Research on Heartbeat Detection Method of Ballistocardiogram based on
Bidirectional long short-term memory network
GENG PANG1,2, DUYAN GENG1,2
1State Key Laboratory of Reliability and Intelligence of Electrical Equipment, Hebei University of Technology,
Tianjin 300130, CHINA
2School of Electrical Engineering, Hebei University of Technology,Tianjin 300130, China
3School of Health Sciences&Biomedical Engineering, Hebei University of Technology, Tianjin 300130, CHINA
1. Introduction
ardiovascular disease is the main cause of death [1]. The
acquisition of heartbeat information through wearable or
non-contact devices is of great significance in sleep quality
assessment, cardiovascular health monitoring and mental state
recognition. In view of the inevitable contact interference and
relatively specialized operation process of electrocardiography
(ECG) acquisition, the non-contact ballistocardiography (BCG)
technology suitable for home environment has attracted much
attention. Heart rate estimation based on ballistocardiogram has
the advantages of non-invasive, simple operation and low cost,
and the resulting problem of heartbeat detection accuracy has
become a bottleneck restricting its application.
BCG signals are mechanical signals produced by heart beats,
which reflects the mechanical characteristics of the heart [2-4],
and can show changes in the body's external pressure or body
surface displacement caused by the heart's pumping activity. A
typical BCG signal is shown in the Fig. 1, which mainly contains
7 peaks such as H, I, J, K, L, M and N peaks. Usually, H, I, J, K
and L peaks are regarded as a BCG signal complex to represent
a heartbeat [5]. The J peak in BCG signal corresponds to the R
peak in electrocardiogram (ECG) [6], which represents the
maximum amplitude point of a cardiac cycle in the signal.
Therefore, accurate J peak positioning is the key to achieve
BCG signals heartbeat detection.
H
I
J
L
M
N
K
Figure 1. Standard ballistocardiograms heartbeat pattern
In recent years, the utilization of BCG signals for heart rate
statistics and prediction has been studied by researchers. These
algorithms are mainly based on threshold detection [7-8],
template matching [9-10] or machine learning theory. The
threshold method generally identifies J peaks by conventional
peak detection methods, and finding local maxima within
windows as potential heartbeat locations. This method is most
susceptible to interference, including individual differences,
acquisition mode differences, and BCG signal oscillations,
among others. The template matching method is to construct a
template database by collecting I-J-K composite wave of BCG,
achieve template matching by means of a locally moving
window function, and detect heartbeats according to correlation
coefficients. This method can obtain better results when
template matching is consistent. However, the adaptability of
the heartbeat template will be reduced due to the influence of
individual differences, or the position movement of subjects and
the change of limb state. Machine learning method [11-12] uses
unsupervised learning clustering or Gaussian mixture model
clustering algorithm to distinguish J peaks and other peaks to
C
Abstract: In order to improve the accuracy and generalization ability of extracting successive heartbeat cycle
based on ballistocardiogram (BCG), this paper proposed a general method for detecting J peak of BCG signals
by using bidirectional long short-term memory network. First, the clustering method is used to establish the
sequence feature set of BCG signals in different sleeping positions, and the data set used contains a variety of
different forms of BCG signals. Then, according to the Bidirectional LSTM (BiLSTM) many-to-many
recognition model, the number of J peaks in the output sequence is counted to achieve real-time heartbeat
detection. The results showed that the deviation rate of BCG heart rate detection was 0.27%, and there was no
significant difference between BCG and ECG in the detection of heartbeat interval. Compared with other
methods, this method has higher robustness and accuracy in detection effect, which provides a new idea for
realizing high-precision unconstrained heartbeat detection.
