A Deep Learning–based PPG Quality Assessment Approach for Heart

Rate and Heart Rate Variability

EMAD KASAEYAN NAEINI, University of California Irvine, USA

FATEMEH SARHADDI, IMAN AZIMI, and PASI LILJEBERG, University of Turku, Finland
NIKIL DUTT and AMIR M. RAHMANI, University of California Irvine, USA

Photoplethysmography (PPG) is a non-invasive optical method to acquire various vital signs, including heart rate (HR) and
heart rate variability (HRV). The PPG method is highly susceptible to motion artifacts and environmental noise. Unfortu-
nately, such artifacts are inevitable in ubiquitous health monitoring, as the users are involved in various activities in their
daily routines. Such low-quality PPG signals negatively impact the accuracy of the extracted health parameters, leading to
inaccurate decision-making. PPG-based health monitoring necessitates a quality assessment approach to determine the sig-
nal quality according to the accuracy of the health parameters. Different studies have thus far introduced PPG signal quality
assessment methods, exploiting various indicators and machine learning algorithms. These methods differentiate reliable
and unreliable signals, considering morphological features of the PPG signal and focusing on the cardiac cycles. Therefore,
they can be utilized for HR detection applications. However, they do not apply to HRV, as only having an acceptable shape
is insufficient, and other signal factors may also affect the accuracy. In this article, we propose a deep learning–based PPG
quality assessment method for HR and various HRV parameters. We employ one customized one-dimensional (1D) and three
2D Convolutional Neural Networks (CNN) to train models for each parameter. Reliability of each of these parameters will be
evaluated against the corresponding electrocardiogram signal, using 210 hours of data collected from a home-based health
monitoring application. Our results show that the proposed 1D CNN method outperforms the other 2D CNN approaches. 24
Our 1D CNN model obtains the accuracy of 95.63%, 96.71%, 91.42%, 94.01%, and 94.81% for the HR, average of normal to
normal interbeat (NN) intervals, root mean square of successive NN interval differences, standard deviation of NN intervals,
and ratio of absolute power in low frequency to absolute power in high frequency ratios, respectively. Moreover, we compare
the performance of our proposed method with state-of-the-art algorithms. We compare our best models for HR-HRV health
parameters with six different state-of-the-art PPG signal quality assessment methods. Our results indicate that the proposed
method performs better than the other methods. We also provide the open source model implemented in Python for the
community to be integrated into their solutions.
CCS Concepts: • Computer systems organization → Embedded systems; Redundancy; Robotics; • Networks → Network
reliability;
Additional Key Words and Phrases: Deep learning, convolutional neural networks, signal quality assessment, health moni-
toring, Internet of Things, photoplethysmogram, heart rate variability, wearable electronics

This work was partially supported by the US National Science Foundation (NSF) grant S&CC CNS-1831918, and Academy of Finland through
the SLIM Project under Grant 316810 and Grant 316811.
Authors’ addresses: E. Kasaeyan Naeini, University of California Irvine, California, USA; e-mail: ekasaeya@[Link]; F. Sarhaddi, I. Azimi, and
P. Liljeberg, University of Turku, Finland; e-mails: {[Link], [Link], pakrli}@[Link]; N. Dutt and A. M. Rahmani, University of
California Irvine, USA; e-mails: {dutt, [Link]}@[Link].

This work is licensed under a Creative Commons Attribution International 4.0 License.
© 2023 Copyright held by the owner/author(s).
2637-8051/2023/11-ART24
[Link]

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
24:2 • E. Kasaeyan Naeini et al.

ACM Reference format:

Emad Kasaeyan Naeini, Fatemeh Sarhaddi, Iman Azimi, Pasi Liljeberg, Nikil Dutt, and Amir M. Rahmani. 2023. A Deep
Learning–based PPG Quality Assessment Approach for Heart Rate and Heart Rate Variability. ACM Trans. Comput. Healthcare
4, 4, Article 24 (November 2023), 22 pages.
[Link]

1 INTRODUCTION
Internet of Things technology is fundamentally changing the delivery of healthcare, enabling health monitor-
ing applications anywhere and anytime [1–5]. Wearables, as intelligent electronic devices, play a key role in
such applications; health data are collected,analyzed, and shared over a network. Such devices are becoming
widely used around the globe, as they are even more miniaturized, smarter, and easier to use. Wearables can
include a variety of sensing resources to continuously monitor body functions. Photoplethysmography (PPG)
is an inexpensive and convenient method that is being employed in a variety of wearables as well as smart-
phones to collect various vital signs such as heart rate (HR), heart rate variability (HRV), SpO2, and respiration
rate [6].
PPG is a simple optical method to measure plethysmogram, showing the variations of blood volume in body
organs. The obtained signal can be tailored to track cardiorespiratory parameters. The PPG method mainly
consists of two components placed on the skin. First, a light source is utilized to reflect light to the skin surface.
The red, infrared, or green light can be selected according to the application. Second, a photodetector collects
the light reflection [7]. The collected signal includes a pulsatile (AC) and a slowly fluctuating (DC) component,
allowing the monitoring of cardiorespiratory parameters non-invasively and continuously. The PPG method is
tailored in many clinically approved devices and commercial wearables (e.g., smartwatches and rings), as it is
easy-to-implement and energy efficient [8–13].
However, input PPG signals might be distorted due to noises caused by motion artifacts and other environmen-
tal sources [14], which are ubiquitous and unavoidable in everyday life settings. For instance, the light sensors
might be exposed by environment light sources, by which the collected PPG signal is distorted and the infor-
mation is concealed within the signal. Specifically, we observe noise affects the collected signal differently as
the users participate in various physical activities while using the PPG-based wearables. Such movements could
negatively impact the signal quality [15, 16]. For example, if the engaged subject is running, then the noise power
is much higher compared to that affecting the same signal acquired when the participant is sleeping [12]. Low-
quality PPG signal (i.e., low signal-to-noise ratio) affects the reliability of the health parameters extracted, e.g.,
HR and HRV parameters. Such unreliable measurements can lead to false alarms or life-threatening decisions in
healthcare applications [17].
In the literature, the PPG signal quality was investigated by proposing assessment methods to discriminate
reliable and unreliable parts of the signal. Various studies introduced signal quality indicators [18–20] or uti-
lized template matching approaches to distinguish the reliable segments of the PPG signal [21–23]. Moreover,
traditional machine learning methods such as support vector machine (SVM), decision tree, and K-nearest neigh-
bors were presented to carry out PPG quality assessment using features extracted from the shape of the signal
[24–30]. Recently, deep learning methods also exploited PPG quality assessment, enabling automatic PPG feature
extraction [31–35].
These studies have mostly assessed the shape of the signal, focusing on the HR. In other words, the signal
is classified as reliable if the cardiac cycle can be detected. We believe that such a quality assessment method
is insufficient for PPG signals, from which many other parameters can also be extracted. PPG signals can be
leveraged to remotely collect HRV parameters whose accuracy might be diminished due to different factors
in the signal. A PPG signal might be reliable for HR detection but, for example, unreliable for standard devi-
ation of NN intervals (SDNN). Therefore, a PPG quality assessment method is essential in health monitoring

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
 A Deep Learning–based PPG Quality Assessment Approach • 24:3

applications, determining the reliability of the PPG signal according to the health parameters. Such a method
needs to provide a confidence value for each health parameter. Consequently, unreliable parameters can be re-
moved, preventing incorrect health decision-making.
Moreover, there are available public datasets, such as WESAD, including physiological and motion data
recorded using wrist- and chest-worn devices in lab settings [36]. PPG signals are often distorted due to motion
artifacts and environmental noise. Therefore, data collected in lab settings are insufficient for PPG signal quality
assessment studies. The signal quality assessment methods should be evaluated using the PPG data collected in
free-living conditions, where the users engage in their daily routines.
In this article, we propose a quality assessment method to distinguish reliable PPG signals according to the
HR and HRV values extracted from the signal. Convolutional Neural Network (CNN) methods are tailored in this
regard to train models for each health parameter. We design (i) a customized 1D CNN method and (ii) three
2D CNN methods enabled by three deep neural networks. The proposed methods are investigated, and the best
architecture is selected. Then, the performance of the selected method is evaluated, in comparison to existing
rule-based, machine learning, and deep learning PPG quality assessment methods. To this end, we perform a
home-based electrocardiogram (ECG) and PPG collection, in which the signals are acquired simultaneously and
remotely for 24 hours. The evaluation includes more than 210 hours of data. For each health parameter, the PPG
quality is defined in comparison to the ECG using an automatic annotation process. The main contributions of
this article are as follows:
• Proposing a CNN-based PPG quality assessment method based on HR and HRV
• Customizing the method with different CNN architectures to investigate the performance
• Presenting an automatic annotation method, where the PPG signals are labeled as “reliable” or “unreliable”
against the ECG as a baseline
• Evaluating the proposed methods in comparison to state-of-the-art PPG quality assessment methods via a
home-based monitoring where more than 210 hours of PPG and ECG signals are collected
• Providing a portable and open source model implemented in Python1 for the community to be integrated
in their solutions
The rest of the article is organized as follows. Section 2 outlines the motivation behind this research and the
related works. We describe the background of this research in Section 3. In Section 4, we present the case study,
including recruitment, data collection, and data annotation. Section 5 introduces the PPG quality assessment
approach. In Section 6, the results are presented. Finally, Section 7 concludes the article.

