing patterns, deep learn-
ing, fast Fourier transform, fetal health, fetal movement, wavelet
transform, wiener filtering.
I. INTRODUCTION
REGULAR monitoring throughout the pregnancy allows
early detection of well-being problems that might arise
and will aid their treatment, improving the chance for the birth
of a healthy baby. The health and condition of both the mother
and the baby is a clear indication of future complications or the
well-being of the fetus [1]. Fetal well-being can be monitored
in different ways [2]–[4], each with its own advantages and
disadvantages. Hence, the availability of multiple methods will
provide the medical professionals with better tools, which
can be used in different specific situations. Also, the mood
and health of the mother can have a dramatic impact on the
Manuscript received February 16, 2022; revised April 2, 2022; accepted
April 20, 2022. Date of publication May 13, 2022; date of current version
May 26, 2022. The Associate Editor coordinating the review process was
Dr. Chao Tan. (Corresponding author: Eranda Somathilake.)
Eranda Somathilake is with the Department of Mechanical Engineer-
ing, University of Peradeniya, Peradeniya 20400, Sri Lanka (e-mail:
eranda.somathilake@eng.pdn.ac.lk).
Upekha Hansanie Delay, Janith Bandara Senanayaka, Samitha Lakmal
Gunarathne, Roshan Indika Godaliyadda, Mervyn Parakrama Ekanayake, and
Janaka Wijayakulasooriya are with the Department of Electrical and Electronic
Engineering, University of Peradeniya, Peradeniya 20400, Sri Lanka.
Samitha Lakmal Gunarathne is with the Department of Electrical and Elec-
tronic Engineering, University of Peradeniya, Peradeniya 20400, Sri Lanka.
Chathura Rathnayake is with the Department of Obstetrics and Gynaecol-
ogy, University of Peradeniya, Peradeniya 20400, Sri Lanka.
Digital Object Identifier 10.1109/TIM.2022.3175041
development of the baby, both in long and short terms [5]. This
article, therefore, proposes a device to assess the condition of
the fetus as well as the mother, which is noninvasive, low cost,
and simpler to use compared to most of the existing methods
available for fetal condition monitoring.
Fetal condition monitoring can be done in multiple ways and
the area of focus in this research is through fetal movement,
which can be used as an indication of future complications [6].
Fetal movement can be identified as a primary indicator of
fetal well-being. Reduction or absence of fetal movement is a
strong indication of fetal compromise [7], [8]. Fetal movement
monitoring methods can be divided into two approaches: active
and passive. Active methods directly observe the fetus using
various imaging techniques, while passive methods measure
the fetal movements indirectly by measuring other responses
as the movements of the fetus. Cardiotocography (CTG) and
ultrasound scanning are examples of active methods, whereas
the use of sensors such as accelerometers or acoustic sensors
is an example of passive methods.
While most of the common active methods such as ultra-
sound and CTG are used extensively, they do possess some
drawbacks that can make them undesirable in certain sit-
uations. Although it is not proved clinically, the use of
high-frequency audio waves that penetrates the uterus may
cause harm to the fetus [9]. Also, equipment used for both
CTG scanning and ultrasound scanning are bulky and require
trained professionals for operation and interpretation, thus
making them impractical to be used on a daily basis over
an extended period of time. Fetal state monitoring using these
methods is only conducted in clinical settings and most of
the time is done after the mother’s admittance to the hospital.
It would, however, be beneficial if fetal movements can be
monitored domestically in an ambulatory setting, which will,
in turn, enable the assessments to be done more frequently
and is ideal in times of a pandemic, where the mothers
can safely stay in their homes. The most common fetal
movement measurement method that is currently used outside
the intervention of medical staff and complicated equipment
is the manual kick counting from the mother. But as stated
in [8], this is unreliable and not accurate and requires a more
reliable alternative. In addition, it is difficult to obtain images
of obese mothers, which makes these methods less viable to
1557-9662 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Taipei Univ. of Technology. Downloaded on November 19,2022 at 09:24:41 UTC from IEEE Xplore. Restrictions apply.