Keywords: Ballistocardiogram, Bidirectional LSTM, Heart rate detection, Sequence feature learning
Received: May 28, 2021. Revised: March 20, 2022. Accepted: April 23, 2022. Published: June 28, 2022.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.16
Geng Pang, Duyan Geng
E-ISSN: 2224-2902
151
Volume 19, 2022
calculate the BCG signals peak group characteristics, and then
calculate the heartbeat interval. Clustering algorithm has strong
dependencies on BCG signals with standard morphology and
single mode, and the accuracy of extracting J peaks tends to be
somewhat lower than supervised learning. However, due to less
training steps, simple process and fast output results of
clustering algorithm, clustering algorithm can efficiently label
BCG signal peak classes, thereby providing a role for
supervised learning.
In the actual acquisition process, BCG signals exhibit
different morphologies due to noise interference, individual
variations, acquisition modalities and acquisition equipment.
The morphological changes are manifested in the relative
changes of the amplitude and distance of each peak, and the
peak group model is no longer obvious. In summary, the
instability of BCG morphology needs to be taken into account
when using BCG signals to extract heart rate. Detection based
on the intrinsic temporal features of BCG complex is the key to
solving such problems. Long Short-Term Memory network
(LSTM) excels at extracting features from sequences and can
make predictions for each data in sequences. Since heart rate
information is rich in temporal characterization, LSTM network
have been widely used in heart rate information research
[13-15]. There have been studies applying LSTM on BCG
signals to calculate heart rate [16], but this method simply
calculates the approximate heart rate of a certain segment of the
signal and cannot accurately localize the heartbeat cycle.
Therefore, this paper proposes a heartbeat detection method
based on BiLSTM for J peak localization of BCG signals. By
dividing BCG sequence segments and extracting peak feature
parameters as input, the BiLSTM advanced semantic
recognition model is used to classify BCG peaks in the test set.
The heart rate is calculated according to the classification
results of the J peaks, and the accuracy is compared with the
labels in the data set to verify the effectiveness of the method in
this paper.
2. Methods
2.1 Dataset
Since there is no uniform standard for the sensors,
measurement positions, and methods used to collect BCG
signals, there is no recognized standard BCG database. The
BCG dataset used in this study need to be collected and labeled
independently. A total of 28 volunteers were recruited in the
experiment, ranging in age from 18 to 50 years old, with 16
males and 12 females. The 28 subjects were defined as P1~P28.
The BCG signals acquisition equipment uses DEEBCG
ballistocardiography (DeE Software Co., Ltd. Zhejiang, China).
As shown in Fig. 2a, after the signal is collected by the device, it
is denoised and uploaded to the cloud, and the data is obtained
at the client. The data acquisition process is shown in Figure 2b.
The subject lies flat on the bed, and the sensor is placed under
the pillow. The data is collected in the four sleeping positions of
supine, prone, left and right for 10min respectively.
a bPressure sensor
Figure 2. (a) Acquisition equipment (b) Signal acquisition process
In order to obtain the ECG signals as the reference standard
synchronously, each subject used the three-lead ECG
acquisition system designed by the research group. The R peaks
are detected using the Pan-Tumpkins algorithm [17], and the
RR interval are calculated as the reference (“ground truth”)
data.
Experimental data need to be labeled for each peak and
entered as a sequence fragment. Clustering algorithm can divide
unlabeled data into several disjoint subsets according to certain
rules [18]. The similarity within the same subset is as large as
possible, the similarity between subsets is as small as possible.
According to the characteristics of BCG complex showing
regularity and periodicity in a continuous period of time,
K-means clustering algorithm is used to distinguish various
types of BCG peaks when labeling each BCG J peaks. The
algorithm steps are as follows:
(1) Search all peaks and troughs of the signal in the interval
and the corresponding index positions,
(2) Use all the peak and trough data for parameter calculation,
as shown in Fig. 3, 4 parameters are calculated for each peak,
which are the amplitude Pa(n) of the peak P(n), the amplitude
Ta(n) of the trough T(n) adjacent to the peak P(n), the distance
Pd(n) between the peak P(n) and the trough T(n), and the
distance Td(n) between the trough T(n) and the next peak
P(n+1). Take the parameters of 5 consecutive peaks as the
eigenvector of the first peak to construct a feature set, each
eigenvector has a total of 20 parameters, the eigenvector fn is:
(1)
(3) The K-means algorithm is used to cluster the feature set.
