Received 15 August 2023, accepted 13 September 2023, date of publication 18 September 2023,

date of current version 27 September 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3316508

A Novel Digital Twin (DT) Model Based on WiFi

CSI, Signal Processing and Machine Learning
for Patient Respiration Monitoring and
Decision-Support
SAGHEER KHAN 1, AAESHA ALZAABI 1, (Member, IEEE), ZAFAR IQBAL1 ,
THARMALINGAM RATNARAJAH 1 , (Senior Member, IEEE),
AND TUGHRUL ARSLAN 1,2 , (Senior Member, IEEE)
1 School of Engineering, The University of Edinburgh, EH9 3FF Edinburgh, U.K.
2 Advanced Care Research Centre (ACRC), The University of Edinburgh, EH16 4UX Edinburgh, U.K.
Corresponding author: Sagheer Khan ([email protected])
This work was supported by the Advanced Care Research Centre (ACRC), The University of Edinburgh, Edinburgh, U.K.

ABSTRACT Digital Twin (DT) in Healthcare 4.0 (H4.0) presents a digital model of the patient with all
its biological properties and characteristics. One of the application areas is patient respiration monitoring
for enhanced patient care and decision support to healthcare professionals. Obtrusive methods of patient
monitoring create hindrances in the patient’s daily routine. This research presents a novel Respiration DT
(ResDT) model based on Wi-Fi Carrier State Information (CSI), improved signal processing, and Machine
Learning (ML) algorithms for monitoring and classification (binary and multi-class) of patient respiration.
A Wi-Fi sensor ESP32 with Wi-Fi CSI was utilized for the collection of respiration data. This provides an
added advantage of unobtrusive monitoring of patient vital signs. The Patient’s Breaths Per Minute (BPM)
is estimated from raw sensor data through the integration of multiple signal processing methodologies for
denoising (smoothing and filtering) and dimensionality reduction (PCA, SVM, EMD, EMD-PCA). Multiple
filters and dimensionality reduction methodologies are compared for accurate BPM estimation. The elliptical
filter provides a relatively better estimation of the BPM with 87.5% accurate estimation as compared to other
bandpass filters such as Butterworth (BF), Chebyshev type 1 Filter (CH1), Chebyshev type 2 Filter (CH2),
and wavelet Decomposition (62.5%, 75%, 68.75%, and 75% respectively). Principal Component Analysis
(PCA) was performed to provide better dimensionality reduction with 87.5% accurate BPM values compared
to EMD, SVD, and EMD-PCA (57%, 44%, and 44% respectively). Additionally, the fine tree algorithm, from
the implemented 21 ML supervised classification algorithms with K-fold cross-validation, was observed to
be the optimal choice for multi-class and binary-class classification problems in the presented ResDT model
with 96.9% and 95.8% accuracy respectively.

INDEX TERMS Digital twin, machine learning (ML), principle component analysis (PCA), respiration rate
estimation, signal processing, unobtrusive Wi-Fi sensor.
I. INTRODUCTION connection between the digital, physical, and biological
Advancements in technology have drastically transformed spheres. The amalgamation of a variety of technologies, such
our world from time to time. Healthcare has been reshaped as Artificial Intelligence (AI) [1], Internet of Things (IoT)
into smart and connected healthcare, known as Healthcare 4.0 [2], [3], Cyber-Physical Systems (CPS) [4], big data [5], [6],
(H4.0). It is a combination of technologies to create a etc., are utilized to establish automation and data exchange.
Fig. 1. represents the revolution in healthcare. Healthcare 1.0
The associate editor coordinating the review of this manuscript and is at the stage of doctor and patient interactions. The
approving it for publication was Ghufran Ahmed . doctor will examine the patient at the clinic and provide

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

103554 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

a diagnosis/medication/care routine based on test results with AI model was implemented in the radiology department
and consultations. Healthcare 2.0 is the implementation to improve department operations of processing sensor data,
of medical devices in monitoring, surgeries, imaging, and patient demands, clinical complexity, and improving infras-
life support. Healthcare 3.0 is about utilizing healthcare tructure. Capacity Command Center was designed by GE
information technologies to employ computer networks and Healthcare for decision support and simulation capabilities at
digitized records to make remote care and telehealth a reality. the John Hopkins Hospital in Baltimore. A DT model of the
Creating a gateway between the patient and doctor to replace heart was created by [19]. The human respiratory system was
face-to-face encounters. Being able to do virtual meetings developed by Oklahoma State University’s Computational
proved to be a major advantage in the recent COVID-19 Biofluidics and Biomechanics Laboratory [20], [21], [22].
pandemic. H4.0 is rising in parallel to I4.0 and is proving Most of the healthcare DT models enabled by AI-ML are
itself in healthcare to be revolutionary. Healthcare evolved of humans [23], [24], [25]. According to [26], at this stage,
with the use of AI, IoT, CPS, big data, and cloud computing replicating the full functionality of a human is not possible.
technologies working together [7] to provide a quick and Thus, the researchers are focusing on utilizing DT on specific
accurate diagnosis of patient big data. Increasing connectivity aspects of human biology. Fig. 2. represents the multiple
between the medical equipment and healthcare personnel to application areas of DT in H4.0.
enhance patient care, recovery, and experience.

FIGURE 1. Healthcare revolution.

In the last decade, healthcare has kept on improving by

providing smart healthcare services. The idea of Digital
Twin (DT) has already advanced industries but is going to
revolutionize H4.0 [7], [8], [9], [10]. DT allows the formation
of a digital representation of the physical system with all FIGURE 2. DT applications in H4.0.
its traits. A bi-directional link between the physical system
and its digital model [6], [11]. Creating such a link is the The paper is distributed in sections for better understanding
task of multiple technologies to work in cooperation with and improving the layout. Section II provides the motivation
each other. Khan et al. [7] provides a substantial review of and contribution of this research. Section III discusses Wi-Fi
such technologies working together for the creation of DT sensing and sensing in DT, Section IV is the preliminary
healthcare. The tasks of real-time predictions, comparisons, discussions on Wi-Fi CSI, PCA, Empirical Mode Decom-
and simulations of a patient and the environment based on position (EMD), and Single Variable Decomposition (SVD),
real-time sensor data can become a reality with DT. DT in Section V provides an overview of data collection with
healthcare can provide early detection of abnormalities, quick ESP32 sensor, Section VI is the detailed performance analysis
and accurate diagnosis, predictive analysis of patient vitals of multiple pre-processing techniques on raw Wi-Fi CSI
and procedures, the most effective and risk-free surgeries, and sensor data in the creation of the ResDT model, Section VII
improved patient monitoring. discusses various ML algorithms accuracies for the multiple
Over the years researchers have worked on implementing classification problem, Section VIII discusses the conclusion
DT in healthcare through various methods. H. Laak et al. [12] and future work.
provided a framework for DT in remote surgeries. To detect
abnormalities and diagnosis, an Electrocardiogram (ECG) II. PROBLEM STATEMENT AND CONTRIBUTION
signal rhythms classifier based on Machine Learning (ML) A. PROBLEM STATEMENT
is implemented [13]. An elaboration of DT in medicine, such Overall, life expectancy has increased, reaching the sixties
as medical CPS, and ML in healthcare is discussed in [14], and beyond. Countries across the world are seeing an increase
[15], and [16]. An overview of challenges and technologies in their older populations. It is predicted by the World Health
in implementing DT in personal healthcare is reviewed in [17] Organization (WHO) that by 2030, one out of six people
and [18]. Siemens Healthineers optimized Mater Private would be 60+ years of age. By 2050, the number of people 60
Hospital (MPH) with the help of DT technology. DT model and older will have doubled [27]. This creates a burden on

