ScienceDirect
Available online at www.sciencedirect.com
Procedia Computer Science 151 (2019) 551–558
1877-0509 © 2019 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the Conference Program Chairs.
10.1016/j.procs.2019.04.074
10.1016/j.procs.2019.04.074 1877-0509
Available online at www.sciencedirect.com
Procedia Computer Science 00 (2018) 000–000
www.elsevier.com/locate/procedia
The 10th International Conference on Ambient Systems, Networks and Technologies (ANT)
April 29 - May 2, 2019, Leuven, Belgium
A Real-time PPG Quality Assessment Approach for
Healthcare Internet-of-Things
Emad Kasaeyan Naeinia,∗, Iman Azimib, Amir M. Rahmania,c, Pasi Liljebergb, Nikil Dutta
aUniversity of California, Irvine, Department of Computer Science, Irvine, California, USA
bUniversity of Turku, Department of Future Technologies, Turku, Finland
cTU Wien, Institute of Computer Technology, Vienna, Austria
Abstract
Photoplethysmography (PPG) as a non-invasive and low-cost technique plays a significant role in wearable Internet-of-Things
based health monitoring systems, enabling continuous health and well-being data collection. As PPG monitoring is relatively
simple, non-invasive, and convenient, it is widely used in a variety of wearable devices (e.g., smart bands, smart rings, smartphones)
to acquire different vital signs such as heart rate and pulse rate variability. However, the accuracy of such vital signs highly depends
on the quality of the signal and the presence of artifacts generated by other resources such as motion. This unreliable performance
is unacceptable in health monitoring systems. To tackle this issue, different studies have proposed motion artifacts reduction and
signal quality assessment methods. However, they merely focus on improvements in the results and signal quality. Therefore, they
are unable to alleviate erroneous decision making due to invalid vital signs extracted from the unreliable PPG signals. In this paper,
we propose a novel PPG quality assessment approach for IoT-based health monitoring systems, by which the reliability of the vital
signs extracted from PPG quality is determined. Therefore, unreliable data can be discarded to prevent inaccurate decision making
and false alarms. Exploiting a Convolutional Neural Networks (CNN) approach, a hypothesis function is created by comparing
heart rate in the PPG with corresponding heart rate values extracted from ECG signal. We implement a proof-of-concept IoT-based
system to evaluate the accuracy of the proposed approach.
c© 2018 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the Conference Program Chairs.
Keywords: Health Monitoring; Internet of Things; Photoplethysmogram; Wearable Electronics; Deep Learning; Convolutional Neural Networks
1. Introduction
Ubiquitous health monitoring systems are increasingly in demand to provide reactive and proactive healthcare
solutions for at-risk patients as well as mitigating medical costs, errors, and workloads [5, 30]. Fortunately, techno-
∗ Corresponding author. Tel.: +1-949-610-6661 .
E-mail address: ekasaeya@uci.edu
1877-0509 c© 2018 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the Conference Program Chairs.
Available online at www.sciencedirect.com
Procedia Computer Science 00 (2018) 000–000
www.elsevier.com/locate/procedia
The 10th International Conference on Ambient Systems, Networks and Technologies (ANT)
April 29 - May 2, 2019, Leuven, Belgium
A Real-time PPG Quality Assessment Approach for
Healthcare Internet-of-Things
Emad Kasaeyan Naeinia,∗, Iman Azimib, Amir M. Rahmania,c, Pasi Liljebergb, Nikil Dutta
aUniversity of California, Irvine, Department of Computer Science, Irvine, California, USA
bUniversity of Turku, Department of Future Technologies, Turku, Finland
cTU Wien, Institute of Computer Technology, Vienna, Austria
Abstract
Photoplethysmography (PPG) as a non-invasive and low-cost technique plays a significant role in wearable Internet-of-Things
based health monitoring systems, enabling continuous health and well-being data collection. As PPG monitoring is relatively
simple, non-invasive, and convenient, it is widely used in a variety of wearable devices (e.g., smart bands, smart rings, smartphones)
to acquire different vital signs such as heart rate and pulse rate variability. However, the accuracy of such vital signs highly depends
on the quality of the signal and the presence of artifacts generated by other resources such as motion. This unreliable performance
is unacceptable in health monitoring systems. To tackle this issue, different studies have proposed motion artifacts reduction and
signal quality assessment methods. However, they merely focus on improvements in the results and signal quality. Therefore, they
are unable to alleviate erroneous decision making due to invalid vital signs extracted from the unreliable PPG signals. In this paper,
we propose a novel PPG quality assessment approach for IoT-based health monitoring systems, by which the reliability of the vital
signs extracted from PPG quality is determined. Therefore, unreliable data can be discarded to prevent inaccurate decision making
and false alarms. Exploiting a Convolutional Neural Networks (CNN) approach, a hypothesis function is created by comparing
heart rate in the PPG with corresponding heart rate values extracted from ECG signal. We implement a proof-of-concept IoT-based
system to evaluate the accuracy of the proposed approach.