4005111 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 71, 2022
be used Benacerraf [10] even though they are more likely
to encounter complications [11]. The lack of experts and the
expensive equipment required, in rural areas, can be a major
hindrance [12] in monitoring fetal movement, thus promoting
the need for low cost and simpler methods.
Simpler passive methods, therefore, can be used as an
alternative or along with existing active methods to have a
better assessment of fetal movement throughout the preg-
nancy. The proposed device detects fetal movement through an
accelerometer. Obtaining the data through this device is more
convenient, noninvasive, and takes minimal time to set up. This
device is designed to be portable, ergonomic, and with modular
electronics for easy repair. As suggested by clinicians, these
design features make the device desirable for both medical
staff and the mother due to its practical applicability and ease
of use.
Furthermore, it was observed that the accelerometer placed
on the abdomen has the ability to capture maternal respiratory
motion data as well. It can, therefore, be used to detect respi-
ratory patterns [13], which can provide valuable information
such as respiratory rate and maternal energy expenditure (EE)
level to medical practitioners. Both respiratory rate [14] and
maternal EE [15] are key indicators of maternal well-being
during pregnancy. Hence, this device can be used as an overall
condition monitoring device rather than solely focusing on
fetal movement.
The data obtained using the sensors consist of information
on many processes that happens in the body, namely fetal
movements, maternal respiratory patterns, and disturbances
such as coughing, laughing, or walking. In this study, the
desired parameters were maternal respiratory motions and fetal
movements. Therefore, analyses were conducted to identify
fetal movements and to extract information on maternal res-
piratory data.
This device is designed to be used by lay users, and the
readings will not be taken in a controlled environment. Hence,
the device should have the capability to identify and filter out
unnecessary data. Therefore, in this study, it is demonstrated
that the artifacts introduced into the accelerometric signal due
to the walking during a session can be successfully removed
using a Wiener filter. Moreover, a peak detecting algorithm
was used to calculate the respiratory rate of the mothers, which
can act as one of the main well-being indicators for maternal
health [16]. Furthermore, analyses were conducted to classify
the EE level of the mother’s body, depending on the respiratory
pattern. To accomplish that, feature extraction was done using
the discrete wavelet transform (DWT) and then classified using
a neural network.
Fetal movement patterns can be complex in nature and
hence machine-learning techniques were used to identify
them [17], [18]. One approach was to use convolutional
neural networks (CNNs), where a scalogram was generated
by taking the wavelet transform of the signal and was
then used to identify fetal movement. The other approach
was to use a recurrent neural network (RNN), where gated
recurrent unit (GRU) cells were used so that the signal
can be analyzed over long-term dependencies to detect fetal
movement.
II. BACKGROUND
Active methods such as ultrasound scanning and CTG
scanning observe the movement of the fetus and give a phys-
ical representation of it. They are, therefore, highly accurate
and experts can assess fetal health fairly accurately [1]. The
negative aspects of these devices, as stated previously, are then
too complex to use, bulky, and might have harmful effects
on both the mother and the baby. Furthermore, interpreting
the data received from these methods requires technical skill
as well as time, which can act as a hindrance to taking
quick actions when necessary. Therefore, passive methods that
observe the surface of the uterus were considered in this study.
Also, active methods do not evaluate fetal movement con-
stantly, but identification of changes in fetal movement patterns
over an extended time period can help identify complications
earlier [8].
Maternal perception of movement identification and mon-
itoring is a widely used and easy-to-implement method but
it is highly subjective and inconsistent [19]. Hence, there is
a need for a viable fetal movement monitoring method that
eliminates these inaccuracies. One such method proposes a
device that uses acoustic sensors and accelerometers attached
to a belt worn around the abdomen [20]. Also, multiple
accelerometer sensors were used in a different study [21], [22].
These multiple sensor approaches have yielded good results.
However, as stated in [20], acoustic sensors are too sensitive
for this particular problem. The use of multiple accelerometers,
although is better than one, did not seem to have a significant
effect on movement detection, and increasing the sensors
had a diminishing impact on the results while it led to an
increase in used equipment and complexity. Hence, in this
study, a single sensor was used and more attention was paid
to optimizing the postprocessing of the data. Most of the
proposed devices mentioned above have performed well in
an experimental environment. However, their capability to be
adapted to practical situations where they are handled by users
who are not technically trained was not considered in most
cases. The proposed device in this study has been designed
for practical implementation with ease of use and simplicity
in mind.
Since inertial sensors measure the physical movement of
the surface of the abdomen, in addition to capturing fetal
movement data, they will simultaneously record other bodily
functions of the mother as well. Hence, the device can observe
multiple states of the mother. In this study, we have focused on
monitoring the breathing patterns of the mother using the same
inertial sensor. The use of accelerometers to measure breathing
patterns is a viable solution and it has been implemented to
classify different breathing patterns [13]. A similar approach
was considered in this study to evaluate the breathing patterns
of pregnant mothers to evaluate their health.
III. DEVICE IMPLEMENTED
The device implemented consists of an accelerometer and
two buttons to get external inputs. One input was used by the
doctor who observes the fetus using ultrasound scanning to
be used as the ground truth to train the neural network, while

SOMATHILAKE et al.: ASSESSMENT OF FETAL AND MATERNAL WELL-BEING DURING PREGNANCY 4005111
Fig. 1. Component configuration of the proposed device.
Fig. 2. Mother wearing the sensor.
the other button is used to record other interference that might
be detected by the accelerometer such as the mother laughing
or coughing. The same device was used in separate sessions
to record breathing patterns and this did not require any user
input while recording. The configuration of the components
of the device is given in Fig. 1.
The components are controlled by an Arduino Uno micro-
controller, as stated, and the device consists of an accelerome-
ter (MPU 9250) and two buttons. In addition, there is a micro
SD card module to store data, a real-time clock (DS3231), and
LEDs to indicate the current state-of-the-art device and button
presses. The accelerometer was attached to a belt which can be
wound around the mother’s uterus as shown in figure Fig. 2.
In this preliminary study, the device was controlled by a PC
to give better control over its operation.
The sensor used here, MPU 9250, consists of a tri-axial
accelerometer, triaxial gyroscope, and a compass. Only one
axis of the tri-axial accelerometer was used in this analysis.
The sensor communicates with the Arduino using the I2
C
interface up to a maximum sampling rate of 32 kHz, but the
readings obtained were at a sampling rate of around 280 Hz
mainly due to the other processes running in the microcon-
troller. The specified temperature range for the operation of
the sensor is from −40 ◦
C to 85 ◦
C, which accommodates the
operating temperatures for this application.
The device operates in two states: the training state and
the prediction state. In the training state, it records both
the sensor readings and the user inputs so that they can be
Fig. 3. Summarized description of the device.
Fig. 4. Different representation of the data obtained while the device is at
an stationery position.
used for creating the predictor models; in the prediction state,
it reads sensor readings and uses the predictor models to make
predictions, as depicted in Fig. 3.
IV. DATA ANALYSIS
A. Preliminary Readings
Following the fabrication of the system, several prelim-
inary tests were conducted on the system to identify its
performance in different environmental conditions as well as
how it responds to different behaviors of pregnant mothers.
Initially, the noise profile of the device is obtained. This is
done by taking readings from the device, while it is in a
stationary position. The time-domain accelerometric readings
of a single axis, as well as the time-frequency variation of
the readings, can be observed in Fig. 4(a) and (b). Moreover,
the probability distribution function (PDF) of the time-domain
data and the power spectrum of the data can be observed
in Fig. 4(c) and (d), respectively. Furthermore, the kurtosis
values of several samples were evaluated and the mean of the
values was 2.9854, which is approximately 3. Therefore, it can
be observed that the stationery noise has a behavior similar to
Gaussian White noise.
The environmental temperature in the region where the
studies were conducted varied from 16 ◦
C to 32 ◦
C. However,
the usual body temperature is 37 ◦
C. Hence, when the sensor
is worn for an extended time, the temperature could vary from
16 ◦
C to 37 ◦
C. In order to observe the effect of temperature
on the noise features of the sensor, readings on different
temperatures within the range of 16 ◦
C–37 ◦
C were taken.
The effect of the temperature cannot be clearly observed in