2 RELATED WORK AND MOTIVATION

Several signal quality assessment techniques have been introduced in the literature to determine whether col-
lected PPG signals are reliable. To this end, the unreliable part of the signal is detected and removed, preventing
misinterpretation of data and invalid decision-making. Such techniques are mostly focused on the morphology
of the bio-signals. Different signal quality indicators—such as skewness, kurtosis, and baseline wandering of
the signal—were exploited to estimate the PPG signal quality [18–20]. Rule-based methods have also been in-
troduced to distinguish low-quality bio-signals, leveraging decision rules, e.g., signal saturation detection and
beat-to-beat-interval evaluation [37–39]. Moreover, template matching techniques have been proposed to dis-
tinguish reliable and unreliable signals. For example, Sun et al. [21] proposed a template matching method
based on dynamic time warping for signal quality assessment. In other similar studies, the quality assessment
was performed by investigating the morphological similarity between the original signal and the template sig-
nals, which were the expected waveform or surrounding pulses [22, 23]. The PPG morphological waveform can
be affected by motion artifacts, environmental noise, and cardiovascular issues. Therefore, these methods are

1 [Link]

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
24:4 • E. Kasaeyan Naeini et al.

inaccurate due to variations in the PPG morphological features. Moreover, these methods usually need prede-
fined thresholds that should set manually based on the data.
A number of machine learning methods have been developed for PPG signal quality assessment. Various stud-
ies in the literature proposed to train machine learning methods, utilizing morphological and time/frequency
domain features of PPG signals. These methods utilized supervised and lightweight unsupervised approaches.
The supervised methods include hierarchical decision rule [40], decision tree [41], random forest [24], SVM
[25, 26], and neural network [29]. With this intention, Sabeti et al. [27] proposed a threshold optimization learn-
ing method and compared the method with a SVM and a decision tree. Furthermore, in Reference [28], the
authors investigated SVM, K-nearest neighbors, and decision tree for the signal quality assessment in the case of
atrial fibrillation. In addition, Mahmoudzadeh et al. [30] proposed a lightweight unsupervised method providing
low-computational real-time PPG quality assessment. The authors of Reference [42] also proposed unsupervised
PPG SQA methods based on the self-organizing map. Machine learning–based methods outperform rule-based
methods in terms of accuracy. However, similarly to the rule-based method, results are affected by morpholog-
ical variation of PPG signals. In addition, machine learning–based methods utilized manual feature extraction.
Therefore, the accuracy and generalization of these methods are restricted due to a limited set of selected features.
Recently, deep learning methods were also introduced for the PPG quality assessment. Kasaeyan Naeini et al.
[33] introduced a real-time PPG assessment approach to classify the signals according to the HR values. Perieira
et al. [31] used 1-dimensional (1D) and 2D deep neural networks for signal quality assessment in case of atrial
fibrillation. They considered the PPG signal as time series (1D) and also as images (2D). Authors in Reference [34]
also proposed a 1D CNN model for determining reliable segment of PPG signals. In another work, Soto et al. [32]
proposed a multi-task CNN called DeepBeat for signal quality assessment and atrial fibrillation detection. They
improved their results by pre-training the model using convolutional denoising auto-encoders (CDAE). In addition,
Roh et al. [35] utilized a 2D CNN for classifying each segment of beat waveform. These methods benefited from
automatic feature extraction. However, they need manual/expert labeling based on HR values or morphological
waveform of PPG signals. They are inappropriate for HRV parameters analysis. A comparison of the PPG quality
assessment methods is indicated in Table 1.
The PPG quality assessment methods, presented in the literature, were mostly designed to analyze the shape
of a signal and assessing the reliability of the PPG signal itself focusing solely on HR. These methods tailor
rules, templates, and hypothesis functions to classify a PPG as reliable if the desired signal (i.e., cardiac cycle) is
appropriately retrieved. Such methods can be utilized for HR detection applications where the pulse is visible
in the reliable signals. However, they are not applicable to HRV analysis, as only having an acceptable shape
is insufficient, and other factors in the signal may affect the accuracy, too. Moreover, such factors could impact
differently on the accuracy of the HRV parameters. For instance, a 5-minute high-quality PPG signal with two
or three noisy peaks could result in an accurate/acceptable HR and SDNN but can lead to an invalid Root Mean
Square of Successive Differences (RMSSD) value.
To emphasize the importance of developing quality assessment models for each HRV parameter, and for clar-
ification, let us show five real motivational examples of PPG signals with different artifacts in Figures 1 and 2.
These figures show that a noisy PPG signal, which results in some unreliable HRV parameters, can still be used
to extract other HRV parameters accurately. Therefore, the quality of PPG signals should be evaluated accord-
ing to the desired HRV parameters. In these examples, we extract the HR, average of normal to normal interbeat
(NN) intervals (AVNN), RMSSD, SDNN, and low frequency/high frequency (LF/HF) ratio of the PPG signals. Then,
we specify their errors by calculating the distance between these parameters and the corresponding parameters
obtained from the baseline ECG signals. In other words, the error is ECG value − PPG value.
(a) Figure 1(a) illustrates an example, in which less than 5 seconds of the PPG signal (highlighted in red) is
distorted due to hand movements. As indicated, a few peaks are detected incorrectly in the noisy part. The
PPG signal provides HR, AVNN, and SDNN values with low errors compared with the ECG. However, the

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
 A Deep Learning–based PPG Quality Assessment Approach • 24:5

Table 1. Comparing Current PPG Quality Assessment Methods with Proposed Method for Their Features, Algorithms,
Annotation Procedures, and Source-code Availability
Automatic
Automatic Source code
Reference Features Features Method Annotation
Labeling Availability
extraction
Proposed Based on HR and
 — 1D CNN/ 2D CNN  
method HRV parameters
Extracted HR
Based on
Orphanidou et al. [37] × and morphological Threshold based rules × ×
the signal shape
Rule based features
Methods
Hierarchical
Reddy et al. [38] × Predictor Coefficient × " ×
decision rules

Statistical parameters Predefined

Tyapochkin et al. [39] × × " ×
of IBIs rules

Amplitude and Hierarchical

Vadrevu et al. [40] × × " ×
time-series features decision rules
Morphological Based on the
Alam et al. [24] × features and Random forest × signal shape ×
Supervised HR value and HR values
methods
Frequency domain
Based on the
Zhang et al. [25] × and time series SVM × ×
signal shape
characteristics

Frequency-domain,
Preira et al. [26] × time-domain, and SVM × " ×
non-linear features

Frequency-domain,
Preira et al. [28] × and time-domain SVM × " ×
features
Unsupervised Mahmoudzadeh et al. [30] × Statistical features Elliptical envelope × " 
methods
Entropy and
Roy et al. [42] × signal complexity Self-organizing map × " ×
features
Preira et al. [31]  — ResNet18 × " ×
Deep learning
Multi-task CNN
methods
Soto et al. [32]  — (pre-training with × " ×
CDAE)

Goh et al. [34]  — 1D CNN × " ×

Roh et al. [35]  — 2D CNN × " ×

Naeini et al. [43]  — 1D CNN  Based on HR ×

error for the RMSSD is high. The reason is that RMSSD is correlated to the short-term variation in the PPG
signal, and a small corrupted part in the signal could affect its accuracy.
(b) Figure 1(b) shows a PPG signal with no distorted peaks. However, the variation of the intervals is not
similar to the ECG. This signal provides reliable HR, AVNN, and RMSSD but unreliable SDNN. The reason
is that SDNN is associated with long-term variation in normal-to-normal interbeat intervals (NNIs), and the
noises that affect the variation of intervals will impact SDNN accuracy.
(c) Figure 1(c) shows a 1-minute window where the PPG signal is partially corrupted (i.e., 20%). This ratio of
noise in the signals affects both short-term and long-term variations in the interval distribution. Therefore,
both RMSSD and SDNN are unreliable in this signal, while the extracted HR and AVNN are acceptable.

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
24:6 • E. Kasaeyan Naeini et al.

Fig. 1. One-minute windows of filtered PPG signals, from which HRV parameters with different accuracy are extracted.

(d) Figure 1(d) depicts a noisy PPG signal. The noise corrupted a considerable part of the signal and affected
detected peaks and long-term and short-term variation of NNIs. Therefore, extracted HR, RMSSD, and
SDNN are inaccurate. However, the error of the AVNN is relatively low as AVNN represents the mean of
the intervals and is more resistant to outliers and noises. This example shows that we can extract some
HRV parameters with acceptable accuracy despite the high ratio of noise.
(e) Figure 2(a) and (b) show two 5-minute samples of PPG signal with accurate HR, AVNN, SDNN, and RMSSD.
However, the signal indicated in Figure 2(a) is reliable for extracting LF/HF, while the signal in Figure 2(b)
is unreliable for the LF/HF. The reason is that noises with same frequency of low- and high-frequency
bands of NNIs affect the accuracy of LF/HF. Figure 2(c) and (d) show power spectral density (PSD) of NNIs
for these two samples in Figure 2(a) and (b), respectively. As indicated, although the two PPG signals (in the
time domain) are similar, the power of the NNIs signals in the low-frequency and high-frequency bands are
different. As indicated in these figures, noise affects the frequency bands of the signals and, consequently,
the accuracy of LF/HF.
These examples show that HRV parameters are of essence for PPG quality assessment and those PPG qual-
ity assessment methods that are solely based on HR or the morphology of the PPG itself are not efficient
and can result in many adverse consequences, some of which may be life-threatening due to not fully correct

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
 A Deep Learning–based PPG Quality Assessment Approach • 24:7

Fig. 2. Two five-minute samples of filtered PPG signals with the corresponding PSD of NNIs.

decisions made by doctors. Moreover, state-of-the-art PPG quality assessment methods were mostly evaluated
using the simulated data or data collected in controlled lab settings with limited motion artifacts. We believe
that the confidence models need to be trained using the data collected in everyday settings where the subjects
engage in several physical activities in various environments. This way, the model can learn about the validity
of the signal in different conditions with different artifacts.