The number of clusters is 5 and the distance is Euclidean
distance. The best arrangement is selected in the five
initializations, and the peak group index after clustering is
confirmed.
Figure 3. Characteristic parameters of peaks and troughs
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.16
Geng Pang, Duyan Geng
E-ISSN: 2224-2902
152
Volume 19, 2022
After classifying each peak, proceed to the next step of
processing and labeling, the steps are as follows:
(1) The J-wave results were manually checked against the
synchronously acquired ECG signals. The criteria are as follows:
whether the J peak is between two adjacent R peaks; whether the
number of J peaks is consistent with the number of R peaks.
After verification, unrecognized or incorrectly recognized J
peaks are corrected according to the peak index, so as to ensure
the quality of the data set.
(2) According to the J peak index, locate the H peak and L
peak index respectively. In each BCG sequence, J peak is
labeled as 1, H peak is labeled as 2, L peak is labeled as 3, and
the rest peaks are labeled as 0. An example of the labeling result
is shown in Figure 4.
(3) After labeling the peaks, each BCG signal is segmented
according to every 100 peaks as a sequence, and the remaining
fragments with less than 100 peaks are filled with 0. The dataset
finally contains 1190 heartbeat sequences, with a total of 19375
heartbeats.
Figure 4. Peaks marking results
2.2 Model Training
BiLSTM consists of forward LSTM and backward LSTM.
LSTM is a special recurrent neural network (RNN), and its cell
structure is shown in Fig. 4. Each cell unit contains three gating
units: input gate, forget gate and output gate. The output is
determined by the current state of the cell and the gate authority,
which can effectively solve the vanishing gradient problem
[19]. When each cell unit works, it first passes through the
forget gate to determine how much information to retain from
the original cell's memory. The calculation formula of forget
door is:
1,
tt
f f f
w h x b


(2)
where,
is the sigmoid activation function,
f
w
is the forget
gate weight matrix,
1t
h
is the hidden layer state at time (t-1),
t
x
is the input at time t, and
f
b
is the forget gate intercept.
Next, the proportion of new information stored in the new cell
state is determined through the input gate. The input gate
expression and cell state correction value expression are:
1,
tt
i i i
w h x b


(3)
~1
tanh ,
t t t
cc
C w h x b



(4)
where,
i
w
and
i
b
are the weight and intercept of input gate,
c
w
and
c
b
are the weight and intercept of cell state correction value.
Finally, the output is determined according to the current
state information, the new cell state is obtained, and the hidden
layer is updated. The expressions of the output gate, the current
cell state and the hidden layer at the current moment are:
1,
tt
o o o
w h x b


(5)
~1t t t
if
C C C
(6)
tanh
tt
o
hC
(7)
where,
o
w
and
o
b
are the weight and intercept of the output gate,
tanh is the hyperbolic tangent function.
As shown in Fig. 5, the BiLSTM model adopts a
many-to-many output mode, each sequence contains 100 steps,
and the output produces 100 classification results.
LSTM LSTM LSTM LSTM
LSTM LSTM LSTM LSTM
y1 y2 y3 y4
x1 x2 x3 x4
t
x
1t
h
tanh
tanh
f
i
o
t
h
t
C
t
y
1t
C
LSTM-Cell
Figure 5. Structure of BiLSTM
The overall framework of the network is shown in Fig. 6,
which consists of an input layer, two BiLSTM layers, two fully
connected layers and an output layer. The input layer is input
from the labeled feature set, and each feature sequence is a
100
3 data matrix, where 100 is the step size, corresponding to
the index information of 100 peaks in the BCG sequence; 3 is
the dimension, corresponding to the characteristic parameters of
each peak, which are the peak amplitude and the distance
between the peak and the adjacent two peaks. In order to play
the role of expanding features, the labels are processed by
one-hot encoding. The hidden layer has two BiLSTM layers and
two fully connected layers. The first BiLSTM layer is set with
128 hidden layer units, and the second BiLSTM layer is set with
64 hidden layer units. The two fully connected layers are set
with 32 and 16 hidden layer units respectively, and the
activation function is set as a Rectified Linear Activation
Function. The output layer uses the softmax activation function
for logistic regression to classify the four output results.