VOLUME 11, 2023 103555

 S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

the healthcare system and resources [28], [29]. Not only for unobtrusive monitoring of the patient/elderly vital signs
with age, but people of any age can present multiple health without the need for clear LOS, no privacy violation,
conditions with breathing being a crucial indicator of the or the need for special arrangements. Wi-Fi CSI has the
health condition. advantage of monitoring the patient’s physical, physiological,
It is elaborated in [30] and [31] that DT in H4.0 has and biological features without the technology creating
application areas of high-quality economical healthcare, obstruction in the daily routine.
improved patient monitoring, accurate diagnosis and medica-
tion, preventing emergencies, and early diagnosis of medical B. CONTRIBUTION
conditions and decision-support. However, it is an open Following are the novel points of this research toward
research field with problems of optimal signal processing implementing DT technology in healthcare for monitoring
methodology, how big data analytics can be implemented, the Breaths Per Minute (BPM) of a patient with respiration
which AI and ML algorithms can improve decision-making problems through unobtrusive Wi-Fi sensing with integration
and predictive analysis, which communication technology is of signal processing methodologies and ML algorithm.
best suited in a specific healthcare scenario, cyber security for 1) We propose a novel DT model (ResDT) based on
sensor and patient data protection, etc. [7]. Wi-Fi CSI for patient respiration rate. ESP32 sensor
In [32], around 5% of the population in prosperous with Wi-Fi CSI allows for an unobtrusive method of
countries suffer from the breathing disorder of bradypnea respiration data collection for the DT model.
(slow breathing), and nearly 30% of the people in the 2) Our study is the investigation of signal process-
study, in their 60s, showed symptoms of multiple breathing ing techniques performance analysis comparison on
diseases. In clinical medicine, impedance pneumography, and complex noisy ESP32 sensor data in the proposed
capnography are two of the conventional methodologies for ResDT model. This will help select the optimal
continuously checking the breathing rate. They are intrusive methodology for estimating accurate BPM from raw
respiration monitoring methods which may not be considered sensor data. Various techniques comparison comprise;
in the future for elderly care home or hospital environment. multiple Infinite Impulse Response (IIR) bandpass
Based on the advancements in sensing, signal processing and filters (0.15Hz to 0.6Hz) and wavelet decomposition
machine learning, they can be incorporation for enhanced performance comparison for denoising the respiration
respiration monitoring methodologies. sensor data; multiple dimensionality reduction method-
Unobtrusive sensing methods have a more promising ologies comparison on complex Wi-Fi CSI respiration
future in healthcare, especially for elderly care. They do datasets in the estimation of the patient’s BPM.
not create hindrance in patient daily life as compared to 3) In our research, 21 ML-based supervised learn-
obtrusive methods. Compared to obtrusive data collection, ing classification algorithm accuracies, along with
one such solution can be unobtrusive sensor data collection. K-fold cross-validation, are assessed for the various
Liu et al. [24] worked on creating cloud-based DT Healthcare classification problem of abnormal and normal respi-
(CloudDTH) for real-time monitoring of the patient. The ration rates in the proposed ResDT model. This will
data acquired through the ECG sensors are sent to the help in implementing decision-support or classification
CloudDTH platform through the ECG device (Huake HKW- between binary and multi-class problems in proposed
10). The authors presented a contact based DT model for data ResDT model.
collection. Jagade et al. [33] worked on monitoring human
respiration rate through Infrared Thermography (IRT) but C. PROPOSED ResDT MODEL
that requires clear LOS between the infrared camera and Fig. 3 represents the proposed Respiration DT (ResDT)
the patient. Chu et al. [34] monitored the respiration rate model. ESP32 Wi-Fi sensor with CSI characteristics collects
of the human through low-powered piezo-resistive sensors patient respiration data unobtrusively. The raw data is not
attached to the human body. The sensors are combined ready to analyzed for the estimation of the patient’s BPM. It is
with Bluetooth units to observe respiration rate from a subjected to multiple denoising methodologies of smoothing,
distance but the sensors can create hindrance in daily routines. bandpass filters, and PCA to acquire the patient BPM from
Ryser et al. [35] utilized a single chest-worn accelerometer raw sensor data. Feature data is extracted, enabling the
to estimate the respiration rate. Respiration and abdominal testing of ML algorithms and neural networks for binary and
bands are attached on the around the chest. Even with multi-class classification in the proposed ResDT model.
unobtrusive sensing methods, there is the limitation of clear
LOS, privacy concerns, or the need of putting on special III. SENSING
equipment by the patient. A. Wi-Fi SENSING
In the above-mentioned literature along with the work The recent COVID-19 pandemic has shown how conven-
presented in [36], [37], and [38], they have the limitation tional methodologies fail in providing healthcare in such
of obtrusive sensing, the need for clear LOS, and restricts harsh conditions [39], [40]. Not only that, with the increasing
the patient’s movements. Hence, unobtrusive sensor data number of elderly patients, it is important to look for
collected with Wi-Fi CSI can provide a better solution new innovative noninvasive techniques to overcome these

103556 VOLUME 11, 2023

S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

FIGURE 3. ResDT: proposed respiration DT model with ESP32 sensor.

challenges. Researchers can make use of cameras indoors wireless signals can still deliver valuable sensor data
to capture images of respiration signals [41]. The major even if there is an obstruction in the middle of the
drawback of this camera-based respiration monitoring is device and the user.
that the images require a clear line of sight and sufficient 4) The wireless signal transmission and reception do not
illumination [42], [43]. An additional drawback is that depend on the light, making it possible to work around
the cameras are a privacy concern. Under the threat of the clock.
cybercrimes, it is difficult to have a promising solution
for indoor sensing [44]. Wearable devices can be utilized B. SENSING IN DT
as a solution for unobtrusive sensing [45]. Even though Wi-Fi sensing is not only important for the unobtrusive
wearable devices have a broad range of applications, they respiration monitoring of a patient but also crucial in the
need exclusive gear to be worn for extended periods and creation of DT models. The DT’s popularity increased with
are only appropriate for extended periods of monitoring an the availability of economical and miniature-size sensors that
immobile patient [46]. A smartphone can be a solution to can provide all kinds of patient vital information. These
acquire data from the subject’s body but it needs to be models are data-driven and can only represent the theoretical
carried all the time [47], [48]. Another solution can be the behavior of a system if the actual state in the physical world
environmentally installed sensors but they require a heavy is unknown. The choice of the sensor itself is critical as
installment setup [49], [50]. it impacts the functionality and accuracy of the DT model.
A real-time economical unobtrusive respiratory monitor- Khan et al. [44] discussed the importance of the sensor’s
ing method is essential. Wi-Fi sensing can be utilized to position in data collection for the creation of DT models.
sense the vital signs of a patient from a distance [51], [52]. If the sensor utilized provides inaccurate measurements of the
But recent studies have shown that Wi-Fi signals tend to get patient’s respiration rate, the DT model will not replicate the
disturbed by the movement of residents and other movements respiration behavior in real-time. This will lead to inaccurate
in a zone [53]. For that, in Wi-Fi sensing data, Channel recommendations with the DT utilized in research or clinical
State Information (CSI) [54] applies to a range of applications trials.
such as sedentary behavior analysis [55], gesture recogni- In the context of DT technology in healthcare, unobtrusive
tion [56], vital sign detection [57], human detection [58], sensing through Wi-Fi CSI has a vital role in building an
occupancy detection [59], crowd counting [60], and human accurate DT model that reflects the current state of the patient.
activity recognition [61], etc. Compared to contact-based The benefits of data sensing in DT include predictive analysis,
sensing methods, Wi-Fi signals have multiple excellent equipment maintenance, optimizing healthcare, accurate
attributes and advantages over wearable devices. real-time monitoring, and improved decision-making. Unob-
1) Only information about the human vital signs, position, trusive Wi-Fi sensing with outstanding characteristics gives
etc. can be obtained from the wireless signals. Wi-Fi it a broad range of potential applications in DT and the
signal protects the user’s privacy as it cannot acquire possibility to be the cornerstone of the following generation
details about appearance, clothes, etc. of sensing technology.
2) Wireless-based sensing technology does not require
specific gears to be worn by the user. It still achieves IV. PRELIMINARIES
accurate sensing and is very appropriate for long-term A. CHANNEL STATE INFORMATION (CSI)
implementation. Numerous wireless network standards e.g., LTE, Wi-Fi,
3) Wireless signals are propagated in a zone and do and WiMAX embrace Orthogonal Frequency-Division Mul-
not depend upon LOS. Diffraction and reflection of tiplexing (OFDM) for communication [62]. OFDM is the