c© 2018 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the Conference Program Chairs.
Keywords: Health Monitoring; Internet of Things; Photoplethysmogram; Wearable Electronics; Deep Learning; Convolutional Neural Networks
1. Introduction
Ubiquitous health monitoring systems are increasingly in demand to provide reactive and proactive healthcare
solutions for at-risk patients as well as mitigating medical costs, errors, and workloads [5, 30]. Fortunately, techno-
∗ Corresponding author. Tel.: +1-949-610-6661 .
E-mail address: ekasaeya@uci.edu
1877-0509 c© 2018 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the Conference Program Chairs.
© 2019 The Authors. Published by Elsevier B.V.
This is an open access article under the CC B - C-  license (http://creativeco ons.org/licenses/by-nc-nd/4.0/)
eer-re ie  er res sibility of the Conference Program Chairs.
552 Emad Kasaeyan Naeini  et al. / Procedia Computer Science 151 (2019) 551–558
2 Kasaeyan Naeini, E. et al. / Procedia Computer Science 00 (2018) 000–000
logical advancements in Internet-of-Things (IoT) systems, wearable electronics, and data analytics create significant
opportunities for the development of such monitoring systems [18, 15]. These IoT-based healthcare systems can tailor
various types of wearable sensors, transmission units, and remote computing resources to deliver end-to-end health
solutions to users [32, 8]. In this regard, Photoplethysmography (PPG) as a non-invasive, low-cost, and convenient
technique can play a significant role in IoT-based health monitoring systems to continuously provide various health
parameters such as heart rate, respiration rate, and oxygen saturation.
PPG is an optical technique that indicates synchronous blood volumetric changes in the microvascular bed of tis-
sues. Such variations in the PPG signal are associated with the oscillation of heartbeat and respiration [2]. The signal
is collected by exposing light (e.g., infrared, red, and green) to the skin surface and acquiring the light reflection via
photodetectors [40]. Due to simplicity and feasibility of the measurement, this technique is widely used these days in
different wearable and mobile applications [29, 33, 3], and in several clinical and commercial medical devices such as
pulse oximeter, smart rings, and smartwatches [23, 6, 12]. However, the PPG signal quality is highly susceptible to ar-
tifacts generated by the user’s motions or surrounding resources. Unfortunately, such background noises are inevitable
in IoT-based health monitoring systems, as users engage in various physical activities in different environmental con-
ditions. For example, the signal is corrupted by hand motions in wrist-based PPG sensors although the distortion
depends on the signal power and the wavelength [26]. The poor signal (i.e., low signal-to-noise ratio) leads to inac-
curate vital signs extraction and erroneous decision making. Since inaccurate or incorrect assessment based on these
signals may risk life-threatening consequences, such unreliable performance is unacceptable for healthcare applica-
tions. There are numerous studies in the literature to reduce the PPG signals artifacts. Several techniques have been
proposed employing different signal processing, filtering, and noise cancellation methods including adaptive filters,
spectral subtraction method, and Fourier series analysis [35, 43, 16, 41, 20]. Moreover, other efforts have tackled the
motion artifact reduction, integrating other resources such as acceleration data to their techniques [34, 19, 42, 7, 11].