Adaptive moment estimation and cross-entropy loss function
are used to optimize the network [18]. The network weights and
biases are optimized after each iteration according to the value
of the loss function.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.16
Geng Pang, Duyan Geng
E-ISSN: 2224-2902
153
Volume 19, 2022
Input layer
Characteristic scale of peak
sequence
100 3
BiLSTM
LSTM LSTM LSTM LSTM
LSTM LSTM LSTM LSTM
BiLSTM
LSTM LSTM LSTM LSTM
LSTM LSTM LSTM LSTM
Fully connected layer
12345 32
Fully connected layer
123 16
Output layer Peak category
Figure 6. Heartbeat detection model framework based on BiLSTM
The data set contains 1190 feature sequences, which are
divided in a ratio of 3:1. Data from 21 subjects was used as the
training set and data from 7 subjects as the test set. The number
of samples is 884 and 306 respectively, and the data is
normalized before training. The main training parameters of the
model are set as follows: the number of training epoch is 100,
the sample batch size is 50, and the learning rate is 0.01. 20% of
the training set data are set as the cross-validation set, and the
two layers of BiLSTM are set with a dropout operation of 0.5
and 0.3 respectively to reduce overfitting.
3. Results
The evaluation method is as follows:
(1) Model evaluation. The classification performance of
BiLSTM network was evaluated using accuracy (Acc),
sensitivity (Se), specificity (Sp), precision (P), F1 score (F1)
and confusion matrix.
(2) Heartbeat detection verification. The purpose of this
paper is to locate the J peak, and the accuracy of J peak
recognition is mainly verified by the number of heartbeats. The
heartbeat detection results were evaluated using deviation rate
(E), false positive rate (FPR), false negative rate (FNR),
precision rate and recall rate.
(3) Statistical analysis. Based on the results of heartbeat
detection, the statistical differences between BCG detection
results and synchronized ECG were analyzed by paired sample t
test and Pearson correlation coefficient.
3.1 Model Evaluation
When building the dataset, the setting of labels is essentially
binary: J peak and non-J peak. Therefore, in the process of
model evaluation, we only evaluate the recognition results of J
wave. The test set data is predicted on the classification model,
and the evaluation calculation results are:
Acc
=99.67%,
Se
=99.14%,
Sp
=99.78%,
P
=98.89%,
1F
=99.01%. The
confusion matrix of test set classification results is shown in
Table I.
Table I. J peak classification confusion matrix of test set
Reference
Prediction
J peak
Non-J peak
J peak
5086
44
Non-J peak
57
25413
According to Table I and the evaluation calculation results,
the model shows high accuracy in the classification of J peak
and non-J peak. In order to verify the advantages of this model,
we compared it with other popular sequence learning models.
We established RNN, LSTM, and Gate Recurrent Unit (GRU)
models for comparative experiments. The structure and
parameter settings of the three models are consistent with the
aforementioned BiLSTM model. Three experiments were
performed for each model, and the results are shown in Table 2.
It can be seen that the BiLSTM model has the highest average
accuracy, which is 7.12%, 6.12%, and 3.67% higher than RNN,
LSTM, and GRU.