3 BACKGROUND
Recent advances in information and communication technology provide an opportunity to enable remote health
monitoring using wearable electronics. Such wearables can measure biomedical signals, allowing continuous
monitoring of the individual’s health condition. ECG is a non-invasive method that can be used to remotely track
cardiorespiratory parameters using portable monitors [8, 44]. The method includes limb and chest electrodes,
which collect electrical signals generated from the action potentials of the heart cells. The ECG is the gold
standard in HR detection and diagnosis of cardiovascular diseases. However, it cannot be performed for long-term
monitoring due to its complicated data collection. Alternatively, PPG is a more convenient method to monitor
cardiorespiratory variables. The PPG acquires the rate of blood flow in the tissue (e.g., wrist) as controlled by the

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
24:8 • E. Kasaeyan Naeini et al.

Fig. 3. A window of a filtered PPG waveform.

heart’s pumping action. The method leverages an optical sensor in conjunction with a light source to collect the
signals. In the following, we outline background on the PPG, HRV, and CNNs as a method we used for the PPG
analysis.

3.1 Photoplethysmography
PPG is an optical measurement method that records the variation of blood flow by emitting a light onto the
surface of the skin and measuring the light absorption. The PPG signal consists of a pulsatile (AC) component
and a non pulsatile (DC) component [6]. The AC component reflects the pulsations in the interrogated blood
volume with each heartbeat, whereas the DC component contains the low frequency fluctuations, including
absorption from the tissue and bones as well as static blood absorption [45]. The AC component—oscillated with
the contraction and relaxation of the heart—enables the measurements of cardiac cycles by detecting the peaks
(i.e., maximum values) in this signal. The PPG signal calculated with this procedure can provide the real-time
measurements of HR. Moreover, variation in time intervals between the successive pulse peaks, called HRV, allow
us to obtain more information about the Autonomic Nervous System (ANS). Figure 3 shows an individual PPG
signal.
The PPG method is convenient, economic, and easy to set up [46] using an optical sensor in conjunction with
a light source such as a green LED. The method is already used in various commercial and clinical devices. The
PPG with green light is utilized in the optical sensors of most consumer wearable devices such as smartwatches
for HR calculation. In addition, the PPG with red and infrared LEDs are employed in pulse oximeters to monitor
peripheral capillary oxygen saturation (SpO2 ).

3.2 HRV Analysis

HRV consists of the fluctuations in the time periods between successive heartbeats [47]. In the literature, HRV
analysis has been introduced to examine the ANS correlated with pain intensity, stress level, sleep quality, to
name but a few [43, 48–50]. Conventionally, HRV values are calculated from the ECG signal by extracting the
cardiac cycles, i.e., the RR-intervals in the signal. Alternatively, the HRV can be also obtained using the PPG
signals, where the peak-to-peak intervals—also called NNIs—indicate the cardiac cycles. Studies show that there
is a high correlation between the HRV obtained from the ECG and PPG [51–53]. The HRV values are extracted
over a period of the ECG/PPG signal, which is in long term over 24 hours, in short term over 5 minutes, or in
ultra-short term over 1 minute [49, 54, 55].
The HRV can be obtained from the signal both in time domain and frequency domain. The time domain
features of HRV are statistical features that quantify the amount of variability in measurements of the inter-beat-
interval (IBI), which is a time interval between adjacent heartbeats. In contrast, the frequency domain features
of HRV are estimations of distribution of absolute or relative power into four frequency bands mainly based
on PSD. Heart rate oscillations is divided into ultra-low-frequency, very low-frequency (VLF), LF, and HF bands
force [56]. Some common HRV features in both time domain and frequency domain are described in Table 2.

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
 A Deep Learning–based PPG Quality Assessment Approach • 24:9

Table 2. Time Domain and Frequency Domain HRV Features and Their Descriptions

Feature Units Description

AVNN ms Mean of NN intervals
SDNN ms Standard deviation of NN intervals
RMSSD ms Root mean square of successive NN interval differences
SDSD ms Standard deviation of successive NN interval differences
nnXX ms Number of NN interval differences greater than the specified threshold
pnnXX % Percentage of successive NN intervals that differ by more than x ms
VLF power s2 Absolute power in very low frequency band (≤0.04)
LF power s2 Absolute power in low frequency band (0.04–0.15)
HF power s2 Absolute power in high frequency band (0.15–0.4)
LF peak Hz Peak frequency in low frequency band (0.04–0.15)
HF peak Hz Peak frequency in high frequency band (0.15–0.4)
Total Power s2 Total power over all frequency bands
LF/HF % Ratio of LF-to-HF power

The HRV features obtained from the PPG show different characteristics within different sampling frequencies.
The lower the sampling frequency is, the more variation in the peak locations, meaning the more errors in the
HRV analysis [57]. Considering this fact and limitation of our PPG data collection (the sampling rate was 20 Hz),
we only focus on the HR and the following four HRV features. These features, which show insignificant error
rate at fs ≥ 20, are AVNN, RMSSD, SDNN from the time domain and LF/HF (ratio of LF power and HF power)
from the frequency domain [57]. We select these HRV in this study, as they are important for various health
application such as stress monitoring [49].

3.3 Convolutional Neural Networks

CNN are a class of deep neural networks, also known as the state-of-the-art models for image recognition prob-
lems [58]. CNNs are capable of automatically learning spatial hierarchies of features from low- to high-level
patterns. CNN is a hierarchical model consists of convolutional, subsampling, and fully connected layers. The
first two layers—convolution and subsampling—carry out the feature engineering part and the third layer (a fully
connected layer) performs the classification.
The state-of-the-art deep architectures such as VGG, ResNet, and MobileNet implement different approaches
for the classification tasks. The VGG was proposed to increase the depth of the convolutional structure of the
model for achieving better performance [59]. VGG obtained a top-5 error rate of 7.32%. However, only increasing
the depth of the network saturates the accuracy and then degrades it rapidly. Therefore, the ResNet was intro-
duced to address this problem, exploiting the shortcuts or parallel blocks of convolutional filters while building
deeper models [60]. ResNet, however, achieved a top-5 error rate of 3.57%. In contrast, MobileNet was proposed
as a lightweight model to run deep neural networks on personal mobile devices. MobileNet obtained a top-5 error
of 7.5%, almost the same as VGG Network. We leverage the significant performance of these deep architectures
for the PPG quality assessment. In this regard, we convert the PPG signals to PPG snapshots to feed the images
to the CNNs.
Furthermore, CNNs can be harnessed on one-dimensional time-series data sharing the same characteristics
and the same approach as in image-based CNNs [61]. The difference is the structure of the input data and how
the filter, i.e., convolution kernel or feature detector, slides across the data. The model learns to extract features
using a technique called sliding window with a one-dimensional filter over the time series, followed by a non-
linear function to learn non-linear decision boundaries. The model can learn an internal representation of the
time-series data automatically.

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
24:10 • E. Kasaeyan Naeini et al.

4 METHODS AND MATERIALS

The data used in this work are part of a multipurpose study on remote health monitoring. In the following, we
briefly describe the participants, recruitment, data collection, and data annotation in this study.

4.1 Participants and Recruitment

The study was conducted in southern Finland during July–August 2019 by inviting healthy individuals who were
between 18 and 55 years old. The exclusion criteria were if the possible candidates had a diagnosed cardiovas-
cular disease, symptoms of illness during the recruitment, and restrictions on the physical activity or using the
devices in the daily routines. The recruitment started by personally contacting students and staff members of
the University of Turku. Then snowball sampling was used to reach a convenient number of participants. We
aimed for an equal number of female and male participants as gender affects HRV parameters. In face-to-face
meetings, the selected candidates were informed about the purpose of the study and the instructions to use the
devices, i.e., a Shimmer device [62] and a Samsung Gear Sport smartwatch [63]. Written informed consent forms
were also provided to the participants. Forty-six individuals, who agreed to participate in this study, were asked
to wear the devices for 24 hours continuously. The data can be shared by signing a data use agreement. This can
be obtained from the corresponding author upon request.

4.2 Ethics
The study was conducted according to the ethical principles based on the Declaration of Helsinki and the Finnish
Medical Research Act (No. 488/1999). The study protocol received a favorable statement from the ethics commit-
tee (Unversity of Turku, Ethics committee for Human Sciences, Statement No. 44/2019). The participants were
informed about the study both orally and in writing, before their consent was obtained. Participation was vol-
untary and all the participants had the right to withdraw from the study at any time and without giving any
reason. To compensate the time used for the study, each participant got a gift card to grocery store (20 euro) at
the end of the monitoring period when returning the devices.

4.3 Data Collection

We performed home-based ECG and PPG collection, in which the signals were acquired simultaneously and
remotely for 24 hours. In this study, the ECG signal was collected, employing the Shimmer3 ECG [62]. The ECG
test included four limb leads placed on the left arm, right arm, left leg, and right leg. This device provides raw
data using medical grade sensors. It also provides accelerometer, gyroscope, and magnetometer data. Participants
were instructed to place the ECG unit on their chest and to attach the four skin electrodes. The 512-Hz sampling
rate was used in this study, as suggested for clinical trials.
In addition, the Samsung Gear Sport smartwatch was selected for collecting the PPG signal, considering avail-
ability of the raw PPG signals and configurability of the data recording. The watch also has a built-in inertial
measurement unit by which daily physical activity data are extracted. Participants were asked to wear the watch
on the non-dominant hand continuously and tightly enough. Considering constraints of the battery in the Gear
sport watch, we programmed the watch to collect 16 minutes PPG signal in every 30 minutes. In this settings,
there is no need to charge the smartwatch during the one-day experiment. Each PPG record contains 1 minute
of unreliable data (due to sensor calibration) and 15 minutes PPG signals. The PPG signals were collected with a
20-Hz sampling frequency that is suitable to extract HR and HRV parameters.