These methods could improve the signal quality and heart rate estimates, although they were unable to guarantee
highly accurate results all the time, particularly when the Signal-to-Noise Ratio was too low. Consequently, they
might still lead to incorrect results and false alarms.
Other studies have tackled the quality assessment of bio-signals [17]. In the context of PPG studies, Elgendi [13]
introduced a method to use signal quality indices for the quality assessment. The PPG signals were manually annotated
in this study. A method was proposed by Sun et al. [39] to use morphological characteristics and variability of the
signal to assess the signal quality. Another study [27] extracted four waveform features and determined the signal
quality using a decision tree. Similarly, Li and Clifford [28] proposed to feed different features extracted from the
signal to a neural network algorithm to determine the quality. These studies mostly classify the signal quality as
“good” or “bad” according to different features extracted from the signal. However, this definition is inaccurate in the
health monitoring systems, considering the reliability of the signal’s outcome (i.e., health parameters) and decision
makings. Let’s take two different examples in this regard. A signal could be tagged as “bad” when the temporal features
are distorted by motion artifact. However, the heart rate values could be detected flawlessly leveraging a robust artifact
reduction algorithm; and subsequently the decision making is accurate. On the other hand, a signal could be tagged as
“good” because of the superb features although pulse rate variability is erroneous because the sampling frequency is
too small [10]. Therefore, the medical decisions can be invalid.
Within the IoT-based health monitoring systems, we believe a quality assessment approach is required to determine
the quality of the PPG signal according to the validity and reliability of the health parameters and subsequently medical
decisions. Consequently, poor data can be discarded and decision making with invalid data and misinterpretation can
be prevented. In other words, missing data can be less hazardous than unnoticed incorrect data as there are many
data imputation techniques to fill the gap [4] (e.g., repeating the imputed heart rate value vs. showing a significantly
incorrect heart rate value).
In this paper, we propose a novel automatic PPG quality assessment approach for IoT-based health monitoring,
focusing on the accuracy of health parameters extracted from PPG. We employ the Convolutional Neural Network
(CNN) deep learning algorithm to learn a customized hypothesis function from PPG signals, labeling the signal as
“reliable” or “unreliable”. In the learning process, the PPG quality is specified by a set of Electrocardiogram (ECG)
signals as the baseline, comparing heart rate in the PPG with heart rate in the ECG signal. Moreover, we evaluate our
proposed approach on a proof-of-concept IoT-based system.
 Emad Kasaeyan Naeini  et al. / Procedia Computer Science 151 (2019) 551–558 553
Kasaeyan Naeini, E. et al. / Procedia Computer Science 00 (2018) 000–000 3
Fig. 1: A window of a filtered PPG waveform
The rest of the paper is organized as follows. Section 2 briefly outlines the background of this work. In Section 3, we
present the proposed PPG quality assessment approach and our setup. The system evaluation is presented in Section 4.
Finally, Section 5 concludes the paper.
2. Background
The two most widely used methods to measure the cardiac cycle and monitor heart rate are Electrocardiography
(ECG) and Photoplethysmography (PPG). The ECG measures the electrical activation that leads to the contraction
of the heart muscle, using electrodes attached to the body, usually at the chest. The PPG uses a small optical sensor
in conjunction with a light source to measure the discoloration of the skin as blood perfuses through it after each
heartbeat. Since our work focuses on PPG, in this section we summarize the background on Photoplethysmography
(PPG) and Convolutional Neural Networks (CNNs).
2.1. Photoplethysmogram (PPG)
The PPG method senses the rate of blood flow by using a pulse-oximeter and measuring light absorption. Con-
traction of the heart causes an increase in blood pressure, so the Oxygen-rich blood is pumped through the body.
Therefore, the cardiac cycle can be measured by detecting the peaks in this signal. Such a procedure can give a real-
time measurement of the heart rate. A typical PPG signal is shown in Figure 1, in which the heart rate value can be
calculated by the amount of time between two consecutive peaks. The signal contains two major components: the
alternating current (AC) and direct current (DC) [2]. The AC component is attributed to synchronous blood variation
oscillating with the heartbeats. The DC component corresponds to the various low-frequency components of blood
flow such as respiration rate.