Fig. 5. Variation of noise power in different temperature levels.
Fig. 6. When the device is worn by a stationery subject.
the time domain. Therefore, the noise power level at each
temperature was calculated and plotted against the temperature
to observe the effects. It can be observed in Fig. 5 that when
the temperature is increased, the noise power also increases
gradually.
Then the device was worn by a nonpregnant subject, and
readings were taken to observe the sensor response to different
types of human behavior. Initially, the subject was advised
to wear the belt, to stay seated in the Fowler’s position
for an extended time, and the tri-axial accelerometric data
were recorded (see Fig. 6). It can be observed that the
subject’s breathing pattern can also be observed very clearly
from the accelerometric signals. While the X-axis and Y-axis
show a slight pattern, the Z-axis indicates a clear pattern for
the subject’s breathing. Hence, it can be concluded that this
body-worn accelerometer can also be utilized to monitor the
subject’s breathing patterns and further breathing anomalies.
Furthermore, the effect of the position of the subject on the
readings was observed by taking readings while the subject
was in different positions such as the Fowler’s position, supine
position, and lateral recumbent position. However, it was
observed that the initial position of the subject does not have
a significant effect on the data.
However, since the main aim of this research is to conduct a
preliminary study on the use of an accelerometric sensor-based
system to monitor fetal and maternal health in the home, the
effect of maternal movements on the sensor data was also
observed. Initially, the effect on time-domain data when the
subject was changing from one position to another position
was observed. The subject was advised to change into different
positions and the tri-axial accelerometric data were observed.
Different transitions have different effects on each axis’s
readings. However, in all cases, it was observed that during the
transition, there is a shift in the time-domain data. It was also
observed that following the transition period, the time-domain
data have a similar variation to time-domain data before the
transition. This can be observed in Fig. 7, where the Z-axis
Fig. 7. Acceleration variation of the Z-axis when the subject transit from
Fowler’s position to the supine position.
Fig. 8. Tri-axial acceleration variation when the subject is walking while
wearing the device.
variation when the subject is moving from Fowler’s position to
the supine position is depicted. Therefore, it can be concluded
that such small movements (in time scale) of the subject will
not have a significant effect on a session. The observed shift
can be easily eliminated by utilizing simple signal processing
techniques in the preprocessing stage. However, if a fetal
movement is to coincide with a maternal movement, the
extraction of the fetal movement signal may be strenuous.
Nevertheless, due to the short duration of such activities,
in practical situations, the probability of such coinciding
occurrences can be considered to be very small.
Thereafter, the effect of frequent maternal movements such
as walking and talking was observed. The readings taken
while the subject is walking can be observed in Fig. 8. It can
be observed that taking steps have a direct impact on the
time-domain data of all the axes. As discussed in the latter part
of this article, filtering methods such as Wiener filters can be
applied to remove the imposed interference due to taking steps.
However, it is advisable for mothers to stay stationary during
a session. Furthermore, the time-domain data obtained during
the subject’s speaking was compared with the data obtained
while not speaking. No noticeable differences were observed.
Therefore, it was concluded that speech has an insignificant
effect on the accelerometric data.
Subsequently, the time-domain accelerometric variation of
several maternal motions such as cough, hiccup, yawn, and
laugh was observed. These were selected due to the fact that
they may have a similar effect on the maternal abdomen
surface as a fetal movement. It can be observed in Fig. 9 that
while cough, laugh, and hiccups have peaks similar to the
fetal movement signal in the time domain, yawns do not have
a similar time-domain profile. Hence, during analysis, more
emphasis must be paid to extracting fetal movement signals

Fig. 9. Acceleration variation due to cough, hiccup, and yawn.
from data contaminated with a maternal laugh, hiccups, and
cough.
B. Respiratory Data Analysis
When observing the accelerometric data, it was noted that
maternal respiratory movements are also present in the data.
Furthermore, during the initial analysis of the device, respira-
tory motion was observed clearly as can be seen in Fig. 10.
In Fig. 10, the time-domain data of the Z-axis is given and
the general trend of the data is indicated in red color.
Respiratory rate is one of the major well-being indicators
of the human body. While the respiratory rate is usually over-
looked, documenting the respiratory rate can aid in predicting
several serious clinical events [14]. Therefore, the viability
of using this device to identify the respiratory rate was also
studied.
In Fig. 6, it can be observed that respiratory motion causes
noticeable peaks in the Z-axis data in the time domain.
Therefore, it was decided to monitor these occurrences of
peaks in the time domain in order to identify the respiratory
rate. Initially, a detrending algorithm was applied to the data
to remove the mean as well as the trend in time domain data.
In this step, a second-degree polynomial trend was identified
in the data and then deducted. The degree of the estimated
polynomial was selected to be second order to match the
general, most common accelerometric data trend. The estima-
tion of the polynomial was done using the native polynomial
estimation tool in MATLAB (R2018a). In Figs. 6 and 10, it can
be observed that the respiratory motion has a magnitude of
approximately 50 accelerometer measurement units (AMUs)
and the peaks are approximately 600 samples apart. This
reading is taken while the subject was wearing the device
at a stationary position. Further readings were taken after
conducting two physical activities: a 10-min walk and a
10-min run. In all these readings, the magnitude of the
respiratory motion was always less than 200 AMU and the
peaks were more than 400 samples apart. Using this infor-
mation, a peak identifying algorithm was implemented. This
algorithm detects local minima by differencing and is the
Native peak detection algorithm found in MATLAB (R2018a).
The algorithm finds the peaks of the given data by observing
the variation of the gradients (gradients should change from
positive to negative). Then the peak with the maximum value
within a given window is selected as the peak; here the window
size used is 400 data points. This process is repeated to find
all the peaks of the given dataset. Subsequently, the identified
number of peaks was divided by the time they occurred to
Fig. 10. Z-axis accelerometric variation of maternal respiratory motion and
fetal movement.
obtain the respiratory rate. The performance of the algorithm
on the obtained data can be observed in Figs. 14 and 15 and
their implications are discussed in Section V.A.
Furthermore, the assessment of the EE of the human body
is considered to be of importance in sports as well as in
other activities [23]. In [13], it is discussed how to estimate
human body EE by monitoring the respiratory patterns of the
subject. They have utilized an accelerometric sensor and have
conducted studies on three levels of EE: low EE, median EE,
and high EE. In this study, it was studied how this sensor
system can be utilized in identifying these three types of EE
activities.
Initially, three types of tasks were selected for the three
levels of EE. These tasks were 10-min rest, 10-min slow
walk, and 10-min run. Eight sessions were held per class, and
during each session, readings were taken for approximately
2 min. Then an algorithm was implemented to identify the
EE level of the body based on the respiratory motion data.
A three-stage algorithm was utilized, where initially the data
were prepossessed, then features were constructed, and finally,
the extracted features were utilized in the classification of
data. Initially, the tri-axial data were detrended and the mean
was removed. Then, (1) was applied to the tri-axial data to
eliminate sensor rotation interference and to combine data of
the three axes [24]
g(n) =