Table II. Comparison of accuracy of different methods
Learning
models
Accuracy
First
time
Second
time
Third
time
Average
RNN
92.52%
91.33%
92.79%
92.21%
LSTM
93.97%
92.33%
93.32%
93.21%
GRU
95.99%
94.98%
96.02%
95.66%
BiLSTM
98.99%
99.67%
99.32%
99.33%
3.2 Heartbeat Detection Results
The J peak test results from one subject are selected for prior
validation. The BCG signal is a mechanical signal of the heart
that itself has non-strict periodicity, so the heartbeat location
detected by the BCG signal lags behind the heartbeat location of
the ECG, and the JJ interval varies slightly from the RR interval.
Figure 7a shows the comparison between the BCG heartbeat
detection results of the subject and the ECG, the square marks in
the upper figure indicates the R peak location of the ECG, which
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.16
Geng Pang, Duyan Geng
E-ISSN: 2224-2902
154
Volume 19, 2022
serves as the ground truth for this experiment. The square marks
in the lower figure indicates the J peak labels of the BCG, and
the vertical marks are the recognition results. It can be seen that
the detection results correspond one-to-one with the ground
truth. The results of JJ interval versus RR interval are shown in
Figure 7b, which shows that the JJ interval and RR interval
substantially coincide. Figure 7c is a Bland-Altman diagram of
JJ interval versus RR interval, and it can be observed that
detection results are basically in the 95% confidence zone,
illustrating good agreement between JJ interval and RR interval
detected herein within a certain margin of error.
a
cb
Amplitude/V
Amplitude/V
0 200 400 600 800
0.4
0.6
0.8
1.0
1.2
Heartbeat(s)
Heartbeat interval/ s
JJ interval
RR interval
0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00
-0.20
-0.15
-0.10
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
JJ interval/s
Difference between JJ interval and RR interval/s
+1.96(0.03751)
-1.96(-0.03750)
Figure 7. Comparison of calculation results between BCG and ECG: (a) Heartbeat detection, (b) Heartbeat interval, (c) Bland-Altman plot
Table III. Heartbeat detection results based on BiLSTM
Subject
Heartbeat(s)
TP
E%
FNR%
FPR%
Se%
P%
P22
674
672
0.30%
3.45%
0.82%
96.55%
96.28%
P23
680
680
0
0.73%
0.15%
99.27%
99.27%
P24
819
812
0.86%
1.69%
0.52%
98.31%
97.48%
P25
702
698
0.57%
0
0.08%
100%
99.57%
P26
714
714
0
0
0
100%
100%
P27
946
945
0.11%
0
0.02%
100%
99.89%
P28
609
609
0
0.16%
0.02%
99.84%
99.84%
Total
5144
5130
0.27%
0.85%
0.22%
99.15%
98.90%
Table III shows the detection results for subjects numbered
P22 to P28 in the test set. The number of detected JJ interval is
taken as the heartbeat detection result, TP is the number of heart
rate labels in the test set, E is the deviation rate of the detection
result, FNR is the false negative rate, FPR is the false positive
rate, Se is the recall rate, and P is the precision rate. As can be
seen from Table III, the average deviation rate of the subjects is
only 0.27%, the accuracy of heartbeat detection is as high as
99.73%, and the average false negative rate does not exceed 1%.
Only individual subjects had more false negative samples and
the average false positive rate did not exceed 0.3%. In terms of
recall rate and precision rate, the identification method herein
reached 99.15% and 98.90%. In addition, we also use the same
dataset to evaluate the performance of traditional threshold
detection, template matching, and hierarchical clustering. Table
IV shows the detection results of each method, the results show
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.16
Geng Pang, Duyan Geng
E-ISSN: 2224-2902
155
Volume 19, 2022
that our method has a higher overall accuracy rate in the same
field of study, can achieve high-precision heart rate monitoring requirements, and has a stronger expansion value.