The PPG is easy to set up, convenient, simple, and economically efficient [9]. A PPG sensor includes a light source
(e.g., green LED) and a light sensor placed on the skin to expose and collect the reflected light, respectively. The
red and infrared lights are conventionally leveraged in finger pulse oximeter devices to obtain heart rate and blood
Oxygen saturation. Moreover, the green light is mostly utilized for wearable devices such as smartwatches for heart
rate detection.
2.2. Convolutional Neural Networks (CNNs)
Similar to how a human learns to recognize objects, a machine learning algorithm needs to be presented with an
object several times before it can classify the inputs and make predictions for the incoming data. Convolutional Neu-
ral Networks, ConvNets or CNN [25] represent a class of deep learning algorithms that are organized as hierarchical
feed-forward neural networks. CNNs differ from traditional neural networks through their hierarchical architecture
and connectivity organization. Each layer of a CNN can be defined via three dimensions: height, width, and depth.
Furthermore, neurons in each convolutional layer are not fully connected to the neurons of the next layer [24]. Clas-
sification is achieved through successive extraction of features at each layer; these features are generated by sliding
filters along the input data at that layer. Finally, the ultimate output is reduced to a single vector of probability scores.
CNNs enable efficient and automatic extraction of features, overcoming the tedious and possibly error-prone task of
manual feature extraction, and also removing the need for an expert human-in-the-loop.
554 Emad Kasaeyan Naeini  et al. / Procedia Computer Science 151 (2019) 551–558
4 Kasaeyan Naeini, E. et al. / Procedia Computer Science 00 (2018) 000–000
Fig. 2: 10-second windows of Empatica E4 and PulseOn PPG signals along with the ECG peak locations showing by the red dashed lines
3. PPG Quality Assessment Approach
In this section, we present how our proposed approach automatically performs the quality assessment of the PPG
signal. We first outline the setup and then describe the detection approach.
3.1. Study Design and Setup
The system is designed to continuously monitor heart rate values in everyday settings. It consists of two wearable
devices for PPG and one for ECG data collection. The devices are PPG-based Empatica E4 (E4) [14], PPG-based
PulseOn(PO) [31], and Shimmer ECG Dev Kit (SHM) [37]. The E4 and PO are wearable wristbands enabled by a
green light source. The output signal of these wrist-worn devices is a fluctuation in voltage, representing the variation
in blood volume. Two views of the PPG signal collected via both devices are shown in Figure 2. In this experiment,
the E4 and PO wristbands were worn tightly enough on the right and left hands, respectively. We select the Shimmer
device to measure ECG signal as a gold standard to calculate heart rate values. ECG monitors the electrical activity
of heart and is not prone to motion artifacts (i.e., hand movements). To prevent motion artifacts, the Shimmer device
was placed on the chest along with two skin electrodes.
For this study, we carried out a continuous ECG and PPG data collection for 5 days. The sampling rates, reported
by device manufacturers, are 201.3 Hz for the Shimmer device, 64 Hz for the Empatica E4, and 25 Hz for the PulseOn.
Considering heart rates between 30 to 180 beats per minute, a Butterworth filter [36] was set to only pass heartbeat
signals (i.e., 0.5-3 Hz). Then, a filter-based technique was utilized to extract heart rate values from the PPG signal [3].
In this technique, first, a filter was designed using the cutoff frequencies that were dynamically selected according
to the peak values in the power spectral density (PSD) of the incoming signal. Second, a peak detection method was
utilized to obtain the heartbeat cycles and subsequently calculate the heart rate values. The processing methods were
performed using the Scipy library [21] in Python.