x(n)2
+ y(n)2
+ z(n)2
. (1)
After prepossessing the data, the time-domain signal was
segmented into epochs of 1000 samples with a 20% overlap.
Then features were calculated for each epoch. When selecting
the features to best represent the data, available time-domain
features and frequency domain features were considered. More
attention was paid to selecting a set of features that are
not redundant as well as inclusive. It can be observed in
Fig. 10 that the time-domain data are very noisy. Therefore,
it was decided to utilize frequency-domain features rather than
time-domain features.
When extracting features from the data in the frequency
domain, initially, the DWT was applied to the time-domain
data, and it was segmented into four frequency bands [25].
Individual features from each band were then computed. Then
different statistical features of these bands were considered.
Mainly two types of statistical features were considered:

Fig. 11. Algorithm implemented to classify EE using respiratory data.
measures of central tendency and measures of variability. Mea-
sures of central tendencies considered were: mean, median,
and the mode. Measures of variability considered were: stan-
dard deviation, variance, quartiles, and skewness. From these
features selected and computed were: the mean, standard devi-
ation, variance, and the skewness of each band. From this, for
a single epoch, 16 features were calculated and these features
were fed into the classifying algorithm. However, in order to
compare the performance of different types of wavelets on
accelerometric signal classification, several types of wavelets
(Daubechies wavelet db2, db4, db6, Symlet wavelet sym2, and
sym4) were used and the accuracy of each one was compared.
This comparison can be observed in Table II.
The resulting features were fed into a simple standard neural
pattern recognition algorithm. The input was set to be the
features constructed in the previous step and the output was
the three classes of activities. This network is a feed-forward
network and the training was done utilizing the scaled conju-
gate gradient back-propagation. From the dataset, 70% of the
data were used for training, 15% were used for validation, and
the remaining 15% was used for testing. The summary of the
implemented algorithm is illustrated in Fig. 11.
C. Fetal Movement Detection
The dataset obtained using the mothers was used for the
analysis involving fetal movement detection. This dataset
consists of readings from 13 pregnant mothers. Each mother
participated in a single session, and each session lasted
approximately 20–30 min. In each session, the occurrence
of fetal movements and maternal external motions such as
laugh were recorded using the two input buttons. The analysis
can be conducted in several different ways as has been stated
from rudimentary methods where the signal is observed for
peaks to the use of machine learning algorithms. This research
focuses on using machine-learning techniques since they can
be implemented in small handheld devices with accurate
results as has been demonstrated in many wearables that are
currently in wide use [26].
The use of CNNs and RNNs is considered here and the
comparison of their performance is given in Table I. As the
input for the neural network, for both RNNs and CNNs, time-
frequency parameters of the accelerometer signal were used.
This is because a time-frequency analysis is an ideal way
Fig. 12. Network architecture. (a) For CNN. (b) For GRU.
to analyze the characteristics of nonstationary signals such
as fetal movement. Hence, here, short-time Fourier transform
and wavelet transform were used to emphasize the required
characteristics of the signal.
CNNs have proved to have generally good results for
fetal movement identification as demonstrated in [17], where
spectrogram images obtained from the accelerometer readings
were used as the input to the CNN. In this article, to explore a
different perspective, wavelet transform was used to generate
the images. As stated in [27], wavelet transform suits better
for nonstationary signal analysis when compared with Fourier
analysis as subtle changes may not be represented well through
short-time Fourier transform. The network used consisted of
three convolution layers, each with 32 filters with a kernel size
of 3, each followed by a max-pooling layer and a dense layer
of 120 units, then a dropout layer, and finally a dense layer of
two units. All the dropout layers had a dropout rate of 0.4 and
all the max-pooling layers had a pool size of 2 × 2 and a
stride of 1. A visual representation of the model is given in
Fig. 12.
The dataset used for the analysis is available at [28]. For
the CNN, a set of accelerometer readings of length 3000 data
points was used to take the wavelet transform. The windowed
data were classified into two different classes 1 and 0, where
class 1 represents images that correspond to fetal movement
and class 0 to readings with no fetal movement. Through
experimentation, both visual and the performance of the net-
work, the Morlet wavelet was chosen. Then in order to remove
any influence from the trimmed edges of the signal having an
effect on the image generated, the 1000 data points at the start
and end of the transformed signal were removed. Finally, the
input was rescaled so that the input matrix varies between
1 and 0 to ensure proper training.
Another approach that is considered here is the use of
RNNs. RNNs have proved to perform best with the use of
spectrograms in [18] for fetal movement detection. Here, as an
extension of that research, the spectrograms were used as
the input to the network, but instead of using long short-
term memory (LSTM) networks, GRU networks were con-
sidered, which are almost similar in performance while being