VOLUME 11, 2023 103557

 S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

suitable choice for achieving high data rates and frequency- that are orthogonal to each other [65]. PCA can be utilized in
selective channels. The total spectrum is split into nar- multiple applications e.g., data compression, noise reduction,
rowband orthogonal subcarriers in OFDM. The channel biological signal processing [66], and image and pattern
fading caused by large delay spreads is lessened by the recognition [67], [68], [69]. Some of the PCA advantages
utilization of Inverse Fast Fourier Transform (IFFT) for are:
data transmission on the subcarriers. The complexity of
the Fast Fourier Transform (FFT) at the receiver can be 1) Lower memory requirements.
reduced by the utilization of cyclic redundancy. Channel 2) Orthogonality decreases the absence of repetition in the
properties of the communication link between the transmitter data by utilizing orthogonal components [70], [71].
and receiver are represented by the CSI. It can be freely 3) Noise reduction as the maximum variation basis is
acquired from the signals of COTS Wi-Fi devices equipped considered. In the background, the small variation is
with a Network Interference Card (NIC). The CSI signals are automatically removed [70].
the superpositions of the signals from different propagation 4) Reducing complexity as data is grouped by the
paths. The outcome of analyzing the CSI is that the PCA [70], [71].
information of the surrounding, encoded in the Wi-Fi signals, 5) PCA does not require complex computations.
can be acquired. The CSI has been measured at different However, in the PCA, until the training data provides
subcarrier frequencies due to the OFDM. Multiple antennas specific information, the simplest invariance could not
are authorized to communicate through multiple data streams be captured [72]. Also, it is difficult to compute the
in a Multiple-Input Multiple-Output (MIMO) system. The covariance matrix accurately in PCA [70]. Khanh et al. [73]
propagation of a Wi-Fi signal can be affected by the utilized Ultra-Wide Band (UWB) impulse radar to sense
micro-movement of the subject. The kind of micro-movement the vital signals of the heart rate of the patient. The data
under observation in this research is respiration. This can thus is subjected to PCA for reducing data complexity and
be detected by CSI. Bao et al. [63] provide a discussion on acquiring the projection on the primary PC. Such a projec-
the phase and amplitude changes toward detection of such tion facilitates substantially enhancing the Signal-to-Noise
micro respiration movements. Mathematically, the signal at (SNR) in comparison to other techniques of complex signal
the receiver can be stated as a function of the transmitted decomposition and Direct Fast Fourier Transform (DFFT).
signal [64]. PhotoPlethysmoGraphy (PPG) signals can be utilized to
yi = Hi xi + ni (1) obtain cardiorespiratory signals i.e., respiration rate and heart
rate which will lessen the number of sensors placed on the
The transmitted signal is represented by xi = R(NT ) and the patient’s body for data collection [74]. The authors proposed
received signal is represented by yi = R(NR ) . The subcarrier Ensemble Empirical Mode Decomposition with Principal
index is represented by i whereas, ni is the noise vector. Component Analysis (EEMD-PCA) for estimating heart rate
NT represents the number of the transmitter antenna. NR are and respiration rate concurrently from the PPG datasets of
the number of receiver antenna. Hi ∈ C(NR xNT ) represents the MIMIC (Physionet ATM data bank) and Capnobase database
complex CSI matrix. with the short length of 30 seconds. The proposed algorithm
provided better results compared to existing methods of
Hi12 . . . Hi1NT
 11 
Hi Correntropy Spectral Density (CSD) [75], Power Spectral
 H 21 H 22 . . . H 2NT 
Himn =  i i
 ... ... ... ... 
i  (2) Density (PSD) [75], Smart Fusion [76], and Empirical Mode
Decomposition (EMD) [77]. Another example of recent
HiNR 1 HiNR 2 . . . HiNR NT work of respiration rate estimation based on Wi-Fi frame
For the communication link, between the n − th transmitter capture [78]. The authors worked on Beamforming Feedback
antenna and m − th receiver antenna, Himn is the channel Matrices (BFMs) contained in the capture frames. BFM is a
frequency response. rotated matrix of CSI. PCA was utilized to separate the data
from the chest movements of the subjects in the BFMs. The
Himn = |Himn | + exp(j̸ Himn ) (3) resulting data were subjected to Discrete Fourier Transform
(DFT) to estimate respiration rates in the frequency domain.
̸ Himn and |Himn | correspond to the phase and amplitude
It is stated by the authors that the frame-capture-based
response.
respiration estimation has a lower estimation error than
3.5 breaths/minute.
B. PRINCIPAL COMPONENT ANALYSIS (PCA)
The complexity of high-dimensional sensor data is simplified
while preserving trends and patterns by using PCA. PCA 1) MATHEMATICAL MODEL
retains information as much as possible from the original Details of mathematical model of PCA is discussed in [79]
signal while deriving the data signal into a fewer number of and [80] but the following mathematical model provides a
decorrelated linear combinations. The general idea of PCA generic overview of how PCA works in our dataset. The input
is to find the Principal Components (PC) of the data signals matrix X of time series respiration sensor data is represented

103558 VOLUME 11, 2023

S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

in Equation 4: D. SINGLE VARIABLE DECOMPOSITION (SVD)

SVD is a characteristic decomposition and massive data
X( t) = [x1 (t) + x2 (t) + x3 (t) . . . .. + xn (t)] (4) compression technique utilized in the fields of data analysis,
The respiration signals are obtained from raw sensor data by ML, and signal processing. SVD disassembles the sensor
using the PCA through the covariance matrix. The covariance data matrix into three components: U (left singular vector),
matrix is defined as Equation 5 or Equation 6: 6, and V (right singular vector). The U and V are the
orthogonal matrices, whereas, 6 is the singular values
C = conv(X ) (5) of the original matrix represented in a diagonal matrix.
Liu et al. [87] provided a detailed mathematical model of
Or,
SVD for signal processing. According to [88], a simpler
n version can be represented in equation 10:
1X
C= Xj XjT (6)
n
j=1 A = U ∗ 6 ∗ VT (10)
This provides an m ∗ m matrix. The answer to Equation 7
where A is an MxN matrix, U is an MxM, 6 is an MxN
provides the eigenvector αj and eigenvalues λj .
matrix, and V T is an NxN matrix.
Cj αj = λj αj j = 1, 2, 3, 4, . . . .n (7) In our research, the purpose of the SVD method is to
achieve denoising, dimensionality reduction, and character-
The PCA acquired are using Equation 8: istic decomposition to estimate the patient’s BPM for the
creation of a respiration DT model.
zj = αj X (8)

They are assembled in order of magnitude of eigenvalues. V. RESPIRATION DATA COLLECTION WITH ESP32
The eigenvalues correspond to the fraction of the total ESP32 Microcontroller units containing Wi-Fi modules using
variance. So, the PC with the most variance is considered the the esp32-CSI tool have been utilized to collect datasets
signal with the most data. i.e., Wi-Fi sending data using CSI for respiration rate
measurements, in a standard 3m x 3m room [89]. With respect
C. EMPIRICAL MODE DECOMPOSITION (EMD) of Fig. 3, this is the initial stage where patient respiration
EMD is a data-driven signal processing technique that data collection is performed with ESP32 sensor. The patient
obtains the oscillatory tones embedded in a signal. No prior chest expansion and contraction during the breathing process
knowledge of the data is necessary, thus making it fully perturbs the Wi-Fi’s propagation environment. The chest
data-driven. EMD or Hilbert-Huang Transform decomposes small perturbations are detected as small changes in the Wi-Fi
multi-component, nonlinear signals into oscillatory modes CSI data [90]. The Wi-Fi CSI-based non-invasive sensor
of the signal called Intrinsic Mode Functions (IMFs) and datasets are collected from the experiments to be utilized for
a monotonic function known as residue [81], [82]. The creating a patient DT model. The experiment was performed
shifting process, an iterative process, is utilized to estimate with the Wi-Fi transmitter and receiver in a 3m x 3m room.
the IMFs [83]. The subject is seated around 0.9m perpendicularly from
the midpoint of the transmitter-to-receiver direct line-of-
M
X sight distance away from the Wi-Fi transmitter and receiver,
X (t) = Ci (t) + rM (t) (9) which were installed at a height of 0.85m from the floor
i=1
level. The subject’s age is 26; 1.56m tall and weighs 47kg.
X (t) represents the discrete-time samples. Where in Sensor data is collected over 2 minutes with 120 packets
Ci (t)M
i=1 , M is the number of obtained IMFs with the same per second sampling rate of the transmitter. The sensor
length as X (t). Residue function is presented by rM (t). data is expected to be 14400 packets according to the
EMD has been used for some time in multiple engineering sampling rate, but it is expected in the experimental setup
areas, including denoising ECG signals, classification of to have losses. The sensor data has 52 subcarriers as the
EEG, and various biomedical engineering fields. The output sensor with Wi-Fi CSI characteristics utilizes the concept of
of the EMD is the variable quantity of components that have OFDM. The beats are set to a range of 12BPM to 28BPM
the same length as that of the original signal. Researchers to depict the respiration rate of older adults. The range
have been using EMD for the estimation of patient vital signs. of normal BPM for older adults is 12 to 16 and beyond
Pinheiro et al. [84] utilized EMD and PCA to acquire patient’s which is an indication of health problems. MATLAB has
cardiovascular vitals from raw data of a chair-based system. been used to acquire Wi-Fi CSI signal from the .csv file,
The performance of EMD for respiration vitals estimation signal processing, and utilizing ML algorithms for BPM
in [85] and [86]. rate calculation, classification of normal and abnormal BPM,
In our research, we aim to utilize EMD as a dimensionality and predicting BPM values. Fig. 4 represents the Wi-Fi CSI
technique methodology with IMFs being utilized to estimate respiration rate data collection of a patient with an ESP32
the patient’s BPM for the creation of a respiration DT model. sensor.

VOLUME 11, 2023 103559

 S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

TABLE 1. Respiration rate of multiple IIR filters.

FIGURE 4. Wi-Fi CSI respiration rate data collection in a room.

VI. RESPIRATION DT MODEL PRE-PROCESSING

A. METHODOLOGY AND FILTERS COMPARISON
Filtering is the central part of extracting the respiration signal
from a strong noisy signal. It is attained by utilizing capable
bandpass digital IIR filters. In [33] and [91], the authors Table. 1 provides a comparison of multiple IIR filters and
provided a comparison of IIR and FIR filters for baseline wavelet decomposition for the calculated BPM after pre-
noise removal in the ECG signals. Due to the computational processing. The resulting data from each of the methodology
complexity of FIR, higher memory requirements, and phase is passed through the PCA for respiration rate estimation.
delay due to higher-order FIR, IIR is considered a better The elliptic filter provides the least number of errors in
option. Fig. 5. provides an overview of the methodology calculating the BPM concerning the inter-observer variability
adopted for pre-processing raw Wi-Fi CSI data for estimating threshold of +−5 BPM [92], [93], [94]. The elliptical filter
patient BPM. provides a relatively better estimation of the BPM with 87.5%
accurate estimation as compared to other bandpass filters
such as Butterworth (BF), Chebyshev type 1 Filter (CH1),
Chebyshev type 2 Filter (CH2), and wavelet Decomposition
(62.5%, 75%, 68.75%, and 75% respectively). Additionally,
for normal BPM (12 to 16), elliptical filter presents no
inaccurate calculation of the BPM based on the inter-observer
variability threshold.