3.2. Methodology
We propose a machine learning-based method to assess the PPG signal quality. We train a classifier in this regard
to distinguish between reliable and unreliable signals. For the training dataset, an automatic method is utilized to label
the collected PPG signals as reliable or unreliable. The automatic method is performed in the system as follows. First,
the PPG signal is synchronized in time with our gold standard ECG reference. Second, we employ a sliding window
strategy to segment the signal into windows of 60 seconds that have a 50% overlap with each other. Third, the signal’s
amplitude in each window is normalized to enhance the peak detection (i.e., heart rate extraction). Then, the windows
(i.e., PPG signal) are labeled by comparing the average heart rate value of each window with the corresponding heart
rate value from the ECG reference. The PPG signals along with the labels are fed to the proposed classification
method.
The PPG Quality Assessment has been proposed in different studies using traditional machine learning algo-
rithms [28, 38]. However, such algorithms need proper features to perform the classification. In these works, feature
 Emad Kasaeyan Naeini  et al. / Procedia Computer Science 151 (2019) 551–558 555
Kasaeyan Naeini, E. et al. / Procedia Computer Science 00 (2018) 000–000 5
Fig. 3: CNN architecture used for the PPG Quality Assessment
values are extracted from the signal at first. Then, feature selection and dimensionality reduction are applied. Even
though, such features might indicate the signal quality according to the variations in temporal and frequency domains,
they cannot reflect the accuracy of signal’s outcome. Therefore, a reliable and automatic method is needed to per-
form the feature extraction process. We believe deep networks are effective in this regard. One can feed the network
with the data directly and discard the challenging feature extraction process, as extracting feature can be automati-
cally done in the architecture itself. One of the most successful deep architectures is Convolutional Neural Networks
(CNN) [25, 24]. CNN can extract features from the 1D signals (PPG signals in our case) and handle the entire feature
engineering part. CNN consists of hierarchical layers for feature extraction. First layers extract basic features from the
signal although higher layers obtain more complex features.
We here design a CNN method to perform PPG quality assessment. Figure 3 shows the overall architecture of our
CNN computational pipeline. The inputs are 60-second windows of PPG signals. The processing begins with 196
convolutional filters with the size of 1 × 16. Each convolution layer takes inputs from the previous layers to convolve
with several kernels and pass the result to the next layers. We leverage a non-linear activation function in this method.
Among all the conventional activation functions, we tend to choose the rectified linear unit (ReLU), as its gradients
are non-saturation and accelerates the convergence of the stochastic gradient descent in comparison with sigmoidal
and hyperbolic tangent [24]. ReLU units are applied to the 196 feature maps. A pooling layer is then used to reduce
the dimensionality, increasing the invariance of the features. Taking the maximum value and average value of the data
are two conventional choices to pool feature maps from the previous layers. We select max-pooling of size 1 × 4 due
to its sensitivity to existence of some patterns in pooled regions.
The pooling layer is subsequently passed to a flattering operation (see Figure 3) which unifies all the outputs and
stacks them together. It makes a feature vector which is passed to the network classifier. The flattened feature vector is
then followed by a fully-connected layer with 1024 neurons. The fully-connected layer performs the classification part.
To minimize overfitting, a dropout technique with the rate of 0.05 is used. Finally, a softmax layer predicts output class
probability distribution over two reliable and non-reliable classes, using the inputs from the fully connected layer. In
our setup, Tensorflow computation library [1] is used for training the model and evaluation. The network’s parameters
are optimized with Adam [22] optimization method. Then, binary cross-entropy loss as a loss function on the softmax
output is used. With l2-norm regularization of CNN weights, the goal of training our model which is minimization of
cross-entropy loss function is satisfied. In each iteration from a given training set of PPG, we randomly pick up the
segments that injects enough noise to each gradient update, while achieving a relative speedy convergence.
4. Experimental Results
In this section, we evaluate the proposed approach in terms of accuracy of the signal quality assessments. The PPG
signals in this setup are measured using two different devices, i.e., Empatica E4 and PulseOn. Heart rate values are
extracted from the signals, and labeling of the windows (i.e., reliable or unreliable) is carried out by comparing the
obtained heart rates with the ECG heart rates. For the collected 5-day data, the Mean Absolute Error (MAE) for the
Empatica E4 and PulseOn are 8.21 and 12.41, respectively, measuring the distance between the PPG heart rate values
and the ECG ones. The binary labels are specified using a threshold for the distance between both heart rate values.