4.4 Automatic PPG Annotation

The PPG signals should be annotated to develop a quality assessment method. Traditionally, the signals are
manually annotated, where experts label windows of the signals into “reliable” or “unreliable” according to the
shape/structure of the signals. As described in Section 2, such a method is inaccurate when multiple parameters

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
 A Deep Learning–based PPG Quality Assessment Approach • 24:11

Fig. 4. Automatic PPG annotation pipeline.

are obtained using the PPG. The PPG signals should be labeled according to the accuracy of the desired appli-
cations/health parameters. Therefore, different labels should be allocated to a window of the signal, e.g., the
window is reliable for HR and SDNN while it is unreliable for RMSSD (see Figures 1 and 2).
To address this issue, we develop an automatic PPG annotation method, where the PPG signals are labeled
according to the health parameters. In this regard, the obtained parameters are compared with the parameters
extracted from the ECG as the baseline method. The signal is labeled as “reliable” for a parameter if the error is
insignificant. Otherwise, it is labeled as “unreliable.” A schematic representation of the automatic PPG annotation
method is shown in Figure 4.
4.4.1 Pre-Processing. The ECG and PPG signals were collected by two different wearables for 24 hours. There-
fore, there might be a time shift between the two signals (e.g., seconds). To address this issue, we first synchronize
the two signals. We used a cross-correlation method to synchronize the data provided by the ECG device and the
smartwatch. The ECG device and PPG-based smartwatch collected acceleration signals with the same frequency
of ECG and PPG signals, respectively. We extracted the cross-correlation of the signal vector magnitudes of the
acceleration signals. The output indicated the possible time shift between the two signals. Then, the ECG signals
was shifted with respect to the PPG signals if needed. We then segment the 24-hour PPG data into 5-minute win-
dows using sliding technique with an stride of 10 seconds. Considering HR between 30 to 220 beats per minute,
a Butterworth filter [64] was set to only pass heartbeat signals (i.e., 0.5–3.7 Hz).
4.4.2 Peak Detection. The first step of the HRV analysis is to calculate the RR intervals. In the ECG analysis,
the RR intervals are calculated by detecting the QRS complex—with the highest peak and slope in the signal—
and extracting the distance between two adjacent R peaks. For the QRS detection, we used the method proposed
by Laitala et al. [65], as it shows more robust and accurate QRS detection in comparison to the traditional QRS
detection methods such as Pan-Tompkins [66], Christov [67], Hamilton [68], and Engzee [69, 70]. This method
uses a Long Short Term Memory network to obtain the probabilities and locations of the peaks, followed by
extra processes to remove outliers based on anomalous peak–peak distances and obtain the valid peaks. The
last step is to remove noise from the RR intervals—the time intervals between two successive R-peaks—obtained
by subtracting the time of two successive peaks. We used a quotient filter [71] to remove outliers from the RR
intervals that are used for the feature extraction.
In the PPG analysis, HRV parameters are obtained from the the subtle change of pulse periods, i.e., IBI, gen-
erated as a result of the heart activity. To extract the IBI from the PPG signals, we use a peak detection method
proposed by Van Gent et al. [72]. They showed the method obtains an acceptable accuracy to extract HRV values
in comparison to reference ECG signals using Pan-Tompkins QRS algorithm [66] and an open source algorithm
called HRVAS ECGViewer [73]. This method uses an adaptive threshold to accommodate morphology variation
in the PPG waveform, followed by an outlier detection/rejection to extract valid peaks in the signal.
4.4.3 Feature Extraction. The HRV parameters can be calculated using the extracted peaks. In this study,
we only extract four HRV parameters, as the PPG sampling frequency is 20 Hz [57] (see Section 3.2). Within
each time window of the both ECG and PPG, HR and the four HRV features—three time domain and one fre-
quency domain—are extracted. The time domain metrics are obtained using the NN intervals: AVNN is the
Average Value of NN intervals, RMSSD is the Root Mean Square of Successive Differences between normal
heartbeats, and SDNN is the Standard Deviation of NN intervals. LF/HF is also the ratio of the low-frequency to

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
24:12 • E. Kasaeyan Naeini et al.

high-frequency power. In our setup, these features are extracted from the ECG and PPG signals using the HeartPy
Python package [72].
4.4.4 Labeling. As previously mentioned, the PPG signals are divided into 5-minute windows, from which
five features—HR, AVNN, RMSSD, SDNN, and LF/HF—are extracted. The features are tailored to label the PPG
windows as “reliable” or “unreliable.” For each window, the extracted features are compared with the values
obtained from the corresponding ECG signal using an Euclidean distance function. The window is labeled as
“reliable” for a feature if the distance is less than a threshold value obtained according to the range of the fea-
ture [74]. Otherwise, the window is labeled as “unreliable.” Noted that the threshold can be selected according to
the desired accuracy of the feature. This process is performed for the five features. Therefore, five binary labels
are allocated to each PPG window.

5 PPG QUALITY ASSESSMENT APPROACH

In this section, we present two deep learning– (DL) based methods using CNNs for PPG Quality Assessment.
CNN architectures are capable of handling the challenging feature engineering of PPG signals automatically.
This is an advantage over the traditional PPG Quality Assessment methods that were mostly designed to extract
features based on the morphology of the PPG itself. The traditional methods use extracted features to generally
classify the PPG signal as “reliable” or “unreliable” solely based on HR; however, we create a specific model for
each of the HR and HRV parameters separately. The classification is performed using the labels generated from
our automatic ECG-based annotation method. In addition, our data include more scenarios and corner cases
with different artifacts compared to a lab setting based data collection, as our data were collected in the course
of everyday events.
CNN can perform feature extraction and classification without having any knowledge about the data collection.
Moreover, CNN architectures naturally can handle an input with any dimensionality; two-dimensional and one-
dimensional inputs are the most common ones. PPG signals as a time series have potential to be converted to
an image if a 2D model is desired. With this privilege, PPG signals can be fed into the CNN in a usual 1D time-
series signal or in an encoded 2D image. We will leverage three state-of-the-art 2D CNN architectures (VGG16,
ResNet50, and MobileNetV2) that are pre-trained on a huge dataset and transfer the extensive knowledge gained
from other image classification tasks by repurposing the models and fine-tuning them for our problem.
We train a separate model for each feature extracted from the PPG window individually. The input to each
model is a 5-minute PPG signal, and the output is a label obtained via the automatic PPG annotation method. An
overview of the CNN-based approaches is shown in Figure 5. This shows that the 5-minute PPG time series and
one of the HR-HRV features are fed to 1D CNN in the training phase. Moreover, the pre-trained 2D CNN models
(VGG16, ResNet50, and MobileNetV2) are fed with the 5-minute PPG images—encoded using Gramian Angular
Field (GAF)—along with one of the HR-HRV features. In the following, we describe the CNN-based approaches
with different architectures leveraged to perform PPG signal quality assessment.

5.1 1D CNN
We design a customized CNN method trained with the PPG time-series segments as having two convolution
layers with one-dimensional filters trailed by a non-linear ReLU activation unit. Our convolutional layers are fol-
lowed by a batch normalization layer [75], by which the changes in the hidden unit values are reduced. Moreover,
the batch normalization reduces overfitting, since it has a slight regularization effect. The output of the convo-
lution block is then fed into a maxpooling layer to reduce the dimensionality of the data. The learned features
obtained through the convolutional block are flattened to one long vector and pass through a fully connected
neural network before the output layer used to make a prediction. The fully connected layer ideally provides a
buffer between the learned features and the output with the intent of interpreting the learned features before
making a prediction. For our customized model, we will use a standard configuration of kernel size of 3 and

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
 A Deep Learning–based PPG Quality Assessment Approach • 24:13

Fig. 5. Overview of the CNN-based architectures for the PPG quality assessment.

32 and 64 parallel feature maps for the first and second convolutional layers, respectively. The feature maps
are the number of times the input is processed or interpreted, whereas the kernel size is the number of input
timesteps considered as the input sequence is read or processed onto the feature maps.
A grid search is also carried out to optimize the hyperparameters of the model. The efficient Adam version of
stochastic gradient descent with a learning rate of 0.00001 is used to optimize the network, and the binary cross
entropy loss function is used given that we are learning a binary-class classification problem. We analyze the
performance of the customized 1D CNN model trained separately on every single feature of HR, AVNN, RMSSD,
SDNN, and LF/HF to describe the reliability and unreliability of the signal with respect to that feature. For the
training phase, 5-minute PPG signals along with the label for each specific feature are fed to the 1D CNN.