Table IV. Comparison with other algorithms
Algorithms
E%
FNR%
FPR%
Se%
P%
Threshold
detection
2.59%
6.04%
4.12%
94.15%
93.83%
Template
matching
0.92%
2.68%
1.95%
97.41%
97.15%
Hierarchical
clustering
1.86%
6.69%
0.72%
94.31%
94.08%
Proposed
0.27%
0.85%
0.22%
99.15%
98.90%
Table V. Comparison of statistical parameters of BCG and ECG signal characteristic parameters
Subject
HRECG/min-1
HRBCG/min-1
P
R
RR interval /s
JJ interval /s
P
R
P22
71.11±2.75
71.33±2.73
0.894
0.945
0.841±0.031
0.840±0.032
0.889
0.923
P23
69.67±1.58
69.88±1.69
0.773
0.976
0.862±0.015
0.861±0.018
0.794
0.937
P24
80.63±3.14
81.18±3.12
0.738
0.916
0.722±0.043
0.718±0.052
0.721
0.835
P25
64.30±1.72
64.72±1.68
0.806
0.941
0.922±0.082
0.917±0.094
0.716
0.859
P26
67.18±0.83
67.22±0.86
0.908
0.985
0.884±0.068
0.885±0.068
0.954
0.966
P27
85.11±2.11
85.20±2.16
0.723
0.938
0.706±0.028
0.710±0.045
0.663
0.902
P28
68.38±1.88
68.81±1.35
0.827
0.948
0.880±0.036
0.877±0.051
0.814
0.821
3.3 Statistical Analysis
Based on the results of J peak detection, paired samples t test
and Pearson correlation coefficient statistics were performed on
the average heart rate and the successive heartbeat interval of
the heartbeat detection results. The statistical results are shown
in Table V. It can be seen from Table V that the heart rate and
heartbeat interval calculated from BCG signals and ECG signals
respectively have no significant difference (P>0.05). The
Pearson correlation coefficients were all between 0.8 and 1.0,
indicating that there was a strong correlation between the
calculated results of the BCG signal and the ECG signal. The
above statistical data show that within a certain error range, the
heart rate calculated by the BCG signals can replace the
reference heart rate of the ECG signals, which further illustrates
the feasibility and accuracy of the heartbeat detection method
herein.
4. Conclusion
This paper proposes a J peak detection method based on
BCG peak sequence through the study of BCG peak sequence.
This method does not rely on local time domain signals. By
calculating the characteristic parameters of each wave peak in
the BCG time series segment, and using the BiLSTM
many-to-many recognition model, the method can directly
classify the characteristic differences of BCG peak groups and
efficiently detect the J peak in BCG. The data samples include
BCG waveforms of different subjects in different sleeping
positions, which are not limited to a single acquisition. In future
work, based on the advantages of deep learning algorithms, the
performance of this method can be further verified and
improved under the condition of expanding the training set. In
addition, in order to achieve heart rate monitoring in groups
with heart-related diseases, we need to continue to study the
feature extraction methods.
References
[1] E. J. Benjamin et al. “Heart disease and stroke
statistics-2019 update a report from the american heart
association,” Circulation, vol. 139, no. 10, pp. e56–e528,
Mar. 2019.
[2] He S, Dajani H R, Meade R D, et al. “Continuous Tracking
of Changes in Systolic Blood Pressure using BCG and
ECG,” Annu Int Conf IEEE Eng Med Biol Soc, vol. 2019,
pp. 6826-6829 Jul. 2019.
[3] González-Landaeta R, Casas O, Pallàs-Areny R. “Heart
rate detection from an electronic weighing scale,”
Physiological measurement, vol. 29, no.8, pp. 979-988, Jul.
2008.
[4] Da H D, Winokur E S, Sodini C G. “A continuous,
wearable, and wireless heart monitor using head
ballistocardiogram (BCG) and head electrocardiogram
(ECG),” Annu Int Conf IEEE Eng Med Biol Soc, vol.
2011, pp. 4729-4732, 2011.
[5] Inan O T, Migeotte P F, Park K S, et al.
“Ballistocardiography and seismocardiography: A review
of recent advances,” IEEE Journal of Biomedical and
Health Informatics, vol.19, no. 4, pp. 1414-1427, Jul. 2015.