We select 10 beats per minute distance as the threshold in this evaluation. Therefore, the window is considered as
reliable when the distance between its heart rate and ECG heart rate is less than 10 beats per minutes.
556 Emad Kasaeyan Naeini  et al. / Procedia Computer Science 151 (2019) 551–558
6 Kasaeyan Naeini, E. et al. / Procedia Computer Science 00 (2018) 000–000
(a) (b)
(c)
Fig. 4: Average ROC of the sensors (a) Empatica E4, (b) PulseOn, and (c) both Empatica E4 and PulseOn
To evaluate the proposed method, three different datasets as ”Empatica E4 data”, ”PulseOn data” and ”Empatica
E4 and Pulseon data” are selected for the training process. In this regard, three models are trained to indicate the
performance of the proposed approach in case of different training datasets. Then, we evaluate the models on 1440
60-second windows of PPG signals as the test data. The performance of the proposed method are shown in Figures 4
and 5, and Table 1.
The Receiver Operating Characteristic (ROC) curves of the three models are indicated in Figure 4. Moreover, Area
Under Curve (AUC) per each class and contiguous flattened array (micro-average) of both reliable and unreliable
classes are shown for the three classifiers. AUC values are respectively 0.77, 0.82, and 0.88 for the classifiers trained
with ”Empatica E4 data” (Figure 4a), ”PulseOn data” (Figure 4b), and ”Empatica E4 and PulseOn data” (Figure 4c).
The AUC of the third model is higher that the other two classifiers. Recall that the test data is selected from the both
Empatica E4 and Pulseon datasets. Consequently, the classification accuracy is increased if the model is trained using
both PPG sources.
Moreover, the CNN architecture is evaluated using different values for the dropout rate. This evaluation is per-
formed to mitigate overfitting in the model. The average ROC curve of the third model (green line) along with the
standard deviation (gray area) is shown in Figure 5a. The average AUC is 0.91 and remains almost the same for all
the dropout rates, indicating accuracy changes are insignificant with different dropout values.
Figure 5b shows the normalized confusion matrix of the third experiment, applying the third model on the test
dataset. As indicated, the proposed method could correctly estimate the “Reliable” and “Unreliable” classes %85 and
%83 of the times, respectively. Table 1 presents precision, recall, F1-score and support values of the two classes.
Micro-average and macro average of the model are also indicated in the table. Our results show an acceptable per-
formance for the proposed CNN-based approach, through which a binary decision is delivered to indicate PPG signal
quality in a real-time manner.
 Emad Kasaeyan Naeini  et al. / Procedia Computer Science 151 (2019) 551–558 557
Kasaeyan Naeini, E. et al. / Procedia Computer Science 00 (2018) 000–000 7
(a) Average ROC for validation (b) Normalized confusion matrix
Fig. 5: Average ROC validation and normalized confusion matrix for the third model on the test set.
Table 1: The precision, recall, F1-score, and support results of the proposed CNN method of both classes alongside with micro avg, macro avg
precision recall F1-score support
Unreliable 0.7674 0.8354 0.8000 79
Reliable 0.8850 0.8333 0.8584 120
micro avg 0.8342 0.8342 0.8342 199
macro avg 0.8262 0.8344 0.8922 199
5. Conclusion
PPG is a simple optical technique utilized in various wearable devices to continuously acquire health parameters
such as heart rate. Such a technique is highly susceptible to environmental noises, particularly motion artifacts. In this
work, we presented an automated PPG quality assessment approach for IoT-based health monitoring systems, focus-
ing on the accuracy of the health parameters extracted from the PPG. Using our approach, unreliable signals could be
removed to prevent false decision making. We leveraged a CNN approach to determine the PPG signal quality accord-
ing to the extracted heart rate values. In the training process, the PPG signal was automatically labeled as reliable and
unreliable by comparing the PPG signal against an ECG signal, as a gold standard for heart rate measurements. The
hypothesis function was created by feeding the PPG signals to the model. We tested our algorithm by implementing
a proof-of-concept IoT-based system. Our results indicated an acceptable performance for the proposed approach to
distinguish between reliable and unreliable PPG signals. The precision and recall of the approach were 0.89 and 0.83
in our experiment, respectively.