5.2 2D CNN
We utilize three powerful pre-trained CNN networks (VGG, ResNet, and MobileNet) and repurpose them and
fine-tune them with the PPG images. To create suitable inputs for the CNN, the PPG time series need to be
converted to PPG images, while preserving the temporal dependency of the time series. Therefore, we encode
the PPG signals as images using the GAF method [76]. The GAF also contains temporal correlations—similarly
to an image—making the image proper for the CNN. The GAF is created based on the Gram Matrix defined by
the dot product of every couple of vectors. The dot product shows the similarity of the set of vectors. Since the
Gram Matrix produces a matrix with the size of time series squared, we need to reduce the dimensionality of the
input to decrease the amount of computation. In this regard, we use Piece-wise Aggregate Approximation (PAA) to
reduce the dimensionality of the input time series by dividing them into equal-sized, non-overlapping windows
and extract the average in each segment [77]. To build GAFs, PPG as a time-series signal is scaled into [−1,1]
with a Min-Max scaler. Next, PPG length is decreased using the PAA algorithm to a 224 × 224 image. Then, PPG
is converted into the polar coordinates rather than keeping it in the typical Cartesian coordinates. The polar
encoding is then followed by a Gram Matrix like operation on the resulting angles. Three stages of the PPG
encoding is shown in Figure 6.
To leverage the significant performance of CNNs, we test our approach with three architectures: VGG, ResNet,
and MobileNet. VGG stacks the convolutional layers with an increasing number of filters but with the same size
of 3 × 3, since two 3 × 3 filters almost cover a 5 × 5 filter and are also more lightweight in multiplications [59].
Among VGG architectures, VGG16 and VGG19 are the most popular, since both followed the same strategy.
VGG19 is deeper than VGG16, but it is observed that the accuracy is not improved and saturated. The other
architecture that we investigate is ResNet [60]. The success recipe of ResNet for training a deep network is the
residual connections, where each layer is connected not only to the previous layer but also the layer behind the

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
24:14 • E. Kasaeyan Naeini et al.

Fig. 6. Three stages of encoding PPG time series to PPG image with GAF.

previous layer. With this intention, each layer has more information. The last architecture considered in this
study is MobileNet [78]. MobileNet brought a novel idea of replacing a standard convolution with a depthwise
convolution, followed by a pointwise convolution. This way of convolving performs a solitary convolution on
every color channel as opposed to joining every one of the three and smoothing it. This way of convolving helps
to build a smaller model and smaller complexity, which makes it suitable to run on mobile devices. In this article,
we implement the three 2D CNN models (VGG16, ResNet50, and MobileNetV2) and train them for each HR-HRV
feature to describe the reliability and unreliability of the signal with respect to that feature. We pass the encoded
PPG images as 2D inputs to these models along with the corresponding label obtained via the automatic PPG
annotation method of each feature.

6 EXPERIMENTAL RESULTS
In this section, we investigate the performance of the four approaches (i.e., one 1D CNN and three 2D CNN)
described in Section 5. The approaches are utilized to classify the PPG signals according to the reliability of five
PPG-based parameters, i.e., HR, AVNN, RMSSD, SDNN, and LF/HF ratio. In this regard, four models are trained
for each health parameter, resulting in a total of 20 models.
Moreover, we evaluate the performance of our method in comparison to existing methods. In this regard, our
best models for the health parameters are compared with six different PPG signal quality assessment methods.
In the following, we first outline our setup and the data used for the training and testing the models. Then, we
evaluate and compare the models in terms of their performance.

6.1 Setup
In our setup, we use Keras Sequential API from TensorFlow to train and evaluate the CNN models [79]. We need
to pass a couple of parameters to the Keras API including an optimizer, a loss function, and metrics that will be
used in the training and validation phase of the model. Adam optimizer [80], binary cross-entropy loss as a loss
function, and accuracy, f1-score, and area under curve (AUC) metrics are used to compile our CNN models. Our
dataset has skew over one label, much more samples are from Reliable class for most of the classification models,
which makes our classification problem to be an imbalance binary classification. Therefore, during the training
and validation phase, we monitor the training and validation AUC and loss in each epoch [81]. The model with
the highest AUC for the validation set during the optimization process is selected as the best model. That model
will be the checkpoint for next epochs.

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
 A Deep Learning–based PPG Quality Assessment Approach • 24:15

Table 3. Distribution of Train and Test Dataset on Each Annotation Label

Train (23) + Validation (6) Test (7)

Total Segments (# of Subjects) 58,588 14,628
Reliable Unreliable Reliable Unreliable
HR 46,175 12,413 11,544 3,104
AVNN 50,925 7,663 12,732 1,916
RMSSD 16,327 42,261 4,082 10,566
SDNN 25,822 32,766 6,456 8,192
LF/HF ratio 42,388 16,200 10,597 4,050

Table 4. Validation Set Accuracy (ACC), f1-score (F1), and AUC Performance Results of CNN Models for
Various HR-HRV Features
Label HR AVNN RMSSD SDNN LF/HF
Metric ACC F1 AUC ACC F1 AUC ACC F1 AUC ACC F1 AUC ACC F1 AUC
Model
1D_CNN 95.63 96.11 96.21 96.71 97.68 97.71 91.42 91.48 91.69 96.01 96.97 97.09 97.71 97.71 97.71
MobileNet 94.58 93.93 94.63 95.68 95.95 96.93 89.68 89.44 89.46 92.66 93.91 93.66 91.92 91.95 92.93
ResNet50 92.49 92.44 93.52 90.43 90.22 91.32 86.47 85.81 87.17 92.52 92.12 93.66 90.22 90.22 91.31
VGG16 92.38 94.42 94.87 93.36 94.29 94.85 85.72 85.09 85.79 93.42 93.82 93.99 94.36 94.29 93.99
In this table, any bold value represents higher value in each column.

6.2 Training and Test Data Distribution

A total of 36 subjects are recruited for this study. We split the dataset into independent train set, validation set,
and test set. Twenty percent of the whole dataset (7 subject) is used for test set, and the rest is split into training
and validation set, 80% (23 subject) and 20% (6 subject), respectively. For the sake of fair comparison, the models
are trained and validated on the same train and validation dataset. To evaluate the performance of the DL models,
we use an independent test set. The detail of the distribution of the train, validation, and test set is shown in
Table 3.

6.3 Proposed Method Evaluation

As described in Section 5, four CNN models are trained for each PPG parameter, i.e., HR, AVNN, RMSSD, SDNN,
and LF/HF. For each parameter, the model with the maximum AUC score is chosen. In the following, we evaluate
the models on the validation set and test set. Finally, the best CNN models on the test set are selected to be used
for the comparison with the state-of-the-art methods.
6.3.1 HR Models. The overall performance of our proposed CNN models created using the automatic an-
notated label for the HR feature is shown in Table 4. Classification accuracy, f1-score, and Receiver Operating
Characteristic (ROC) area of all DL models for the test set is shown in Figure 7(a). It can be seen that for the HR
feature the 1D CNN model works the best among all the DL models with an accuracy of 95.63%, f1-score of 96.10%,
and AUC of 96.21%. As a result, the assessment of the reliability of PPG signal w.r.t. HR can be performed with
high confidence using a one-dimensional CNN classifier that is less complex compared to the two-dimensional
models.
6.3.2 AVNN Models. Table 4 summarizes the overall performance of our proposed CNN models created using
the automatic annotated label for the AVNN feature, and Figure 7(b) shows the classification accuracy, f1-score,
and ROC area of the DL models for the test set. It can be seen that for the AVNN feature, the DL models show
promising performances in all the metrics. However, the 1D CNN model outperforms the other DL models with

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
24:16 • E. Kasaeyan Naeini et al.

Fig. 7. Comparison of test set accuracy, f1-score, and AUC performance metrics for different DL models.

an accuracy of 96.71%, f1-score of 97.71%, and AUC of 97.71%. Following these results, quality assessment of PPG
signals w.r.t. the AVNN can be done using a one-dimensional CNN classifier with a marginal error.
6.3.3 RMSSD Models. The overall performance of the proposed CNN-based models for RMSSD is shown
in Table 4. The classification accuracy, f1-score, and ROC area of the DL models for the test set are shown in
Figure 7(c). The DL models, built for the RMSSD feature, show a decent performance classifying the PPG signals
as Reliable or Unreliable. For this feature, the 2D models show a mediocre performance comparing to the 1D CNN
model. The performance of the RMSSD feature is worse than HR and AVNN due to the distribution of Reliable
and Unreliable labels, which can be found in Table 3. Unlike HR and AVNN models with a very high Reliable
to Unreliable class ratio, RMSSD models have the lowest Reliable to Unreliable ratio. This makes it difficult for
the optimization process to find the best model. The 1D CNN model outperforms the other DL models with an
accuracy of 91.42%, f1-score of 91.48%, and AUC of 91.68%. Consequently, PPG quality classification based on the
RMSSD feature can still be performed using a one-dimensional CNN classifier.

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
 A Deep Learning–based PPG Quality Assessment Approach • 24:17

Fig. 8. ROC curve of the reliable class for 1D CNN models of all parameters.

6.3.4 SDNN Models. Table 4 summarizes the overall performance of the proposed models for the SDNN fea-
ture. Moreover, Figure 7(d) shows the classification accuracy, f1-score, and ROC area of the DL models for the test
set. It can be seen that all the 2D DL models trained for SDNN have a promising performance. The 1D CNN model
is the best classifier among all DL models with an accuracy of 96.01%, f1-score of 96.97%, and AUC of 97.08%.
Therefore, the reliability of PPG signal w.r.t SDNN can be assessed with high confidence using a one-dimensional
CNN classifier, a smaller and simpler model than the 2D models.