B. ELLIPTIC FILTER PARAMETRIC ANALYSIS

Table. 2 represents the values of metrics such as SNR,
Peak Signal-to-Noise Ratio (PSNR), and Shannon entropy
for all the experimental and pre-processed BPM signals.
This provides a performance overview of the pre-processing
methodology implemented to acquire the subject respiration
information from raw sensor data. The SNR values are shown
FIGURE 5. Block diagram of pre-processing raw respiration Wi-Fi CSI data.
for both the raw and the pre-processed signal. It indicates an
increase in signal strength with the pre-processing method-
In this research, we aim to utilize bandpass digital IIR ology implemented. PSNR is the ratio of the maximum
filters for acquiring the appropriate frequency range (0.1Hz possible value of a signal and the noise level. PSNR is
to 0.6Hz) for respiration rate from the noisy Wi-Fi CSI interpreted as the higher the value, the better the quality
sensor data. A comparison of multiple IIR filters and of the signal. According to Table. 2, the PSNR provides
wavelet decomposition method is additionally carried out. positive values for all the BPM showing a positive signal
The following are compared in terms of acquiring accurate strength over the noise. The Shannon entropy is considered
BPM from raw sensor data: a natural choice for quantifying the complexity of a signal.
1) Butterworth Filter (BF) Higher entropy represents higher uncertainty and a more
2) Chebyshev type 1 Filter (CT1F) unpredictable signal. Shannon entropy is considered for the
3) Chebyshev type 2 Filter (CT2F) raw and pre-processed signal. According to Table 2, for the
4) Elliptic Filter (EF) majority of the signals, there is an increase in entropy value
5) Wavelet Decomposition after pre-processing is applied to the raw signal. A higher

103560 VOLUME 11, 2023

S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

TABLE 2. Signal parametric analysis before and after pre-processing with 5) A scree plot has been used to visualize all the principal
elliptic filter.
components of the filtered CSI signal. This will help in
deciding which PC to retain and which to let go of. The
PCs of the data signal will have variable variance. PC1
will have the highest variance, PC2 with the second
highest variance, and so on. No definite rule is present
for deciding when to stop including the variables but
in [95], two recommendations are provided. The first
recommendation is to select stop selecting the variables
when an elbow in the scree plot appears - the plot
begins to flatten out. The second recommendation is
to stop selecting the variables when the variable falls
below 1. Yu et al. [96] provided a discussion in terms
of the bandpass filter and PCA. With the bandpass filter
removing the noise, this allows the PCA to maximize
the amplitude of the signal rather than the noise. The
authors calculated the scree plot and selected the first
PC component since it has the highest variance. Based
on the discussion of [95] and [96] and according to the
entropy can be an indication of pre-processing adding noise scree plot of Fig. 6(f)., the first PC will be selected to
or complexity to raw sensor data. However, the BPM values calculate the BPM of the patient.
acquired for the majority of the sensor data are within the 6) The first PC is represented in Fig. 6(g). The final stage
inter-observer variability threshold of +−5. is to implement the Power Spectral Density (PSD).
It is the method of characterizing the distribution
of the signal frequency components in an easier
C. STEP-BY-STEP PERFORMANCE ANALYSIS
visual interpretation than a complex Discrete Fourier
Fig. 6 provides the results throughout the multiple stages of Transform (DFT) [95].
pre-processing presented in Fig. 5. Following is the step-by- 7) Fig. 6(h) represents the PSD graph has a frequency,
step detail elaboration of the pre-processing output: with the maximum power, 0.205 is multiplied by 60 to
1) Fig. 6(a) is the raw 12 BPM Wi-Fi CSI sensor data acquire 12.3 BPM. According to [92], [93], and [94],
considered as an example of pre-processing techniques. the clinical limits can be defined as +/− 5 BPM,
The DC offset is removed from the sensor data set by or the inter-observer variability in the clinical setting
subtracting the mean from it, Fig. 6(b). can result in a BPM difference between 2 to 5 BPM.
2) The sensor data is then subjected to a smoothing
process i.e., the moving average [32] in Fig. 6(c). D. COMPARISON OF MULTIPLE TECHNIQUES AND ResDT
Smoothing is implemented to discover important MODEL
patterns in data while removing the unnecessary In our research, we aim to perform a comparison to analyze
information (i.e., noise). the suitability of multiple techniques such as EMD, PCA,
3) The smoothed data is passed through the bandpass EMD-PCA, and SVD in the development of the ResDT
filter (frequency range 0.1Hz to 0.6Hz [73], [76], [77]) model. The elliptic filter will be used as a bandpass filter
shown in Fig. 6(d). A bandpass filter allows the signal in comparison to EMD, PCA, EMD-PCA, and SVD. EMD-
in a specific band of frequencies, called a passband PCA is a modification to the initial EMD method. This
but blocks the components with frequencies above and offers a harmonious approach toward denoising and finding
below the band. This is a contrast to the high pass out prominent features in raw sensor data. EMD applied
filter and low pass filter, which allows the components as a pre-processing step, generates IMFs. These IMFs are
above and below the specific frequencies, respectively. further analyzed by PCA to identify the most relevant IMFs
An ideal bandpass filter will have a completely flat and extract their corresponding features. The combination
passband with all the frequencies passing to the output of these techniques allows for extracting salient features,
without attenuation or amplification and complete a robust framework for denoising, and a compact and
attenuation for frequencies outside the passband. But meaningful representation of data.
achieving an ideal bandpass filter is not possible in the Table 3 provides a comparison between EMD, PCA, EMD-
real world. PCA, and SVD in terms of estimation of the patient BPM.
4) As for dimensionality reduction, by using PCA, the According to the interobserver variability of +−5BPM, PCA
processed data is projected in the form of principal provides the least errors in BPM estimation as compared
components. The total number of principal components to other methodologies. Incorrect BPM estimations are
is 52 presented in Fig. 6(e). presented as bold in Table 3. PCA is the suitable option

VOLUME 11, 2023 103561

 S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

FIGURE 6. Pre-processing of raw respiration Wi-Fi CSI data (a) Raw respiration Wi-Fi CSI data (b) DC-off set
removed from respiration Wi-Fi CSI data (c) respiration Wi-Fi CSI data after moving average filter (d)
Respiration Wi-Fi CSI data after Elliptical bandpass filter (e) All principal components after PCA (f) Scree
plot (g) First Principal Component (PC1) (h) PSD of First Principal Component.

to be utilized as a pre-processing methodology for the VII. MULTI-CLASS AND BINARY-CLASS SUPERVISED
respiration data acquired through ESP32 unobtrusive sensor. CLASSIFICATION FOR ResDT MODEL
The resulting BPM through the PCA is presented in Fig. 7 as Various combinations of generic data processing techniques
a respiration DT model, ResDT. precede ML classification and regression processes before the
Similar to [24], the ResDT model will be presented in the respiration rate DT. Pre-processing is the first stage before
form of a estimated BPM graph. The BPMs are recorded any ML implementation in terms of learning and under-
by the subject in a linear format starting from 12BPM to standing depending on the features [97]. Some of the recent
28BPM. The respiration data is currently from a single research work involves the utilization of ML algorithms
subject to create the ResDT model. It is performed on purpose for the accurate prediction of numerous medical method-
to know that the data collected can be labeled a certain ologies with real-time data and processing capability [98].
BPM for the implementation of supervised ML. The ResDT Lee et al. [15] elaborated on a context-aware healthcare
model is represented in Fig. 7. According to the interobserver system using the DT framework. ECG-based rhythms clas-
variability of +−5BPM, the least number of experimental sifier model is created with ML to detect cardiac problems
BPM (15, 21, 27) is incorrectly estimated after applying the and diseases. In [99], authors utilized ML algorithms to
pre-processing methodologies. accurately measure the bio-signals.

103562 VOLUME 11, 2023

S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

TABLE 3. Comparison based on BPM estimation.

FIGURE 8. 10-fold cross-validation in ResDT model.

is 408 BPM values. This will help analyze the ML algorithm’s

accuracy in the ResDT model in terms of how successful they
are in the classification of a certain BPM value (abnormal or
normal) if continuous 30 secs pre-processed data is provided
as input. The data is shuffled, and the total data set is kept at
80% training and 20% test datasets respectively. The training
dataset is split into training and validation data sets. For
validation, the k-fold cross-validation method is utilized [33].
The purpose of the validation data set is to provide an estimate
of the model skill. The capability of the ML model on an
unseen data sample is estimated by utilizing the statistical
method of cross-validation. Cross-validation 10-fold is used
in our ML ResDT model. Fig. 8 provides a representation of
the 10-fold cross-validation for ML algorithms in the ResDT
model.The validation set will be considered as the first fold,
and the remaining n-1 folds (n=10 here) will be utilized for
fitting the model and then the second fold and so on.