Acknowledgements
This work was partially supported by the US National Science Foundation (NSF) grants WiFiUS CNS-1702950
and S&CC CNS-1831918, and Academy of Finland grants 311764, 311765, 313448, and 313449.
References
[1] M. Abadi, , et al. Tensorflow: A system for large-scale machine learning. In OSDI’16, pages 265–283, 2016.
[2] J. Allen. Photoplethysmography and its application in clinical physiological measurement. Physiol Meas., 28(3):1–39, 2007.
[3] D. Amiri et al. Edge-Assisted Sensor Control in Healthcare IoT. In IEEE GLOBECOM SAC EH, 2018.
[4] I. Azimi et al. Missing data resilient decision-making for healthcare iot through personalization: A case study on maternal health. Future
Generation Computer Systems, 96:297–308, 2019.
[5] M.M. Baig and H. Gholamhosseini. Smart health monitoring systems: An overview of design and modeling. Journal of Medical Systems,
37(2):9898, 2013.
558 Emad Kasaeyan Naeini  et al. / Procedia Computer Science 151 (2019) 551–558
8 Kasaeyan Naeini, E. et al. / Procedia Computer Science 00 (2018) 000–000
[6] S. Benedetto et al. Assessment of the fitbit charge 2 for monitoring heart rate. PLoS ONE, 13(2), 2018.
[7] M. Boloursaz-Mashhadi et al. Heart rate tracking using wrist-type photoplethysmographic (ppg) signals during physical exercise with simul-
taneous accelerometry. IEEE Signal Processing Letters, 23(2):227–31, 2016.
[8] L. Catarinucci et al. An iot-aware architecture for smart healthcare systems. IEEE Internet of Things Journal, 2(6):515–26, 2015.
[9] P. Cheang and P. Smith. An overview of non-contact photoplethysmography. Electronic Systems and Control Division Research, pages 57–9,
2003.
[10] A. Choi and H. Shin. Photoplethysmography sampling frequency: pilot assessment of how low can we go to analyze pulse rate variability with
reliability? Physiological Measurement, 38(3):586, 2017.
[11] S.S. Chowdhury et al. Real-time robust heart rate estimation from wrist-type ppg signals using multiple reference adaptive noise cancellation.
IEEE Journal of Biomedical and Health Informatics, 22(2):450–9, 2018.
[12] N. Constant et al. Pulse-glasses: An unobtrusive, wearable hr monitor with internet-of-things functionality. In IEEE 12th International
Conference on Wearable and Implantable Body Sensor Networks, 2015.
[13] M. Elgendi. Optimal signal quality index for photoplethysmogram signals. Bioengineering, 3(4):21, 2016.
[14] Empatica. Real-time physiological signals/ e4 eda/gsr sensor, Retrieved on January 2019. https://www.empatica.com/research/e4/.
[15] F. Firouzi et al. Internet-of-things and big data for smarter healthcare: From device to architecture, applications and analytics. Future Genera-
tion Computer Systems, 78:583–6, 2018.
[16] A. Galli et al. Measuring heart rate during physical exercise by subspace decomposition and kalman smoothing. IEEE Transactions on
Instrumentation and Measurement, 67(5):1102–10, 2018.
[17] N. Gambarotta et al. A review of methods for the signal quality assessment to improve reliability of heart rate and blood pressures derived
parameters. Med Biol Eng Comput, 54:1025, 2016.
[18] J. Gubbi et al. Internet of things (iot): A vision, architectural elements, and future directions. Future Gener Comput Syst, 29(7):1645–60, 2013.
[19] H. Han et al. Development of real-time motion artifact reduction algorithm for a wearable photoplethysmography. In Conf Proc IEEE Eng
Med Biol Soc., 2007.