TABLE I
COMPARISON BETWEEN THE NEURAL NETWORKS
less computationally intensive [29]. This makes this network
architecture better suited to be implemented in wearable
devices where computational power is limited. The network
implemented here consisted of initially a 1-D convolution
layer of kernel size 20 and a stride of 5 followed by batch
normalization and a dropout layer; next, a GRU layer of
80 units followed by a dropout layer and a batch normalization
layer, then finally a dense layer of 80 units followed again
by another dropout layer and a batch normalization layer.
All the dropout layers had a dropout rate of 0.2. A visual
representation is given in Fig. 12(b).
The data from the mothers were windowed to a length of
35370 data points. The length of the sample was obtained from
the longest existing fetal movement recorded from the labeled
data. The windowing was done so that the ratio of regions
of the dataset with and without fetal movement is within an
acceptable range to ensure proper training of the network. The
dataset used for training consisted of 36.51% fetal movement
information. To ensure that the trained network is generalized
well, the data was augmented such that the data is windowed
at random points relative to labels indicating fetal movement.
Next, a short-time Fourier transform of the accelerometer data
was obtained with a window size of 16 and a stride of 1 which
was used as the input to the network.
The train test split of the data was obtained for both the
CNN and the GRU from a dataset of 13 mothers. Data from
ten mothers were used for training and three mothers were
used for testing.
D. Interference Removal
The objective of this research is to develop a device to
be used at home by mothers, without medical supervision.
Hence, the posture and movements of the mother will not
be as restricted as they would be in a hospital environment.
This will require the device to have the ability to filter any
interferences.
One of the common disturbances that can occur is walking,
if the motion that the accelerometer picks up due to the phys-
ical movement is considered as noise, this can be filtered to
obtain the signal for breathing pattern analysis. Both breathing
and walking patterns are of a periodic nature and hence they
can be assumed to be stationary, making the Wiener filter a
good choice to filter and obtain the pure breathing signal.
In an instance where the required signal, in this case,
the breathing signal, is corrupted due to interferences such
as walking, the stochastic interference cancellation of the
Wiener filter can be ideally used as a different version of the
ideal signal can easily be acquired by obtaining accelerometer
Algorithm 1 Implementation of the Wiener Filter
1: m, total length of the input signals X, v
2: for n ≤ m do
3: x ← X(n − N + 1 : n), stationary reference signal of
length N
4: v(n), reference signal at nth
position
5: p ← E[xv(n)], cross correlation vector
6: R ← E[xxT
], Auto-correlation matrix
7: w0 ← R−1
p (optimal weight vector)
8: y(n) ← wT
0 x, nth
output
9: n ← n + 1
10: end for
Fig. 13. Signal estimation using Wiener filter.
Fig. 14. Peaks detected in the Z-axis accelerometric data of respiratory
motion.
readings of a person at rest. Due to the simplicity of the Wiener
filter, it is a popular choice and it was chosen for this particular
case as well. The filtering algorithm using the Wiener filter
is given in Algorithm 1. The order of the filter used was
200 (N = 200). The optimal Wiener filter implemented
filters out the walking signal by using accelerometer readings
from a stationary subject to estimate the breathing signal
superimposed with the accelerometer signals due to walking
as depicted in Fig. 13. The filter output y can be used as the
desired signal.
V. RESULTS
A. Respiratory Data Analysis
Initially, an algorithm was implemented to identify the peaks
of the respiratory motion signal. The identified peaks can be
observed in Fig. 14. The identified peaks are indicated using
red arrowheads.
However, upon further investigation, it was observed that
since the target of the algorithm is to detect peaks, it may
identify peaks at the occurrence of a fetal movement as
well. This can be observed in Fig. 15. When considering
the frequency of occurrence of fetal movement, the error
introduced to the respiratory rate due to fetal movement can
be considered to be negligible.

Fig. 15. Peaks detected when the signal contains fetal movement signal.
TABLE II
TRUE-POSITIVE RATE OBTAINED FOR DIFFERENT WAVELET TYPES
Subsequently, it was attempted to estimate the EE level
utilizing respiratory data. During this, initially preprocessing
and data segmentation were conducted. Thereafter, DWT was
implemented on individual epochs to segment the time-domain
data into four frequency bands and four features were extracted
from each band. These features were then used in the classi-
fication algorithm. When applying the DWT, different types
of Wavelet signals were used and their performance was
evaluated to identify the best wavelet to be utilized. Following
the feature extraction, classification was conducted and the
true-positive rate for each wavelet type is obtained. The
true-positive rate obtained during training, validation, and
testing is given in Table II.
B. Fetal Movement Detection
The wavelet transforms generated to be used as the inputs
to the CNN showed significant visual differences for the two
instances of the presence and absence of fetal movement
(Fig. 16), although this was not as evident in some instances.
Although the network trained showed good results for the
original dataset with higher accuracies (90%), significant per-
formance degradation can be observed since the testing of
the results was done using a different mother. Hence, it can
be inferred that the fetal movement patterns are somewhat
unique to specific cases. Therefore, for better generalization,
a significantly larger dataset must be considered or the model
should be fine-tuned so that it fits each mother individually.
The confusion matrix for the data is given in Fig. 17, where
1 and 0 represent the presence and absence of fetal movement,
respectively, for the test data.
The same observation can be seen when an RNN was used,
but with slightly better accuracies, which may be due to the
fact that it gave more emphasis on the variation of the signal
over time. A mean square error of 0.1 was observed by the
final model. Also, the performance of the model compared
to the ground-truth values which are the labels of the data is
given in Fig. 18. Fig. 18 shows how the predicted labels of
Fig. 16. Wavelet scalograms generated from the data. (a) Scalogram with
fetal movement. (b) Scalogram without fetal movement. (c) Scalogram during
mother’s laughter.
Fig. 17. Confusion matrix for the CNN.
Fig. 18. Comparison between the predicted and ground-truth values.
the network are on par with the labels of the data. The figure
indicates the presence of fetal movement as 1 and absence as
0. The ground-truth values are the labeled data that correspond
to the accelerometer readings used for the predictions and the
plot of the predictions is the prediction made by the RNN on
the accelerometer data.
The training and testing information with the number of
epochs is given in Fig. 19 for the CNN and the GRU neural
networks. From the results obtained, it is clear that while
training, the networks, both CNN and GRU, have been able
to identify features that correlate to fetal movement. But the
test set, since it has been taken from readings from different
mothers than the ones that were used for training, a significant
change in the accuracies was not observed. This indicates that