6.3.5 LF/HF Models. The performance of the proposed models for LF/HF is shown in Table 4 (validation set)
and in Figure 7(e) (test set). As indicated, the 2D DL models trained for LF/HF parameter have a promising
performance; however, 1D CNN model outperforms the other DL models with an accuracy of 97.71%, f1-score
of 97.71% and AUC of 97.71%. As a result, the assessment of the reliability of PPG signal w.r.t. LF/HF can be
performed with high confidence using a one-dimensional CNN classifier that is less complex compared to the
two-dimensional models.
Figure 8, represents the ROC curves of the best model for each HR-HRV metric along with the AUC of the
corresponding feature. It can be seen that different features have different characteristic in terms of PPG signal
reliability. The 1D CNN performs exceptionally well classifying all HR-HRV features models into reliable and
unreliable classes. In addition, MobileNetV2 did also a great job for all HR-HRV models to classify them into reli-
able and unreliable classes. Our results show a promising performance for the proposed CNN-based approaches,
through which a binary decision is delivered to indicate PPG signal quality among five different features, HR,
AVNN, RMSSD, SDNN, and LF/HF ratio in a real-time manner.

6.4 Comparison with State-of-the-art Methods

The results in Section 6.3 show that the 1D CNN model outperforms the other DL models in terms of performance
w.r.t to the HR and HRV features. Therefore, the proposed 1D CNN model is selected for the comparison with
the state-of-the-art models. As outlined in Section 2, there is a broad variety of PPG signal quality assessment
methods in the literature. We compared our proposed method with six different methods. First, a rule-based
method [40] is selected for comparison to distinguish low-quality signals, leveraging hierarchical decision rules
combined with simple features, such as absolute amplitude, threshold crossing rate, and autocorrelation function
features. Second, we compare the proposed method with Support Vector Machine, K nearest neighbors, and de-
cision trees algorithms as supervised machine learning algorithms [28]. Furthermore, we compare our proposed
method with an unsupervised method using elliptical envelope [30]. Finally, our proposed method is compared
with Xception [31] as a deep learning approach.

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
24:18 • E. Kasaeyan Naeini et al.

Table 5. Comparison between the Proposed Method and the State-of-the-art PPG
Quality Assessment Methods on Test Set
Label HR AVNN RMSSD SDNN LF/HF
Metric ACC F1 AUC ACC F1 AUC ACC F1 AUC ACC F1 AUC ACC F1 AUC
model
Rule Based
61.19 77.28 67.80 63.08 72.44 68.35 52.13 59.95 61.81 63.08 75.44 71.35 67.65 76.15 65.70
[40]
KNN [26] 87.56 92.20 79.66 92.47 95.71 80.69 93.70 88.62 92.01 85.78 83.60 85.40 79.56 86.23 72.38
SVM [26] 78.70 88.08 50.00 86.84 92.95 50.00 89.54 79.52 84.46 78.39 75.84 78.24 72.42 84.00 50.00
DT [26] 78.70 88.08 50.00 86.84 92.95 50.00 89.54 79.52 84.46 78.39 75.84 78.24 72.42 84.00 50.00
Eliptical
71.54 81.94 57.39 78.89 87.85 53.64 67.35 41.28 59.33 61.76 56.54 61.19 62.54 74.14 53.10
Envelope [30]
Xception [31] 89.30 90.28 85.80 89.54 90.44 85.35 81.03 70.95 77.81 88.54 90.44 85.35 85.09 88.15 84.70
Proposed
95.64 96.12 97.71 96.71 97.68 97.71 91.43 91.48 93.59 94.02 94.98 95.02 94.82 95.22 95.31
Method
In this table, any bold value represents higher value in each column.

Table 5 shows the performance of the proposed method and the aforementioned state-of-the-art algorithms.
The state-of-the-art quality assessment algorithms are only defined for the PPG signal itself and HR feature. How-
ever, our proposed method can assess quality of the PPG signals based on HR and HRV features using separate
models. To make a fair comparison, the labels of the state-of-the-art algorithms are created using our automatic
annotated labeling method described in Section 4.4.4. The proposed method (with the 1D CNN architecture) re-
sults in more accurate signal quality assessment compared to the state-of-the-art. As indicated in the table, the
rule-based and the unsupervised algorithms had the lowest overall performance. However, supervised traditional
machine learning methods, such as KNN, showed promising results for AVNN and RMSSD. The accuracy of KNN
for the RMSSD feature was slightly better than the proposed method. However, for the other features/metrics,
the proposed method performs considerably better. The Xception algorithm, as a deep learning approach, had
better performance, in general, than the traditional supervised machine learning methods. It was able to pursue a
stable performance around 90%. However, the proposed method outperformed the Xception algorithm. However,
the rule-based method [40] is the only algorithm that requires no training phase.

6.5 Limitations and Future Works

This study is limited to HR and a couple of HRV parameters. Future work should include the quality assess-
ment of PPG signal w.r.t. to other frequency domain HRV parameters such as LF power, HF power, and total
power and nonlinear domain HRV features such as SD1 and SD2 (Standard Deviation of major and minor axis
in poincare plot) and sample entropy. These HRV features might have potential use cases in various healthcare
applications.
The length of the data collection is also another limitation of this study. In this study we collected the data for
a period of 24 hours. Future work can be a long-term data collection for such signal quality assessment studies.
Another limitation of our study is that the results does not generalize to unhealthy individuals, as the data
was collected from healthy individuals and models were trained with no prior information about abnormality in
PPG signals. There are studies in the literature showing that the reliability of PPG signals from wearable devices
may be different for different population groups [82, 83]. For instance, patients with cardio vascular diseases have
abnormal PPG morphology with irregular heart beats [47]. This abnormality results in different PPG parameters,
i.e., HR and HRV features compared to the healthy individuals [84]. As a result, further investigation is required
in future work to perform PPG quality assessment of unhealthy subjects.

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
 A Deep Learning–based PPG Quality Assessment Approach • 24:19

7 CONCLUSION
PPG is a non-invasive optical method utilized in various wearable devices, enabling home-based health moni-
toring systems, to continuously acquire vital signs such as HR and HRV. PPG is highly susceptible to motion
artifacts and environmental noise. Low-quality PPG signals negatively impact the accuracy of the extracted
health parameters, leading to inaccurate decision-making. Different studies introduced approaches considering
morphological features of the PPG to determine decision rules for designing PPG quality assessment systems.
Such methods can be only applied on HR detection application, since having an acceptable shape is insufficient
in measuring HRV parameters. We proposed DL-based PPG quality assessment methods using CNNs for (1) HR,
(2) AVNN, (3) RMSSD, (4) SDNN, and (5) LF/HF ratio. We utilized our customized 1D CNN and three 2D deep neu-
ral networks architectures. We presented an automatic annotation method for each HR-HRV parameter, where
the PPG quality was labeled automatically as “reliable” or “unreliable” by comparing against an ECG signal, as the
baseline method for HR and HRV measurements. We evaluated our trained models using 210 hours of PPG data
collected from a home-based health monitoring application on an independent test dataset. Our results indicated
that the proposed 1D CNN method outperforms the other 2D CNN approaches. The accuracy of the customized
1D CNN was 95.63%, 96.71%, 91.42%, 96.01%, and 97.71 for the HR, AVNN, RMSSD, SDNN, and LF/HF ratio, re-
spectively. In addition, we compared the performance of our proposed method with six different state-of-the-art
signal quality assessment methods. Our results indicate that the proposed method performs considerably better
(at least 6% improvement in all performance metrics) than the existing algorithms. We also provided a portable
and open source model implemented in Python for the community to be integrated into their solutions.

REFERENCES
[1] J. Gubbi et al. 2013. Internet of Things (IoT): A vision, architectural elements, and future directions. Fut. Gener. Comput. Syst. 29, 7 (2013),
1645–60.
[2] R. Mieronkoski et al. 2017. The Internet of Things for basic nursing care—A scoping review. Int. J. Nurs. Stud. 69 (2017), 78–90.
[3] M. M. Baig and H. Gholamhosseini. 2013. Smart health monitoring systems: An overview of design and modeling. J. Med. Syst. 37,
2 (2013), 9898.
[4] Iman Azimi, Olugbenga Oti, Sina Labbaf, Hannakaisa Niela-Vilén, Anna Axelin, Nikil Dutt, Pasi Liljeberg, and Amir M Rahmani. 2019.
Personalized maternal sleep quality assessment: An objective IoT-based longitudinal study. IEEE Access 7 (2019), 93433–93447.
[5] Milad Asgari Mehrabadi, Iman Azimi, Fatemeh Sarhaddi, Anna Axelin, Hannakaisa Niela-Vilén, Saana Myllyntausta, Sari Stenholm,
Nikil Dutt, Pasi Liljeberg, and Amir M Rahmani. 2020. Sleep tracking of a commercially available smart ring and smartwatch against
medical-grade actigraphy in everyday settings: Instrument validation study. JMIR Mhealth Uhealth 8, 10 (2 Nov. 2020), e20465. DOI:
[Link]
[6] J. Allen. 2007. Photoplethysmography and its application in clinical physiological measurement. Physiol. Meas. 28, 3 (2007), 1–39.
[7] T. Tamura et al. 2014. Wearable photoplethysmographic sensors—Past and present. Electronics 3, 2 (2014), 282–302.
[8] Sumit Majumder, Tapas Mondal, and M Jamal Deen. 2017. Wearable sensors for remote health monitoring. Sensors 17, 1 (2017), 130.
[9] A. Pantelopoulos and N. G. Bourbakis. 2010. A survey on wearable sensor-based systems for health monitoring and prognosis. Trans.
Syst. Man Cyber. C 40, 1 (2010), 1–12.
[10] N. Constant et al. 2015. Pulse-glasses: An unobtrusive, wearable HR monitor with Internet-of-Things functionality. In Proceedings of
the IEEE 12th International Conference on Wearable and Implantable Body Sensor Networks. IEEE, Cambridge, MA, USA, 1–5.
[11] Delaram Amiri, Arman Anzanpour, Iman Azimi, Marco Levorato, Amir M. Rahmani, Pasi Liljeberg, and Nikil Dutt. 2018. Edge-assisted
sensor control in healthcare IoT. In Proceedings of the IEEE Global Communications Conference (GLOBECOM’18). IEEE, 1–6. DOI:http://
[Link]/10.1109/GLOCOM.2018.8647457
[12] Delaram Amiri, Arman Anzanpour, Iman Azimi, Marco Levorato, Pasi Liljeberg, Nikil Dutt, and Amir M. Rahmani. 2020. Context-aware
sensing via dynamic programming for edge-assisted wearable systems. ACM Trans. Comput. Healthcare 1, 2 (2020), 1–25.
[13] Seyed Amir Hossein Aqajari, Rui Cao, Amir Hosein Afandizadeh Zargari, and Amir M. Rahmani. 2021. An end-to-end and accurate PPG-
based respiratory rate estimation approach using cycle generative adversarial networks. In Proceedings of the 43rd Annual International
Conference of the IEEE Engineering in Medicine and Biology Society (EMBC’21). IEEE, 744–747.
[14] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. 2021. Smil: Multimodal learning with severely missing
modality. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35, 2302–2310.
[15] C. Zong and R. Jafari. 2015. Robust heart rate estimation using wrist-based PPG signals in the presence of intense physical activities.
In Proceedings of the IEEE Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC’15). IEEE,
8078–8082.