A. CLASSIFICATION ANALYSIS
Table 4 provides the accuracies of multiple classification
algorithms on the calculated BPM, pre-processed raw Wi-Fi
CSI data, of the ResDT model. These 21 algorithms are
utilized for supervised learning classification problems.
MATLAB classification learner app allows for built-in
classifiers. Overall, these algorithms are a strong starting
FIGURE 7. ResDT of the patient’s BPM.
point for many supervised learning problems with a range
of techniques that can be adapted for many different types
In our research, the focus is to utilize numerous ML based of applications and data. Table 4 provides the binary-class
algorithms for supervised classification in the ResDT model. and multi-class classification accuracies of the 21 super-
This will help analyze the suitability of of algorithms, classi- vised classification algorithms. The ML algorithms have to
fication problems, and understanding toward a multi-patient classify the abnormal BPM(17BPM to 28BPM) and normal
hospital DT model. BPM(12BPM to 16BPM). In multi-class classification, the
A thriving DT healthcare model depends on effective ML algorithms have to classify between 13 different BPM
and precise ML algorithms to accurately perform multiple ranging from 12BPM to 28BPM. Free Tree algorithm has
processes. Pre-processing has its importance. The signal-to- the best accuracies of 96.9% and 95.8% for multi-class
noise ratio is improved by pre-processing, which directly and binary-class classification respectively in the proposed
impacts the decision-making ability of the ML model ResDT model. A comparison in terms of classification can
for classification and regression. Data heterogeneity, data be carried with [64]. Hu et al. utilized the Convolutional
redundancy, and noisy data are some of the problems with ML Neural Network (CNN) on respiration Wi-Fi CSI sensor
implemented directly on raw datasets [100]. Experimental data for classification problems to show the potential of
data is of a single subject over 2 minutes for every BPM. learning-based algorithms in unobtrusive vital sign detection.
However, to increase the size of the data, the data is split into Our proposed ResDT has multi-class and binary-class clas-
30 secs. The features selected are of, according to Fig. 6(h), sification accuracies of with 96.9% and 95.8% respectively
frequency and power value. The frequency value selected compared to [64] 96.05%. Additionally, the classification
will correspond to the highest power value in that PSD for is performed on 16 different classes as compared to [64]
those specific 30-sec data selected. The total data size for ML 6 classes.

VOLUME 11, 2023 103563

 S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

TABLE 4. ML and DL algorithms accuracies for DT model.

FIGURE 10. Confusion matrix of fine tree algorithm for binary-class

classification.

FIGURE 9. Confusion matrix of fine tree algorithm for multi-class

classification.

FIGURE 11. ROC and AUC curve for binary classification problem of
normal and abnormal BPM for Fine Tree algorithm.
The confusion matrix of the fine tree algorithm for
multi-class and binary-class classification is Fig. 9 and
Fig. 10. According to [101], The k-fold cross-validation
method is considered to have the strongest bias performance and AUC (Area Under Curve) of the classes for the binary
estimates with a small data set. The authors performed a classification problem. The ROC curve of the classes is
comparison with multiple validation methods in terms of bias different, but the AUC is the same. That identifies that the
ML performance estimation. two classes have achieved the same level of performance but
For binary classification, a metric of performance analysis in different ways.
is the Receiver Operating Characteristic (ROC). The ROC
is the relationship between the True Positive Rate (TPR) VIII. CONCLUSION
and the False Positive Rate (FPR) for the different threshold This research developed a novel DT model (ResDT) based
values applied to classify the data. Whereas the AUC (Area on Wi-Fi Carrier State Information (CSI), improved signal
Under the Curve) provides the area under this curve as a processing, and Machine Learning (ML) algorithms for
measure of the overall performance of the classifier. As the monitoring and classification (binary and multi-class) of
ROC is based on a binary classification system, the abnormal patient respiration.
BPM (17 onward) is labeled as class 0, and the normal Multiple pre-processing techniques of DC-offset removal,
BPM (12 to 16) and as class 1. Fig. 11 provides the ROC smoothing filter, multiple IIR bandpass filters along with

103564 VOLUME 11, 2023

S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

wavelet decomposition, and Principal Component Analy- [10] M. Shafto, M. Conroy, R. Doyle, E. Glaessgen, C. Kemp, J. LeMoigne,
sis (PCA) for removing noise and dimensionality reduc- and L. Wang, ‘‘Modeling, simulation, information technology & pro-
cessing roadmap,’’ Nat. Aeronaut. Space Admin., vol. 32, pp. 1–38,
tion were implemented on the patient respiration data Apr. 2012.
acquired with ESP32 Wi-Fi CSI. Multiple IIR filters and [11] K. Wen, H. Xu, M. Xu, Y. Pei, Y. Lu, H. Zheng, and Z. Li, ‘‘Digital twin-
wavelet comparison is provided in the ResDT model to driven intelligent control of natural gas flowmeter calibration station,’’
Measurement, vol. 217, Aug. 2023, Art. no. 113140.
choose the optimal denoising technique for the sensor
[12] H. Laaki, Y. Miche, and K. Tammi, ‘‘Prototyping a digital twin for real
data. The elliptic filter provides the most accurate Breaths time remote control over mobile networks: Application of remote surgery,’’
Per Minute (BPM) calculation in the ResDT model by IEEE Access, vol. 7, pp. 20325–20336, 2019.
87.5% as compared to Butterworth (62.5%), Chebyshev [13] H. Elayan, M. Aloqaily, and M. Guizani, ‘‘Digital twin for intelligent
context-aware IoT healthcare systems,’’ IEEE Internet Things J., vol. 8,
1(75%), Chebyshev 2(68.75%), and wavelet (75%). PCA no. 23, pp. 16749–16757, Dec. 2021.
is compare with Empirical Mode Decomposition (EMD), [14] J. I. Jimenez, H. Jahankhani, and S. Kendzierskyj, Health Care in the
Single Variable Decomposition (SVD), EMD-PCA in terms Cyberspace: Medical Cyber-Physical System and Digital Twin Challenges