Fig. 19. Train test accuracies with the number of epochs during training.
(a) For CNN. (b) For GRU.
Fig. 20. Comparison between filtered, reference, and original signals using
the plot of AMU against time and the probability distribution of the data
points.
the fetal movement patterns are unique to different mothers
and to ensure a well-generalized model for estimation, a much
larger dataset than what was used here must be considered.
Although the test set accuracies do not seem to improve
with training, Figs. 17 and 18 show that the predictions from
the networks have acceptable accuracies even with the limited
data available for training.
C. Interference Removal
To showcase the feasibility of using the Wiener filter,
the results are presented in Fig. 20 that shows how well
the Wiener filter has extracted the desired signal using the
breathing pattern at rest from the signal corrupted by walking.
Fig. 20 compares the filtered signal with the desired reference
signal and the original signal in two different methods: a time
distribution and a probability distribution of the AMUs. It is
clear from the time distribution and the probability distribution
that the required information has been extracted from the
original signal.
This shows that the Wiener filter is suitable to be used to
extract the breathing signal from a person wearing the device
while walking by filtering out the accelerometer readings.
The algorithms, although not ideal, have been able to extract
relevant data by filtering out the noise from walking, which
has a much larger amplitude.
VI. CONCLUSION
In conclusion, the simplicity of the device makes it cheap
and easy to use and therefore the mother can have an active
role in her health monitoring and can provide the medical
experts with valuable information. This is mainly because the
device depends on different filtering and analysis techniques
rather than on complex hardware. Here, we have proposed
the device to be used to monitor the respiratory patterns of
the mother and fetal movement. From the results, it can be
noted that by using a peak detecting algorithm on the data, the
respiratory rate of the subject can be obtained. Additionally,
this device coupled with a three-step algorithm has the ability
to estimate the EE after activity with an approximate accuracy
of 98%. Furthermore, implementing a GRU-based algorithm
on the data collected resulted in identifying the occurrence of
fetal movements with a training accuracy of 90% and testing
accuracy of 75%. While this accuracy level is not sufficient for
a clinical setup, it is for day-to-day monitoring in an in-house
setting. All of these metrics can provide valuable information
when measured over a long period, which is not possible in the
existing methods and the doctors have had to solely depend
on the mother’s observations which can be inconsistent and
unreliable.
The device, along with the stated different analysis methods,
can be implemented to obtain reasonably accurate results for
pregnant mothers outside the hospital environment. It can
give a rough estimate of the mother’s and baby’s health.
Moreover, experts can observe the data from the device for
further analysis as well. This device, therefore, allows a more
personalized and long-term monitoring solution that, although
not highly accurate, can be a good addition to the existing
health monitoring systems. The most common fetal movement
analysis method that is done in in-house conditions is the
manual kick counting done by the mother. As stated earlier,
this is unreliable. Hence, the proposed device can act as an
aid to provide medical professionals with more reliable data.
In addition, the device since it captures the breathing patterns
of the mother can be used to evaluate the mother’s health as
well.
The proposed device, therefore, in conclusion, is expected
to be an aid to the existing methods rather than a substitute.
It is intended to provide basic information that is more reliable
than the current manual method of fetal movement monitoring
which can be used as an aid to the diagnosis of the patients.
REFERENCES
[1] C. Gribbin and D. James, “Assessing fetal health,” Current Obstetrics
Gynaecol., vol. 15, no. 4, pp. 221–227, Aug. 2005. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S0957584705000478
[2] P. Hamelmann, R. Vullings, M. Mischi, A. F. Kolen, L. Schmitt, and
J. W. M. Bergmans, “An extended Kalman filter for fetal heart location
estimation during Doppler-based heart rate monitoring,” IEEE Trans.
Instrum. Meas., vol. 68, no. 9, pp. 3221–3231, Sep. 2019.
[3] R. Liston et al., “Fetal health surveillance: Antepartum and intrapartum
consensus guideline,” J. Obstetrics Gynaecol. Canada, vol. 29, no. 9,
pp. S3–S4, 2007.