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
24:20 • E. Kasaeyan Naeini et al.

[16] H. Han et al. 2007. Development of real-time motion artifact reduction algorithm for a wearable photoplethysmography. In Proceedings
of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC’07) IEEE, Lyon, France, 1538–1541.
[17] Simhadri Vadrevu and M. Sabarimalai Manikandan. 2019. Real-time PPG signal quality assessment system for improving battery life
and false alarms. IEEE Trans. Circ. Syst. II: Expr. Briefs 66, 11 (2019), 1910–1914.
[18] M. Elgendi. 2016. Optimal signal quality index for photoplethysmogram signals. Bioengineering 3, 4 (2016), 21.
[19] Jiajia Song, Dan Li, Xiaoyuan Ma, Guowei Teng, and Jianming Wei. 2019. PQR signal quality indexes: A method for real-time photo-
plethysmogram signal quality estimation based on noise interferences. Biomed. Sign. Process. Contr. 47 (2019), 88–95.
[20] Hyeon Seok Seok, Sangjin Han, Junyung Park, Donggeun Roh, and Hangsik Shin. 2018. Photoplethysmographic pulse quality assess-
ment methods based on similarity analysis. In Proceedings of the Joint 10th International Conference on Soft Computing and Intelligent
Systems (SCIS’18) and 19th International Symposium on Advanced Intelligent Systems (ISIS’18). IEEE, 350–353.
[21] X. Sun et al. 2012. Assessment of photoplethysmogram signal quality using morphology integrated with temporal information approach.
In Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC’12). IEEE, 3456–3459.
[22] Gabriele B. Papini, Pedro Fonseca, Xavier L. Aubert, Sebastiaan Overeem, Jan W. M. Bergmans, and Rik Vullings. 2017. Photoplethys-
mography beat detection and pulse morphology quality assessment for signal reliability estimation. In Proceedings of the 39th Annual
International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC’17). IEEE, 117–120.
[23] Arlene John, Barry Cardiff, and Deepu John. 2020. A generalized signal quality estimation method for IoT sensors. In Proceedings of the
IEEE International Symposium on Circuits and Systems (ISCAS’20). IEEE, 1–5.
[24] Shahnawaz Alam, Shreyasi Datta, Anirban Dutta Choudhury, and Arpan Pal. 2017. Sensor agnostic photoplethysmogram signal quality
assessment using morphological analysis. In Proceedings of the 14th EAI International Conference on Mobile and Ubiquitous Systems:
Computing, Networking and Services. Association for Computing Machinery, New York, NY, 176–185.
[25] Yue Zhang and Junjun Pan. 2017. Assessment of photoplethysmogram signal quality based on frequency domain and time series pa-
rameters. In Proceedings of the 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-
BMEI’17). IEEE, 1–5.
[26] Tania Pereira, Kais Gadhoumi, Mitchell Ma, Rene Colorado, Kevin J. Keenan, Karl Meisel, and Xiao Hu. 2018. Robust assessment of
photoplethysmogram signal quality in the presence of atrial fibrillation. In Proceedings of the Computing in Cardiology Conference
(CinC’18), Vol. 45. IEEE, 1–4.
[27] Elyas Sabeti, Narathip Reamaroon, Michael Mathis, Jonathan Gryak, Michael Sjoding, and Kayvan Najarian. 2019. Signal quality mea-
sure for pulsatile physiological signals using morphological features: Applications in reliability measure for pulse oximetry. Inf. Med.
Unlocked 16 (2019), 100222.
[28] Tania Pereira, Kais Gadhoumi, Mitchell Ma, Liu Xiuyun, Ran Xiao, Rene A. Colorado, Kevin J. Keenan, Karl Meisel, and Xiao Hu. 2019.
A supervised approach to robust photoplethysmography quality assessment. IEEE J. Biomed. Health Inf. 45 (2019), 1–4.
[29] Q. Li and G. D. Clifford. 2012. Dynamic time warping and machine learning for signal quality assessment of pulsatile signals. Physiol.
Meas. 33 (2012), 1491–1501.
[30] Aysan Mahmoudzade, Iman Azimi, Amir M. Rahmani, and Pasi Liljeberg. 2021. Lightweight photoplethysmography quality assessment
for real-time IoT-based health monitoring using unsupervised anomaly detection. Proced. Comput. Sci. 184 (2021), 140–147.
[31] Tania Pereira, Cheng Ding, Kais Gadhoumi, Nate Tran, Rene A. Colorado, Karl Meisel, and Xiao Hu. 2019. Deep learning approaches
for plethysmography signal quality assessment in the presence of atrial fibrillation. Physiol. Meas. 40, 12 (2019), 125002.
[32] Jessica Torres Soto and Euan Ashley. 2020. DeepBeat: A multi-task deep learning approach to assess signal quality and arrhythmia
detection in wearable devices. arXiv:2001.00155. Retrieved from [Link]
[33] Emad Kasaeyan Naeini, Iman Azimi, Amir M. Rahmani, Pasi Liljeberg, and Nikil Dutt. 2019. A real-time PPG quality assessment ap-
proach for healthcare Internet-of-Things. Proc. Comput. Sci. 151, C (2019).
[34] Choon-Hian Goh, Li Kuo Tan, Nigel H. Lovell, Siew-Cheok Ng, Maw Pin Tan, and Einly Lim. 2020. Robust PPG motion artifact detection
using a 1-D convolution neural network. Comput. Methods Progr. Biomed. 196 (2020), 105596.
[35] Donggeun Roh and Hangsik Shin. 2021. Recurrence plot and machine learning for signal quality assessment of photoplethysmogram
in mobile environment. Sensors 21, 6 (2021), 2188.
[36] Philip Schmidt et al. 2018. Introducing wesad, a multimodal dataset for wearable stress and affect detection. In Proceedings of the ACM
International Conference on Multimodal Interaction (ICMI’18). Association for Computing Machinery, New York, NY.
[37] Christina Orphanidou, Timothy Bonnici, Peter Charlton, David Clifton, David Vallance, and Lionel Tarassenko. 2014. Signal-quality
indices for the electrocardiogram and photoplethysmogram: Derivation and applications to wireless monitoring. IEEE J. Biomed. Health
Inf. 19, 3 (2014), 832–838.
[38] Gangireddy Narendra Kumar Reddy, M. Sabarimalai Manikandan, and NVL Narasimha Murty. 2020. On-device integrated PPG quality
assessment and sensor disconnection/saturation detection system for IoT health monitoring. IEEE Trans. Instrum. Meas. 69, 9 (2020).
[39] Konstantin Tyapochkin, Evgeniya Smorodnikova, and Pavel Pravdin. 2019. Smartphone PPG: Signal processing, quality assessment, and
impact on HRV parameters. In Proceedings of the 41st Annual International Conference of the IEEE Engineering in Medicine and Biology
Society (EMBC’19). IEEE, 4237–4240.
[40] Simhadri Vadrevu and M. Sabarimalai Manikandan. 2019. A new quality-aware quality-control data compression framework for power
reduction in IoT and smartphone PPG monitoring devices. IEEE Sens. Lett. 3, 7 (2019), 1–4.
ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
 A Deep Learning–based PPG Quality Assessment Approach • 24:21