(Digital Twin Technologies and Smart Cities). Berlin, Germany: Springer,
of dimensionality reduction of data for the estimation of Jan. 2020.
patient’s BPM. [15] I. Lee, O. Sokolsky, S. Chen, J. Hatcliff, E. Jee, B. Kim, A. King,
ML supervised learning classification algorithms accu- M. Mullen-Fortino, S. Park, A. Roederer, and K. K. Venkatasubramanian,
racies, multi-class and binary class with 10-fold cross- ‘‘Challenges and research directions in medical cyber-physical systems,’’
Proc. IEEE, vol. 100, no. 1, pp. 75–90, Jan. 2012.
validation, are provided for the proposed ResDT model. [16] N. Dey, A. S. Ashour, F. Shi, S. J. Fong, and J. M. R. S. Tavares, ‘‘Medical
This provides an overview of the algorithm’s accuracies in cyber-physical systems: A survey,’’ J. Med. Syst., vol. 42, no. 4, p. 64,
terms of multi-class and binary-class classification of BPM Mar. 2018.
[17] N. Bagaria, F. Laamarti, H. Badawi, A. Albraikan, R. M. Velazquez, and
in decision-making for healthcare professionals. Among
A. E. Saddik, Health 4.0: Digital Twins for Health and Well-Being. Berlin,
the implemented algorithms, fine tree algorithm provides Germany: Springer, 2020.
binary and multi-class classification accuracies of 96.9% [18] G. Ahmadi-Assalemi, H. Al-Khateeb, C. Maple, G. Epiphaniou,
and 95.8%. Z. A. Alhaboby, S. Alkaabi, and D. Alhaboby, ‘‘Digital twins for
precision healthcare,’’ in Cyber Defence in the Age of AI, Smart
In our research of the proposed ResDT model, the Societies and Augmented Humanity. Berlin, Germany: Springer, Apr. 2020,
performance accuracies of ML algorithms with k-fold cross- pp. 133–158.
validation are indeed high but k-fold cross-validation is [19] R. Martinez-Velazquez, R. Gamez, and A. El Saddik, ‘‘Cardio twin:
A digital twin of the human heart running on the edge,’’ in Proc. IEEE
tested as the first validation method for the ResDT model. Int. Symp. Med. Meas. Appl. (MeMeA), Jun. 2019, pp. 26–28.
There is a need to implement and test multiple validation [20] X. Chen, Y. Feng, W. Zhong, and C. Kleinstreuer, ‘‘Numerical investigation
methodologies and big data of the subject’s respiration to of the interaction, transport and deposition of multicomponent droplets
completely evaluate ML algorithms in the ResDT model. Not in a simple mouth-throat model,’’ J. Aerosol Sci., vol. 105, pp. 108–127,
Mar. 2017.
only that, with big respiration data (experimental or synthetic [21] Y. Feng, J. Zhao, X. Chen, and J. Lin, ‘‘An in Silico subject-variability
generated), multi feature analysis along with ML algorithm study of upper airway morphological influence on the airflow regime
implementation on a larger training data. in a tracheobronchial tree,’’ Bioengineering, vol. 4, no. 4, pp. 90–93,
2017.
[22] X. Chen, Y. Feng, W. Zhong, B. Sun, and F. Tao, ‘‘Numerical
REFERENCES investigation of particle deposition in a triple bifurcation airway due
[1] T. Wuest, C. Irgens, and K.-D. Thoben, ‘‘An approach to monitoring to gravitational sedimentation and inertial impaction,’’ Powder Technol.,
quality in manufacturing using supervised machine learning on product vol. 323, pp. 284–293, Jan. 2018.
state data,’’ J. Intell. Manuf., vol. 25, no. 5, pp. 1167–1180, Oct. 2014. [23] B. Björnsson, C. Borrebaeck, N. Elander, T. Gasslander, D. R. Gawel,
[2] I. Lee and K. Lee, ‘‘The Internet of Things (IoT): Applications, M. Gustafsson, R. Jörnsten, E. J. Lee, X. Li, S. Lilja, D. Martínez-Enguita,
investments, and challenges for enterprises,’’ Bus. Horizons, vol. 58, no. 4, A. Matussek, P. Sandström, S. Schäfer, M. Stenmarker, X. F. Sun,
pp. 431–440, Jul. 2015. O. Sysoev, H. Zhang, and M. Benson, ‘‘Digital twins to personalize
[3] M. Rezaei, M. A. Shirazi, and B. Karimi, ‘‘IoT-based framework for medicine,’’ Genome Med., vol. 12, no. 1, pp. 1–4, Dec. 2020.
performance measurement: A real-time supply chain decision alignment,’’ [24] Y. Liu, L. Zhang, Y. Yang, L. Zhou, L. Ren, F. Wang, R. Liu, Z. Pang, and
Ind. Manag. Data Syst., vol. 117, no. 4, pp. 688–712, May 2017. M. J. Deen, ‘‘A novel cloud-based framework for the elderly healthcare
[4] D. Tang, K. Zheng, H. Zhang, Z. Sang, Z. Zhang, C. Xu, services using digital twin,’’ IEEE Access, vol. 7, pp. 49088–49101,
J. A. Espinosa-Oviedo, G. Vargas-Solar, and J.-L. Zechinelli-Martini, 2019.
‘‘Using autonomous intelligence to build a smart shop floor,’’ Proc. CIRP, [25] N. K. Chakshu, J. Carson, I. Sazonov, and P. Nithiarasu, ‘‘A semi-active
vol. 56, pp. 354–359, Jan. 2016. human digital twin model for detecting severity of carotid stenoses from
[5] M. Karatas, L. Eriskin, M. Deveci, D. Pamucar, and H. Garg, ‘‘Big data for head vibration—A coupled computational mechanics and computer vision
healthcare Industry 4.0: applications, challenges and future perspectives,’’ method,’’ Int. J. Numer. Methods Biomed. Eng., vol. 35, no. 5, p. e3180,
Exp. Syst. Appl., vol. 200, Aug. 2022, Art. no. 116912. May 2019.
[6] C. Thuemmler and C. Bai, Health 4.0: How Virtualization and Big Data [26] M. M. Rathore, S. A. Shah, D. Shukla, E. Bentafat, and S. Bakiras,
are Revolutionizing Healthcare. Berlin, Germany: Springer, 2017. ‘‘The role of AI, machine learning, and big data in digital twinning:
[7] S. Khan, T. Arslan, and T. Ratnarajah, ‘‘Digital twin perspective of A systematic literature review, challenges, and opportunities,’’ IEEE
fourth industrial and healthcare revolution,’’ IEEE Access, vol. 10, Access, vol. 9, pp. 32030–32052, 2021.
pp. 25732–25754, 2022. [27] World Health Organization. Ageing and Health. Accessed: May 12,
[8] F. Tao, Q. Qi, L. Wang, and A. Y. C. Nee, ‘‘Digital twin in industry: State- 2023. [Online]. Available: https://www.who.int/news-room/fact-
of-the-art,’’ IEEE Trans. Ind. Informat., vol. 15, no. 4, pp. 2405–2415, sheets/detail/ageing-and-health#:~:text=Common%20conditions%20in
Apr. 2019. %20older%20age,conditions%20at%20the%20same%20time
[9] E. Negri, L. Fumagalli, and M. Macchi, ‘‘A review of the roles of [28] V. Atella, A. Piano Mortari, J. Kopinska, F. Belotti, F. Lapi, C. Cricelli,
digital twin in CPS-based production systems,’’ Proc. Manuf., vol. 11, and L. Fontana, ‘‘Trends in age-related disease burden and healthcare
pp. 939–948, Jan. 2017. utilization,’’ Aging Cell, vol. 18, no. 1, p. e12861, Feb. 2019.