[4] M. J. Rooijakkers, C. Rabotti, S. G. Oei, and M. Mischi, “Low-
complexity R-peak detection for ambulatory fetal monitoring,” Physiol.
Meas., vol. 33, no. 7, p. 1135, 2012.
[5] C. Monk et al., “Effects of maternal breathing rate, psychiatric status,
and cortisol on fetal heart rate,” Develop. Psychobiol., vol. 53, no. 3,
pp. 221–233, Apr. 2011.
[6] E. Saastad, B. A. Winje, B. Stray Pedersen, and J. F. Frøen, “Fetal
movement counting improved identification of fetal growth restriction
and perinatal outcomes—A multi-centre, randomized, controlled trial,”
PLoS ONE, vol. 6, no. 12, Dec. 2011, Art. no. e28482.
[7] J. F. Pearson and J. B. Weaver, “Fetal activity and fetal wellbeing: An
evaluation,” Brit. Med. J., vol. 1, no. 6021, pp. 1305–1307, May 1976.
[8] R. Brown, J. H. B. Wijekoon, A. Fernando, E. D. Johnstone, and
A. E. P. Heazell, “Continuous objective recording of fetal heart rate
and fetal movements could reliably identify fetal compromise, which
could reduce stillbirth rates by facilitating timely management,” Med.
Hypotheses, vol. 83, no. 3, pp. 410–417, 2014.
[9] U. H. Delay et al., “Non invasive wearable device for fetal move-
ment detection,” in Proc. IEEE 15th Int. Conf. Ind. Inf. Syst. (ICIIS),
Nov. 2020, pp. 285–290.
[10] B. Benacerraf, “The use of obstetrical ultrasound in the obese gravida,”
in Seminars Perinatology, vol. 37, no. 5. Amsterdam, The Netherlands:
Elsevier, 2013, pp. 345–347.
[11] I. Guelinckx, R. Devlieger, K. Beckers, and G. Vansant, “Maternal
obesity: Pregnancy complications, gestational weight gain and nutrition,”
Obesity Rev., vol. 9, no. 2, pp. 140–150, Mar. 2008.
[12] M. B. Moyimane, S. F. Matlala, and M. P. Kekana, “Experiences of
nurses on the critical shortage of medical equipment at a rural district
hospital in South Africa: A qualitative study,” Pan Afr. Med. J., vol. 28,
no. 1, p. 157, 2017.
[13] G.-Z. Liu, Y.-W. Guo, Q.-S. Zhu, B.-Y. Huang, and L. Wang, “Estimation
of respiration rate from three-dimensional acceleration data based on
body sensor network,” Telemed. e-Health, vol. 17, no. 9, pp. 705–711,
Nov. 2011.
[14] M. A. Cretikos, R. Bellomo, K. Hillman, J. Chen, S. Finfer, and
A. Flabouris, “Respiratory rate: The neglected vital sign,” Med. J.
Austral., vol. 188, no. 11, pp. 657–659, 2008.
[15] C. Savard, A. Lebrun, S. O’Connor, B. Fontaine-Bisson, F. Haman, and
A.-S. Morisset, “Energy expenditure during pregnancy: A systematic
review,” Nutrition Rev., vol. 79, no. 4, pp. 394–409, Mar. 2021.
[16] R. Elkus and J. Popovich, “Respiratory physiology in pregnancy,”
Clinics Chest Med., vol. 13, no. 4, pp. 555–565, Dec. 1992.
[17] U. Delay et al., “Novel non-invasive in-house fabricated wearable system
with a hybrid algorithm for fetal movement recognition,” PLoS ONE,
vol. 16, no. 7, pp. 1–22, Jul. 2021, doi: 10.1371/journal.pone.0254560.
[18] E. Somathilake et al., “Fetal movement detection using long short-term
memory network,” in Proc. 10th Int. Conf. Inf. Autom. Sustainability
(ICIAfS), Aug. 2021, pp. 464–469.
[19] O. O’sullivan, G. Stephen, E. Martindale, and A. E. P. Heazell, “Pre-
dicting poor perinatal outcome in women who present with decreased
fetal movements,” J. Obstetrics Gynaecol., vol. 29, no. 8, pp. 705–710,
Jan. 2009.
[20] J. Lai et al., “Performance of a wearable acoustic system for fetal
movement discrimination,” PLoS ONE, vol. 13, no. 5, May 2018,
Art. no. e0195728.
[21] N. D. Zakaria, P. E. Numan, and M. Malarvili, “Fetal movements
recording system using accelerometer sensor,” ARPN J. Eng. Appl. Sci.,
vol. 13, pp. 1022–1032, Jan. 2018.
[22] S. Layeghy, G. Azemi, P. Colditz, and B. Boashash, “Non-
invasivemonitoring of fetal movements using time-frequency features of
accelerometry,” in Proc. IEEE Int. Conf. Acoust., Speech Signal Process.
(ICASSP), May 2014, pp. 4379–4383.
[23] A. P. Hills, N. Mokhtar, and N. M. Byrne, “Assessment of physical
activity and energy expenditure: An overview of objective measures,”
Frontiers Nutrition, vol. 1, p. 5, Jun. 2014.
[24] M. Mesbah, M. S. Khlif, C. East, J. Smeathers, P. Colditz, and
B. Boashash, “Accelerometer-based fetal movement detection,” in Proc.
Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., Aug. 2011, pp. 7877–7880.
[25] X. Zhao, X. Zeng, L. Koehl, G. Tartare, and J. D. Jonckheere, “A wear-
able system for in-home and long-term assessment of fetal movement,”
IRBM, vol. 41, no. 4, pp. 205–211, Aug. 2020.
[26] L. Meng, K. Ge, Y. Song, D. Yang, and Z. Lin, “Wearable electrocar-
diogram signal monitoring and analysis based on convolutional neural
network,” IEEE Trans. Instrum. Meas., vol. 70, pp. 1–11, 2021.
[27] M. Akin, “Comparison of wavelet transform and FFT methods in the
analysis of EEG signals,” J. Med. Syst., vol. 26, no. 3, pp. 241–247,
2002.
[28] U. Delay et al., “Fetal Movement detection dataset recorded using
MPU9250 tri-axial accelerometer,” Mendeley Data, V2, 2019, doi:
10.17632/7svcy4cscy.2.
[29] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical evaluation
of gated recurrent neural networks on sequence modeling,” 2014,
arXiv:1412.3555.
Eranda Somathilake received the B.Sc. (Eng.)
degree in mechanical engineering from the Univer-
sity of Peradeniya, Peradeniya, Sri Lanka, in 2020.
He currently works as an Instructor with the
Department of Mechanical Engineering, University
of Peradeniya. His research interests include signal
processing, machine learning, and control theory.
Upekha Hansanie Delay received the B.Sc. (Eng.)
degree (Hons.) in electrical and electronic engineer-
ing from the University of Peradeniya, Peradeniya,
Sri Lanka, in 2020.
She currently works as an Instructor with the
Department of Engineering Mathematics, University
of Peradeniya. Her primary focus is on biomed-
ical signal processing. Presently, she is involved in
research on the application of noninvasive technol-
ogy for wellbeing monitoring. She has numerous
publications in IEEE conferences along with a mul-
tidisciplinary journal publication (PLOS One). Her research interests include
image processing, signal processing, communication, machine learning, and
deep learning.
Janith Bandara Senanayaka received the B.Sc.
(Eng.) degree (Hons.) in electrical and electronic
engineering from the University of Peradeniya, Per-
adeniya, Sri Lanka, in 2020.
He then became an Instructor with the Department
of Electrical and Electronic Engineering, University
of Peradeniya. He currently works as a Consulting
Research and Development Engineer with Farbe
Technologies, Gardena, CA, USA. His research
interests include signal processing, image process-
ing, graph signal processing, computational imaging,
machine learning, and deep learning.
Samitha Lakmal Gunarathne received the B.Sc.
(Eng.) degree in electrical and electronic engineering
from the University of Peradeniya, Peradeniya, Sri
Lanka, in 2017.
Immediately after, he joined the Department of
Engineering Mathematics, University of Peradeniya,
as a Teaching Instructor. He has numerous publi-
cations in IEEE conferences. His research interests
include computer vision, image and signal process-
ing, and machine learning.