[41] K. Li et al. 2012. Onboard tagging for real-time quality assessment of photoplethysmograms acquired by a wireless reflectance pulse
oximeter. IEEE Trans. Biomed. Circ. Syst. 6, 1 (2012), 54–63.
[42] Monalisa Singha Roy, Rajarshi Gupta, and Kaushik Das Sharma. 2020. Photoplethysmogram signal quality evaluation by unsupervised
learning approach. In Proceedings of the IEEE Applied Signal Processing Conference (ASPCON’20). IEEE, 6–10.
[43] Emad Kasaeyan Naeini, Sina Shahhosseini, Ajan Subramanian, Tingjue Yin, Amir M. Rahmani, and Nikil Dutt. 2019. An edge-assisted
and smart system for real-time pain monitoring. In Proceedings of the IEEE/ACM International Conference on Connected Health: Applica-
tions, Systems and Engineering Technologies (CHASE’19). IEEE, 47–52.
[44] Vega Pradana Rachim and Wan-Young Chung. 2016. Wearable noncontact armband for mobile ECG monitoring system. IEEE Trans.
Biomed. Circ. Syst. 10, 6 (2016), 1112–1118.
[45] Michael T. Petterson, Valerie L. Begnoche, and John M. Graybeal. 2007. The effect of motion on pulse oximetry and its clinical signifi-
cance. Anesthes. Analges. 105, 6 (2007), S78–S84.
[46] P. Cheang and P. Smith. 2003. An overview of non-contact photoplethysmography. Electr. Syst. Contr. Div. Res. 1, 1 (2003).
[47] Rollin McCraty and Fred Shaffer. 2015. Heart rate variability: New perspectives on physiological mechanisms, assessment of self-
regulatory capacity, and health risk. Glob. Adv. Health Med. 4, 1 (2015), 46–61.
[48] Mingzhe Jiang, Riitta Mieronkoski, Amir M Rahmani, Nora Hagelberg, Sanna Salanterä, and Pasi Liljeberg. 2017. Ultra-short-term
analysis of heart rate variability for real-time acute pain monitoring with wearable electronics. In Proceedings of the IEEE International
Conference on Bioinformatics and Biomedicine (BIBM’17). IEEE, 1025–1032.
[49] Lizawati Salahuddin, Jaegeol Cho, Myeong Gi Jeong, and Desok Kim. 2007. Ultra short term analysis of heart rate variability for mon-
itoring mental stress in mobile settings. In Proceedings of the 29th Annual International Conference of the IEEE Engineering in Medicine
and Biology Society. IEEE, 4656–4659.
[50] Sandrine Devot, Anna M. Bianchi, Elke Naujoka, Martin O. Mendez, Andreas Braurs, and Sergio Cerutti. 2007. Sleep monitoring through
a textile recording system. In Proceedings of the 29th Annual International Conference of the IEEE Engineering in Medicine and Biology
Society. IEEE, 2560–2563.
[51] X. F. Teng and Y. T. Zhang. 2003. Study on the peak interval variability of photoplethysmogtaphic signals. In Proceedings of the IEEE
EMBS Asian-Pacific Conference on Biomedical Engineering. IEEE, 140–141.
[52] Nandakumar Selvaraj, Ashok Jaryal, Jayashree Santhosh, Kishore K. Deepak, and Sneh Anand. 2008. Assessment of heart rate variability
derived from finger-tip photoplethysmography as compared to electrocardiography. J. Med. Eng. Technol. 32, 6 (2008), 479–484.
[53] M. Bolanos, H. Nazeran, and E. Haltiwanger. 2006. Comparison of heart rate variability signal features derived from electrocardiography
and photoplethysmography in healthy individuals. In Proceedings of the International Conference of the IEEE Engineering in Medicine
and Biology Society. IEEE, New York, NY, 4289–4294.
[54] Musa Sesay, Georges Robin, Patrick Tauzin-Fin, Oumar Sacko, Edouard Gimbert, Jean-Rodolphe Vignes, Dominique Liguoro, and Karine
Nouette-Gaulain. 2015. Responses of heart rate variability to acute pain after minor spinal surgery: Optimal thresholds and correlation
with the numeric rating scale. J. Neurosurg. Anesthesiol. 27, 2 (2015), 148–154.
[55] Arto J. Hautala, Jaro Karppinen, and Tapio Seppänen. 2016. Short-term assessment of autonomic nervous system as a potential tool
to quantify pain experience. In Proceedings of the 38th Annual International Conference of the IEEE Engineering in Medicine and Biology
Society (EMBC’16). IEEE, 2684–2687.
[56] A. John Camm, Marek Malik, J. Thomas Bigger, Günter Breithardt, Sergio Cerutti, R. J. Cohen, Philippe Coumel, E. L. Fallen, H. L.
Kennedy, R. E. Kleiger, et al. 1996. Heart rate variability: Standards of measurement, physiological interpretation and clinical use. Task
Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology. Circulation. 93, 5 (1996),
1043–1065.
[57] Ahyoung Choi and Hangsik Shin. 2017. Photoplethysmography sampling frequency: Pilot assessment of how low can we go to analyze
pulse rate variability with reliability? Physiol. Meas. 38, 3 (2017), 586.
[58] Y. LeCun and Y. Bengio. 1998. Convolutional networks for images, speech, and time series. In The Handbook of Brain Theory and Neural
Networks. MIT Press, Cambridge, MA, 255–258.
[59] Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks for large-scale image recognition. arXiv:1409.1556.
Retrieved from [Link]
[60] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition. IEEE, 770–778.
[61] Nagarajan Ganapathy, Ramakrishnan Swaminathan, and Thomas M. Deserno. 2018. Deep learning on 1-D biosignals: A taxonomy-based
survey. Yearbk. Med. Inf. 27, 01 (2018), 098–109.
[62] Shimmer. ECG Sensor Development Kit, Wearable ECG Sensor, Wireless ECG. Retrieved January 2019 from [Link]
[Link]/products/ecg-development-kit
[63] Samsung. Samsung Gear Sport Smartwatch. Retrieved March 2020 from [Link]
[64] I. W. Selesnick and C. S. Burrus. 1998. Generalized digital Butterworth filter design. IEEE Trans. Sign. Process. 46, 6 (1998), 1688–1694.

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.
24:22 • E. Kasaeyan Naeini et al.

[65] Juho Laitala, Mingzhe Jiang, Elise Syrjälä, Emad Kasaeyan Naeini, Antti Airola, Amir M. Rahmani, Nikil D. Dutt, and Pasi Liljeberg.
2020. Robust ECG R-peak detection using LSTM. In Proceedings of the 35th Annual ACM Symposium on Applied Computing. Association
for Computing Machinery, New York, NY, 1104–1111.
[66] Jiapu Pan and Willis J. Tompkins. 1985. A real-time QRS detection algorithm. IEEE Trans. Biomed. Eng. 32, 3 (1985), 230–236.
[67] Ivaylo I. Christov. 2004. Real time electrocardiogram QRS detection using combined adaptive threshold. Biomed. Eng. Online 3, 1 (2004),
28.
[68] Pat Hamilton. 2002. Open source ECG analysis. In Computers in Cardiology. IEEE, 101–104.
[69] Willem A. H. Engelse and Cees Zeelenberg. 1979. A single scan algorithm for QRS-detection and feature extraction. J. Biomed. Sci. Eng.
1, 1 (1979), 37–42.
[70] André Lourenço, Hugo Silva, Paulo Leite, Renato Lourenço, and Ana L. N. Fred. 2012. Real time electrocardiogram segmentation for
finger based ECG biometrics. In Biosignals. Semantic Scholar, NA, 49–54.
[71] Jarosław Piskorski and Przemysław Guzik. 2005. Filtering poincare plots. Comput. Methods Sci. Technol. 11, 1 (2005), 39–48.
[72] Paul van Gent, Haneen Farah, Nicole van Nes, and Bart van Arem. 2019. HeartPy: A novel heart rate algorithm for the analysis of noisy
signals. Transport. Res. Part F: Traffic Psychol. Behav. 66 (2019), 368–378.
[73] Ramshur. ECG Viewer. Retrieved March 2020 from [Link]
[74] Tania Pereira, Pedro R. Almeida, Joao P. S. Cunha, and Ana Aguiar. 2017. Heart rate variability metrics for fine-grained stress level
assessment. Comput. Methods Progr. Biomed. 148 (2017), 71–80.
[75] Sergey Ioffe and Christian Szegedy. 2015. Batch normalization: Accelerating deep network training by reducing internal covariate shift.
In International Conference on Machine Learning. PMLR, 448–456.
[76] Zhiguang Wang and Tim Oates. 2015. Imaging time-series to improve classification and imputation. In Proceedings of the 24rth Interna-
tional Joint Conference on Artificial Intelligence. AAAI Press, 3939–3945.
[77] Eamonn J. Keogh and Michael J. Pazzani. 2000. Scaling up dynamic time warping for datamining applications. In Proceedings of the 6th
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Association for Computing Machinery, New York, NY,
285–289.
[78] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. 2018. Mobilenetv2: Inverted residuals
and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE Computer Society, Los
Alamitos, CA, 4510–4520.
[79] M. Abadi et al. 2015. Tensorflow: Large-Scale Machine Learning on Heterogeneous Systems. Retrieved from [Link]
org/
[80] D. Kinga, Jimmy Ba Adam, et al. 2015. A method for stochastic optimization. In International Conference on Learning Representations
(ICLR), Vol. 5, San Diego, CA, 6 page.
[81] Marina Sokolova and Guy Lapalme. 2009. A systematic analysis of performance measures for classification tasks. Inf. Process. Manage.
45, 4 (2009), 427–437.
[82] Jesse D. Cook, Michael L. Prairie, and David T. Plante. 2018. Ability of the multisensory jawbone UP3 to quantify and classify sleep in
patients with suspected central disorders of hypersomnolence: A comparison against polysomnography and actigraphy. J. Clin. Sleep
Med. 14, 5 (2018), 841–848.
[83] Jesse D. Cook, Michael L. Prairie, and David T. Plante. 2017. Utility of the fitbit flex to evaluate sleep in major depressive disorder: A
comparison against polysomnography and wrist-worn actigraphy. J. Affect. Disord. 217 (2017), 299–305.
[84] Jay Chen, Stephen L. Wasmund, and Mohamed H. Hamdan. 2006. Back to the future: The role of the autonomic nervous system in atrial
fibrillation. Pacing Clin. Electrophysiol. 29, 4 (2006), 413–421.

Received 18 June 2021; revised 8 March 2023; accepted 14 July 2023

ACM Transactions on Computing for Healthcare, Vol. 4, No. 4, Article 24. Publication date: November 2023.