[6] Shin J H, Hwang S H, Chang M H, Park K S. “Heart rate
variability analysis using a ballistocardiogram during
Valsalva manoeuvre and post exercise.” Physiological
Measurement, vol.32, no. 8, pp. 1239-1264, Jul. 2011.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.16
Geng Pang, Duyan Geng
E-ISSN: 2224-2902
156
Volume 19, 2022
[7] Choi B H, Chung G S, Lee J S, Jeong D U, Park K S.
“Slow-wave sleep estimation on a load-cell-installed bed: a
non-constrained method,” Physiol Meas, vol. 30, no. 11,
pp. 1163-1170, Oct. 2009.
[8] D. Heise and M. Skubic. “Monitoring pulse and respiration
with a non-invasive hydraulic bed sensor,” Annu Int Conf
IEEE Eng Med Biol Soc, vol. 2010, pp. 2119–2123, 2010.
[9] Shin J H, Choi B H, Lim Y G, Jeong D U, Park K S.
“Automatic ballistocardiogram (BCG) beat detection using
a template matching approach,” Annu Int Conf IEEE Eng
Med Biol Soc, vol. 2008, pp. 1144-1146, 2008.
[10] Q Xie et al. “A personalized beat-to-beat heart rate
detection system from ballistocardiogram for smart home
applications,” IEEE Trans Biomed Circuits Syst, vol. 13,
no. 6, pp. 1593–1602, Dec. 2019.
[11] Brüser C, Stadlthanner K, Brauers A, Leonhardt S.
“Applying machine learning to detect individual heart beats
in ballistocardiograms,” Annu Int Conf IEEE Eng Med
Biol Soc, vol. 2010, pp. 1926-1929, 2010.
[12] Shen G, Yang M Q, Zhang B Y.
“Ballistocardiogram-Based Heart Rate Variation
Monitoring Using Unsupervised Learning,” Advances in
Transdisciplinary Engineering, 2018, 7.
[13] Murat F, Yildirim O, Talo M, Baloglu U B, Demir Y,
Acharya U R. “Application of deep learning techniques for
heartbeats detection using ECG signals-analysis and
review,” Computers in Biology and Medicine, vol. 120, pp.
103726, Mar. 2020.
[14] Faust O, Shenfield A, Kareem M, San T R, Fujita H,
Acharya U R. “Automated detection of atrial fibrillation
using long short-term memory network with RR interval
signals,” Computers in Biology and Medicine, vol. 102, no.
2018, pp. 327-335, Nov. 2018.
[15] Yildirim Ö. “A novel wavelet sequence based on deep
bidirectional LSTM network model for ECG signal
classification,” Computers in Biology and Medicine, vol.
96, no. 2018, pp. 189-202, May. 2018.
[16] Jiao C, Chen C, Gou S, et al.“Non-Invasive Heart Rate
Estimation From Ballistocardiograms Using Bidirectional
LSTM Regression,” IEEE J Biomed Health Inform, vol.
25, no. 9, pp. 3396-3407, Sep.2021.
[17] Malešević N, Petrović V, Belić M,
Antfolk C, Mihajlović V, Janković M. “Contactless
Real-Time Heartbeat Detection via 24 GHz
Continuous-Wave Doppler Radar Using Artificial Neural
Networks,” Sensors (Basel), Vol. 20, no. 8, pp. 2351, Apr.
2020.
[18] Wael Ahmad A1Zoubi. “A Survey of Clustering
Algorithms in Association Rules Mining,” International
Journal of Computer Science and Information Technology,
vol.11, no. 2, 2019.
[19] Wu S, et al. “Deep learning in clinical
natural language processing: a methodical review,”
Journal of the American Medical Informatics Association,
vol. 27, no. 3, pp. 457-470, Mar. 2020.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the Creative
Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en_US
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.16
Geng Pang, Duyan Geng
E-ISSN: 2224-2902
157
Volume 19, 2022