Roshan Indika Godaliyadda (Senior Member,
IEEE) received the B.Sc. (Eng.) degree in electrical
and electronic engineering from the University of
Peradeniya, Peradeniya, Sri Lanka, in 2005, and the
Ph.D. degree in electrical and computer engineering
from the National University of Singapore, Singa-
pore, in 2011.
He is currently attached to the Department of
Electrical and Electronic Engineering, Faculty of
Engineering, University of Peradeniya, as a Pro-
fessor. His current research interests include image
and signal processing, biomedical signal processing, bioimaging, bio-metrics,
computational epidemiology, pattern recognition, computer vision, smart grid,
remote-sensing applications, and algorithms.
Dr. Godaliyadda was a recipient of the Sri Lanka President’s Award for
Scientific Publications for 2018 and 2019. He was a recipient of multiple
grants through the National Science Foundation (NSF) Sri Lanka and Inter-
national Development Research Centre (IDRC) Canada for research activities.
He is also the recipient of multiple best paper awards from international
conferences for his work. His previous works have been extensively published
in numerous international journals such as the IEEE TRANSACTIONS ON
SMART GRID, the IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE
SENSING, Applied Energy, the IEEE JOURNAL OF SELECTED TOPICS IN
APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, Journal of Food
Engineering, PLOS One, and so on.
Mervyn Parakrama Ekanayake (Senior Member,
IEEE) received the B.Sc. (Eng.) degree in electrical
and electronic engineering from the University of
Peradeniya, Peradeniya, Sri Lanka, in 2006, and the
Ph.D. degree from Texas Tech University, Lubbock,
TX, USA, in 2011.
He is currently attached to the University of Per-
adeniya, as a Senior Lecturer. His current research
interests include applications of signal processing
and system modeling in remote sensing, hyperspec-
tral imaging, and smart grid.
Dr. Ekanayake was a recipient of the Sri Lanka President’s Award for
Scientific Publications in 2018 and 2019. He has obtained several grants
through the National Science Foundation (NSF) for Research Projects. His
previous works have been published in IEEE TRANSACTIONS ON GEO-
SCIENCE AND REMOTE SENSING and several other IEEE-GRSS conferences
including WHISPERS and IGARSS. He also has multiple publications in
many IEEE TRANSACTIONS,Elsevier, and IET journals and has been awarded
several best paper awards at international conferences.
Janaka Wijayakulasooriya (Member, IEEE)
received the B.Sc. degree in electrical engineering
from the University of Peradeniya, Peradeniya,
Sri Lanka, in 1994, and the Ph.D. degree in
pattern recognition from Northumbria University,
Newcastle upon Tyne, U.K., in 2000.
He is currently a Senior Lecturer with the
University of Peradeniya. His current research
interests include artificial intelligence and signal
processing.
Chathura Rathnayake received the M.B.B.S.
(Bachelor of Medicine, Bachelor of Surgery) degree
from the Faculty of Medicine, University of
Colombo, Colombo, Sri Lanka, in 1996, and the
M.S. degree in obstetrics and gynaecology from the
Postgraduate Institute of Medicine, University of
Colombo, in 2002.
He was trained in the Overseas Doctors Fellowship
in the U.K., in 2003–2005. He was a member of the
Royal College of Obstetricians and Gynecologists,
London, U.K., in 2005. He was board certified as a
Specialist with the Department of Obstetrics and Gynaecology, University
of Peradeniya, Peradeniya, Sri Lanka, in 2004. He has served with the
Ministry of Health, as a Consultant Obstetrician, and then joined the Faculty
of Medicine, University of Colombo, as a Senior Lecturer, in 2008. He is
currently working as a Professor with the Department of Obstetrics and
Gynaecology. His research interests are in high-risk pregnancy, subfertility
and reproductive toxins, biomedical instrumentation in pregnancy, and using
molecular techniques in diagnosis and treatment in obstetrics and gynecology.
Dr. Rathnayake was a recipient of the National Science Foundation (NSF)
and the National Research Council (NRC) grants for research in Sri Lanka.
He received Presidential Research Award for Research Publication of NRC
in 2017.

Assessment_of_Fetal_and_Maternal_Well-Being_During_Pregnancy_Using_Passive_Wearable_Inertial_Sensor.pdf

Recomendados

Recomendados

Mais conteúdo relacionado

Semelhante a Assessment_of_Fetal_and_Maternal_Well-Being_During_Pregnancy_Using_Passive_Wearable_Inertial_Sensor.pdf

Semelhante a Assessment_of_Fetal_and_Maternal_Well-Being_During_Pregnancy_Using_Passive_Wearable_Inertial_Sensor.pdf (20)

Mais de ssuserb4d806

Mais de ssuserb4d806 (20)

Último

Último (20)

Assessment_of_Fetal_and_Maternal_Well-Being_During_Pregnancy_Using_Passive_Wearable_Inertial_Sensor.pdf