VOLUME 11, 2023 103565

 S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

[29] T. M. Dall, P. D. Gallo, R. Chakrabarti, T. West, A. P. Semilla, and [49] J. D. P. Ribeiro Filho, F. J. Da Silva e Silva, L. R. Coutinho, and
M. V. Storm, ‘‘An aging population and growing disease burden will B. D. T. P. Gomes, ‘‘MHARS: A mobile system for human activity
require alarge and specialized health care workforce by 2025,’’ Health recognition and inference of health situations in ambient assisted living,’’
Affairs, vol. 32, no. 11, pp. 2013–2020, Nov. 2013. J. Appl. Comput. Res., vol. 5, no. 1, pp. 44–58, Jul. 2016.
[30] P. Kumar and K. Silambarasan, ‘‘Enhancing the performance of healthcare [50] X. Hong and C. D. Nugent, ‘‘Segmenting sensor data for activity
service in IoT and cloud using optimized techniques,’’ IETE J. Res., vol. 68, monitoring in smart environments,’’ Pers. Ubiquitous Comput., vol. 17,
no. 2, pp. 1475–1484, Aug. 2019. no. 3, pp. 545–559, Mar. 2013.
[31] I. Mistry, S. Tanwar, S. Tyagi, and N. Kumar, ‘‘Blockchain for 5G- [51] Q. Zhou, J. Xing, Q. Yang, Y. Chen, and B. Feng, ‘‘Measuring
enabled IoT for industrial automation: A systematic review, solutions, intrinsic human activity information using WiFi-based attention model,’’
and challenges,’’ Mech. Syst. Signal Process., vol. 135, Jan. 2020, Measurement, vol. 195, May 2022, Art. no. 111084.
Art. no. 106382. [52] G. Cosoli, S. Spinsante, F. Scardulla, L. D’Acquisto, and L. Scalise,
[32] S. D. Min, J. K. Kim, H. S. Shin, Y. H. Yun, C. K. Lee, and M. Lee, ‘‘Wireless ECG and cardiac monitoring systems: State of the art, available
‘‘Noncontact respiration rate measurement system using an ultrasonic commercial devices and useful electronic components,’’ Measurement,
proximity sensor,’’ IEEE Sensors J., vol. 10, no. 11, pp. 1732–1739, vol. 177, Jun. 2021, Art. no. 109243.
Nov. 2010. [53] J. Liu, G. Teng, and F. Hong, ‘‘Human activity sensing with wireless
[33] P. Jagadev and L. I. Giri, ‘‘Non-contact monitoring of human respiration signals: A survey,’’ Sensors, vol. 20, no. 4, p. 1210, Feb. 2020.
using infrared thermography and machine learning,’’ Infr. Phys. Technol., [54] D. Halperin, W. Hu, A. Sheth, and D. Wetherall, ‘‘Tool release: Gathering
vol. 104, Jan. 2020, Art. no. 103117. 802.11n traces with channel state information,’’ ACM SIGCOMM Comput.
[34] M. Chu, T. Nguyen, V. Pandey, Y. Zhou, H. N. Pham, R. Bar-Yoseph, Commun. Rev., vol. 41, no. 1, p. 53, Jan. 2011.
S. Radom-Aizik, R. Jain, D. M. Cooper, and M. Khine, ‘‘Respiration rate [55] J. Yang, H. Zou, H. Jiang, and L. Xie, ‘‘CareFi: Sedentary
and volume measurements using wearable strain sensors,’’ npj Digit. Med., behavior monitoring system via commodity WiFi infrastructures,’’
vol. 2, no. 1, p. 8, Feb. 2019. IEEE Trans. Veh. Technol., vol. 67, no. 8, pp. 7620–7629,
[35] F. Ryser, S. Hanassab, O. Lambercy, E. Werth, and R. Gassert, Aug. 2018.
‘‘Respiratory analysis during sleep using a chest-worn accelerometer: A [56] J. Yang, H. Zou, Y. Zhou, and L. Xie, ‘‘Learning gestures from WiFi:
machine learning approach,’’ Biomed. Signal Process. Control, vol. 78, A Siamese recurrent convolutional architecture,’’ IEEE Internet Things J.,
Sep. 2022, Art. no. 104014. vol. 6, no. 6, pp. 10763–10772, Dec. 2019.
[36] H. Tang, J. Gao, C. Ruan, T. Qiu, and Y. Park, ‘‘Modeling of heart [57] Y. Zeng, D. Wu, R. Gao, T. Gu, and D. Zhang, ‘‘FullBreathe:
sound morphology and analysis of the morphological variations induced Full human respiration detection exploiting complementarity of
by respiration,’’ Comput. Biol. Med., vol. 43, no. 11, pp. 1637–1644, CSI phase and amplitude of WiFi signals,’’ Proc. ACM Interact.,
Nov. 2013. Mobile, Wearable Ubiquitous Technol., vol. 2, no. 3, pp. 1–19,
[37] C. O’Brien and C. Heneghan, ‘‘A comparison of algorithms for estimation Sep. 2018.
of a respiratory signal from the surface electrocardiogram,’’ Comput. Biol. [58] H. Zou, Y. Zhou, J. Yang, W. Gu, L. Xie, and C. J. Spanos, ‘‘Wifi-
Med., vol. 37, no. 3, pp. 305–314, Mar. 2007. based human identification via convex tensor shapelet learning,’’ in Proc.
[38] E. P. Doheny, B. P. O’Callaghan, V. S. Fahed, J. Liegey, C. Goulding, 32nd AAAI Conf. Artif. Intell., New Orleans LA, USA, Apr. 2018,
S. Ryan, and M. M. Lowery, ‘‘Estimation of respiratory rate and exhale pp. 1711–1718.
duration using audio signals recorded by smartphone microphones,’’ [59] H. Zou, Y. Zhou, J. Yang, W. Gu, L. Xie, and C. Spanos, ‘‘FreeDetector:
Biomed. Signal Process. Control, vol. 80, Feb. 2023, Art. no. 104318. Device-free occupancy detection with commodity WiFi,’’ in Proc. IEEE
[39] I. F. Miller, A. D. Becker, B. T. Grenfell, and C. J. E. Metcalf, ‘‘Disease Int. Conf. Sens., Commun. Netw. (SECON Workshops), San Diego, CA,
and healthcare burden of COVID-19 in the United States,’’ Nature Med., USA, Jun. 2017, pp. 1–5.
vol. 26, no. 8, pp. 1212–1217, Aug. 2020. [60] H. Zou, Y. Zhou, J. Yang, and C. J. Spanos, ‘‘Device-free occupancy
[40] M. Chandra, K. Kumar, P. Thakur, S. Chattopadhyaya, F. Alam, and detection and crowd counting in smart buildings with WiFi-enabled IoT,’’
S. Kumar, ‘‘Digital technologies, healthcare and COVID-19: insights Energy Buildings, vol. 174, pp. 309–322, Sep. 2018.
from developing and emerging nations,’’ Health Technol., vol. 12, no. 4, [61] J. Yang, H. Zou, H. Jiang, and L. Xie, ‘‘Device-free occupant activity
pp. 1–22, Mar. 2022. sensing using WiFi-enabled IoT devices for smart homes,’’ IEEE Internet
[41] M. Bartula, T. Tigges, and J. Muehlsteff, ‘‘Camera-based system Things J., vol. 5, no. 5, pp. 3991–4002, Oct. 2018.
for contactless monitoring of respiration,’’ in Proc. 35th Annu. Int. [62] W. Li, M. J. Bocus, C. Tang, S. Vishwakarma, R. J. Piechocki,
Conf. IEEE Eng. Med. Biol. Soc. (EMBC), Osaka, Japan, Jul. 2013, K. Woodbridge, and K. Chetty, ‘‘A taxonomy of WiFi sensing: CSI vs
pp. 3–7. passive WiFi radar,’’ in Proc. IEEE Globecom Workshops (GC Wkshps,
[42] A. Jalal, Y.-H. Kim, Y.-J. Kim, S. Kamal, and D. Kim, ‘‘Robust human Dec. 2020, pp. 1–6.
activity recognition from depth video using spatiotemporal multi-fused [63] N. Bao, J. Du, C. Wu, D. Hong, J. Chen, R. Nowak, and Z. Lv, ‘‘Wi-breath:
features,’’ Pattern Recognit., vol. 61, pp. 295–308, Jan. 2017. A WiFi-based contactless and real-time respiration monitoring scheme for
[43] M. H. Siddiqi, R. Ali, M. S. Rana, E. K. Hong, E. S. Kim, and remote healthcare,’’ IEEE J. Biomed. Health Informat., vol. 27, no. 5,
S. Lee, ‘‘Video-based human activity recognition using multilevel wavelet pp. 2276–2285, May 2023.
decomposition and stepwise linear discriminant analysis,’’ Sensors, vol. 14, [64] J. Hu, J. Yang, J.-B. Ong, D. Wang, and L. Xie, ‘‘ResFi: WiFi-
no. 4, pp. 6370–6392, Apr. 2014. enabled device-free respiration detection based on deep learning,’’
[44] S. Khan, I. M. Saied, T. Ratnarajah, and T. Arslan, ‘‘Evaluation in Proc. IEEE 17th Int. Conf. Control Autom. (ICCA), Jun. 2022,
of unobtrusive microwave sensors in healthcare 4.0—Toward the pp. 510–515.
creation of digital-twin model,’’ Sensors, vol. 22, no. 21, p. 8519, [65] J. Lever, M. Krzywinski, and N. Altman, ‘‘Principal component analysis,’’
Nov. 2022. Nature Methods, vol. 14, pp. 641–642, Jun. 2017.
[45] D. Dias and J. P. S. Cunha, ‘‘Wearable health devices—Vital sign [66] F. Castells, P. Laguna, L. Sörnmo, A. Bollmann, and J. M. Roig, ‘‘Principal
monitoring, systems and technologies,’’ Sensors, vol. 18, no. 8, p. 2414, component analysis in ECG signal processing,’’ EURASIP J. Adv. Signal
Jul. 2018. Process., vol. 2007, no. 1, pp. 1–21, Dec. 2007.
[46] M. Zhang and A. A. Sawchuk, ‘‘Human daily activity recognition with [67] L. Ren, H. Wang, K. Naishadham, O. Kilic, and A. E. Fathy, ‘‘Phase-
sparse representation using wearable sensors,’’ IEEE J. Biomed. Health based methods for heart rate detection using UWB impulse Doppler radar,’’
Informat., vol. 17, no. 3, pp. 553–560, May 2013. IEEE Trans. Microw. Theory Techn., vol. 64, no. 10, pp. 3319–3331,
[47] R. Mafrur, I. G. D. Nugraha, and D. Choi, ‘‘Modeling and discovering Oct. 2016.
human behavior from smartphone sensing life-log data for identification [68] J. Meng and Y. Yang, ‘‘Symmetrical two-dimensional PCA with image
purpose,’’ Hum.-Centric Comput. Inf. Sci., vol. 5, no. 1, pp. 1–18, measures in face recognition,’’ Int. J. Adv. Robotic Syst., vol. 9, no. 6,
Dec. 2015. p. 238, Dec. 2012.
[48] J.-L. Reyes-Ortiz, L. Oneto, A. Samà, X. Parra, and D. Anguita, [69] S. Karamizadeh, S. M. Abdullah, A. A. Manaf, M. Zamani, and
‘‘Transition-aware human activity recognition using smartphones,’’ Neu- A. Hooman, ‘‘An overview of principal component analysis,’’ J. Signal Inf.
rocomputing, vol. 171, pp. 754–767, Jan. 2016. Process., vol. 4, pp. 173–175, Aug. 2013.