[20] S. Islam et al. Preheat: Precision heart rate monitoring from intense motion artifact corrupted ppg signals using constrained rls and wavelets.
Biomedical Signal Processing and Control, 38:212–23, 2017.
[21] E. Jones et al. SciPy: Open source scientific tools for Python, 2001, Retrieved on February 2019. https://www.scipy.org/.
[22] D.P. Kingma and J. Ba. Adam: A Method for Stochastic Optimization. CoRR, abs/1412.6980, 2014.
[23] H. Koskima¨ki et al. How do we sleep: A case study of sleep duration and quality using data from oura ring. In UbiComp ’18, 2018.
[24] A. Krizhevsky et al. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems,
pages 1097–1105, 2012.
[25] Y. LeCun and Y. Bengio. The Handbook of Brain Theory and Neural Networks, chapter Convolutional Networks for Images, Speech, and Time
Series, pages 255–258. MIT Press, 1998.
[26] J. Lee et al. Comparison between red, green and blue light reflection photoplethysmography for heart rate monitoring during motion. In Conf
Proc IEEE Eng Med Biol Soc., pages 1724–7, 2013.
[27] K. Li et al. Onboard tagging for real-time quality assessment of photoplethysmograms acquired by a wireless reflectance pulse oximeter. IEEE
Transactions on Biomedical Circuits and Systems, 6(1):54–63, 2012.
[28] Q. Li and G. D. Clifford. Dynamic time warping and machine learning forsignal quality assessment of pulsatile signals. Physiol. Meas.,
33:1491–1501, 2012.
[29] S. Majumder et al. Wearable sensors for remote health monitoring. Sensors, 17(1):130, 2017.
[30] R. Mieronkoski et al. The internet of things for basic nursing carea scoping review. International Journal of Nursing Studies, 69:78–90, 2017.
[31] Pulse ON. OHR Tracker PulseOn, Retrieved on January 2019. http://pulseon.com/ohr/ohr-tracker.
[32] A. Pantelopoulos and N. G. Bourbakis. A survey on wearable sensor-based systems for health monitoring and prognosis. Trans. Sys. Man
Cyber Part C, 40(1):1–12, 2010.
[33] S. Patel et al. A review of wearable sensors and systems with application in rehabilitation. J Neuroeng Rehabil., 9:21, 2012.
[34] M.R. Ram et al. A novel approach for motion artifact reduction in ppg signals based on as-lms adaptive filter. IEEE Transactions on
Instrumentation and Measurement, 61(5):1445–57, 2012.
[35] K. A. Reddy et al. Use of fourier series analysis for motion artifact reduction and data compression of photoplethysmographic signals. IEEE
Transactions on Instrumentation and Measurement, 58(5):1706–11, 2009.
[36] I.W. Selesnick and C.S. Burrus. Generalized digital butterworth filter design. IEEE Transactions on signal processing, 46(6):1688–1694, 1998.
[37] Shimmer. Ecg sensor development kit, wearable ecg sensor, wireless ecg, Retrieved on January 2019. https://www.shimmersensing.
com/products/ecg-development-kit.
[38] J.A. Sukor et al. Signal quality measures for pulse oximetry through waveform morphology analysis. Physiol Meas., 32(3):369–84, 2011.
[39] X. Sun et al. Assessment of photoplethysmogram signal quality using morphology integrated with temporal information approach. In Conf. of
the IEEE EMBS, 2012.
[40] T. Tamura et al. Wearable photoplethysmographic sensorspast and present. Electronics, 3(2):282–302, 2014.
[41] Q. Wang et al. Artifact reduction based on empirical mode decomposition (emd) in photoplethysmography for pulse rate detection. In Conf
Proc IEEE Eng Med Biol Soc., 2010.
[42] R. Wijshoff et al. Reduction of periodic motion artifacts in photoplethysmography. IEEE Trans Biomed Eng, 64(1):196–207, 2017.
[43] C. Zong and R. Jafari. Robust heart rate estimation using wrist-based ppg signals in the presence of intense physical activities. In IEEE EMBC,
2015.