103566 VOLUME 11, 2023

S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

[70] P. J. Phillips, P. J. Flynn, T. Scruggs, K. W. Bowyer, J. Chang, K. Hoffman, [90] S. Mosleh, J. B. Coder, C. G. Scully, K. Forsyth, and M. O. A. Kalaa,
J. Marques, J. Min, and W. Worek, ‘‘Overview of the face recognition ‘‘Monitoring respiratory motion with Wi-Fi CSI: Characterizing per-
grand challenge,’’ in Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern formance and the BreatheSmart algorithm,’’ IEEE Access, vol. 10,
Recognit., San Diego, CA, USA, 2005, pp. 947–954. pp. 131932–131951, 2022.
[71] S. Asadi, C. D. V. S. Rao, and V. Saikrishna, ‘‘A comparative study [91] S. Rani, A. Kaur, and J. Ubhi, ‘‘Comparative study of FIR and IIR filters
of face recognition with principal component analysis and cross- for the removal of Baseline noises from ECG signal,’’ Int. J. Comput. Sci.
correlation technique,’’ Int. J. Comput. Appl., vol. 10, no. 8, pp. 17–21, Inf. Technol., vol. 2, no. 3, pp. 1105–1108, 2011.
Nov. 2010. [92] G. B. Drummond, D. Fischer, and D. K. Arvind, ‘‘Current clinical methods
[72] C. Li, Y. Diao, H. Ma, and Y. Li, ‘‘A statistical PCA method for face of measurement of respiratory rate give imprecise values,’’ ERJ Open Res.,
recognition,’’ in Proc. 2nd Int. Symp. Intell. Inf. Technol. Appl., Shanghai, vol. 6, no. 3, pp. 00023–2020, Jul. 2020.
China, Dec. 2008, pp. 376–380. [93] M. Piotrowski and M. Derén, ‘‘Towards the development of a patient
[73] L. D. Khanh and P. X. Duong, ‘‘Principal component analysis for heart rate monitoring system: Review of available solutions and assumptions for
measurement using UWB radar,’’ Int. J. FUZZY Log. Intell. Syst., vol. 20, building a functionally optimal system,’’ Polish J. Aviation Med., Bioeng.
no. 3, pp. 211–218, Sep. 2020. Psychol., vol. 26, no. 1, pp. 26–34, Mar. 2023.
[74] M. A. Motin, C. K. Karmakar, and M. Palaniswami, ‘‘Ensemble empirical [94] N. Rubio, R. A. Parker, E. M. Drost, H. Pinnock, C. J. Weir, J. Hanley,
mode decomposition with principal component analysis: A novel approach L. C. Mantoani, W. MacNee, B. McKinstry, and R. A. Rabinovich, ‘‘Home
for extracting respiratory rate and heart rate from photoplethysmographic monitoring of breathing rate in people with chronic obstructive pulmonary
signal,’’ IEEE J. Biomed. Health Informat., vol. 22, no. 3, pp. 766–774, disease: Observational study of feasibility, acceptability, and change after
May 2018. exacerbation,’’ Int. J. Chronic Obstructive Pulmonary Disease, vol. 12,
[75] A. Garde, W. Karlen, J. M. Ansermino, and G. A. Dumont, ‘‘Estimat- pp. 1221–1231, Apr. 2017.
ing respiratory and heart rates from the correntropy spectral density [95] D. J. Bartholomew, Principal Components Analysis. Amsterdam,
of the photoplethysmogram,’’ PLoS ONE, vol. 9, no. 1, pp. 1–11, The Netherlands: Elsevier, 2010.
Jan. 2014. [96] B. Yu, Y. Wang, K. Niu, Y. Zeng, T. Gu, L. Wang, C. Guan, and
[76] W. Karlen, S. Raman, J. M. Ansermino, and G. A. Dumont, ‘‘Multipa- D. Zhang, ‘‘WiFi-sleep: Sleep stage monitoring using commodity Wi-
rameter respiratory rate estimation from the photoplethysmogram,’’ IEEE Fi devices,’’ IEEE Internet Things J., vol. 8, no. 18, pp. 13900–13913,
Trans. Biomed. Eng., vol. 60, no. 7, pp. 1946–1953, Jul. 2013. Sep. 2021.
[77] A. Garde, W. Karlen, P. Dehkordi, J. Ansermino, and G. Dumont, [97] A. Holzinger, ‘‘Big data calls for machine learning,’’ Encyclopedia
‘‘Empirical mode decomposition for respiratory and heart rate estimation Biomed. Eng., vol. 3, pp. 258–264, 2019.
from the photoplethysmogram,’’ in Proc. Comput. Cardiol., Zaragoza, [98] E. J. Topol, ‘‘High-performance medicine: The convergence of human and
Spain, Sep. 2013, pp. 22–25. artificial intelligence,’’ Nature Med., vol. 25, no. 1, pp. 44–56, Jan. 2019.
[78] T. Kanda, T. Sato, H. Awano, S. Kondo, and K. Yamamoto, ‘‘Res- [99] I. Al-Zyoud, F. Laamarti, X. Ma, D. Tobon, and A. E. Saddik, ‘‘Towards
piratory rate estimation based on WiFi frame capture,’’ in Proc. a machine learning-based digital twin for non-invasive human bio-signal
IEEE 19th Annu. Consum. Commun. Netw. Conf. (CCNC), Jan. 2022, fusion,’’ Sensors, vol. 22, p. 7457, Dec. 2022.
pp. 881–884. [100] L. Zhou, S. Pan, J. Wang, and A. V. Vasilakos, ‘‘Machine learning
[79] P. Langley, E. J. Bowers, and A. Murray, ‘‘Principal component analysis on big data: Opportunities and challenges,’’ Neurocomputing, vol. 237,
as a tool for analyzing beat-to-beat changes in ECG features: Application pp. 350–361, May 2017.
to ECG-derived respiration,’’ IEEE Trans. Biomed. Eng., vol. 57, no. 4, [101] A. Vabalas, E. Gowen, E. Poliakoff, and A. J. Casson, ‘‘Machine learning
pp. 821–829, Apr. 2010. algorithm validation with a limited sample size,’’ PLoS ONE, vol. 14,
[80] D. Widjaja, C. Varon, A. Dorado, J. A. K. Suykens, and S. Van Huffel, no. 11, Nov. 2019, Art. no. e0224365.
‘‘Application of kernel principal component analysis for single-lead-
ECG-derived respiration,’’ IEEE Trans. Biomed. Eng., vol. 59, no. 4,
pp. 1169–1176, Apr. 2012.
[81] N. U. Rehman and D. P. Mandic, ‘‘Empirical mode decomposition
for trivariate signals,’’ IEEE Trans. Signal Process., vol. 58, no. 3,
SAGHEER KHAN received the bachelor’s degree
pp. 1059–1068, Mar. 2010.
in electrical, with a focus on electronics from
[82] P. Trnka and M. Hofreiter, ‘‘The empirical mode decomposition in real-
the Army Public College of Management and
time,’’ in Proc. 18th Int. Conf. Process Control. Tatranská Lomnica,
Slovakia: Slovak Univ. of Technology in Bratislava, 2011, pp. 14–17.
Sciences (APCOMS), University of Engineering
[83] S. A. Anapagamini and R. Rajavel, ‘‘Removal of artifacts in ECG using
and Technology (UET), Taxila, Pakistan, in 2016,
empirical mode decomposition,’’ in Proc. Int. Conf. Commun. Signal and the master’s degree in electrical engineering,
Process., Melmaruvathur, India, Apr. 2013, pp. 288–292. with a focus on communication and signal pro-
[84] E. Pinheiro, O. Postolache, and P. Girão, ‘‘Empirical mode decompo- cessing from the National University of Sciences
sition and principal component analysis implementation in processing and Technology (NUST), Islamabad, Pakistan,
non-invasive cardiovascular signals,’’ Measurement, vol. 45, no. 2, in 2019. He is currently pursuing the Ph.D. degree
pp. 175–181, Feb. 2012. with the Scottish Microelectronic Centre (SMC), School of Engineering, The
[85] S. Hadiyoso, E. M. Dewi, and I. Wijayanto, ‘‘Comparison of emd, vmd and University of Edinburgh. His research interests include digital twin (DT),
eemd methods in respiration wave extraction based on ppg waves,’’ in Proc. industry 4.0, 5G, and the IoT.
2nd Int. Conf. Electron. Represent. Algorithm ‘Innovation Transformation
Best Practices Global Community’, Yogyakarta, Indonesia, Dec. 2019,
pp. 12–13.
[86] A. Karagiannis, L. Loizou, and P. Constantinou, ‘‘Experimental respiratory
signal analysis based on empirical mode decomposition,’’ in Proc. 1st
AAESHA ALZAABI (Member, IEEE) received
Int. Symp. Appl. Sci. Biomed. Commun. Technol., Aalborg, Denmark,
Oct. 2008, pp. 1–5. the B.Sc. degree in electrical engineering from
[87] S. Liu, Q. Qi, H. Cheng, L. Sun, Y. Zhao, and J. Chai, ‘‘A vital signs fast the American University of Sharjah, United Arab
detection and extraction method of UWB impulse radar based on SVD,’’ Emirates, in 2017, and the M.Sc. degree in signal
Sensors, vol. 22, no. 3, p. 1177, Feb. 2022. processing and communications from The Univer-
[88] Y. Wang and L. Zhu, ‘‘Research and implementation of SVD in machine sity of Edinburgh, in 2020, where she is currently
learning,’’ in Proc. IEEE/ACIS 16th Int. Conf. Comput. Inf. Sci. (ICIS), pursuing the Ph.D. degree in engineering with the
Wuhan, China, May 2017, pp. 471–475. Institute for Integrated Micro and Nano Systems.
[89] Alzaabi, T. Arslan, and N. Polydorides, ‘‘Respiration rate measurement Her research interests include RF sensing, wireless
validity and repeatability of ubiquitous non-contact Wi-Fi sensing for older communications, embedded systems, and sensor
adults in care,’’ IEEE Dataport, Jun. 2022. fusion.

VOLUME 11, 2023 103567

 S. Khan et al.: Novel DT Model Based on WiFi CSI, Signal Processing and Machine Learning

ZAFAR IQBAL received the bachelor’s degree TUGHRUL ARSLAN (Senior Member, IEEE)
from the University of Engineering and Technol- holds the Chair of integrated electronic systems
ogy (UET), Lahore, Pakistan, and the master’s with the School of Engineering, The University of
degree in power systems and its automation from Edinburgh, Edinburgh, U.K. He is a member of
North China Electric Power University, Beijing. the Integrated Micro and Nano Systems (IMNS)
He is currently pursuing the Ph.D. degree with Institute, and leads the Embedded Mobile and
the Institute of Energy Systems, School of Engi- Wireless Sensor Systems (Ewireless) Group with
neering, The University of Edinburgh, Edinburgh. the University (ewireless.eng.ed.ac.uk). He is the
His bachelor’s degree thesis focused on design, author of more than 500 refereed articles and an
implementation and monitoring of photovoltaic inventor of more than 20 patents. His research
system. He was a Research Assistance in multi-vector mini-grid analysis interest includes developing low power radio frequency sensors for wearable
and modeling with The University of Edinburgh, from October 2022 to and portable biomedical applications. He has been a member of the IEEE
February 2023. He performed the duties of a Lecturer with the University of CAS Executive Committee on VLSI Systems and Applications, since 1999,
Management and Technology, Lahore, from October 2017 to July 2018. His and the steering and technical committees of a number of international
research interests include electrical power systems, grid-connected inverters, conferences. He is a Co-Founder of the NASA/ESA Conference on Adaptive
and machine learning. Hardware and Systems (AHS) and currently serves as a member for its
steering committee. He is an Associate Editor of the IEEE TRANSACTIONS
ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS. Previously, he was an
Associate Editor of the IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I:
REGULAR PAPERS, from 2005 to 2006, and IEEE TRANSACTIONS ON CIRCUITS
THARMALINGAM RATNARAJAH (Senior AND SYSTEMS—II: EXPRESS BRIEFS, from 2008 to 2009.
Member, IEEE) is currently with the Institute
for Digital Communications, The University of
Edinburgh, Edinburgh, U.K., as a Professor of
digital communications and signal processing.
He was the Head of the Institute for Digital
Communications, from 2016 to 2018. He has
supervised 16 Ph.D. students and 21 postdoctoral
research fellows and raised more than 11 million
USD of research funding. He was a Coordinator of
the EU Projects ADEL (3.7Me) in the area of licensed shared access for 5G
wireless networks, HARP (4.6Me) in the area of highly distributed MIMO,
and EU Future and Emerging Technologies Projects HIATUS (3.6Me) in
the area of interference alignment, and CROWN (3.4Me) in the area of
cognitive radio networks. His research interests include signal processing
and information theoretic aspects of 6G wireless networks, full-duplex radio,
mmwave communications, random matrices theory, interference alignment,
statistical and array signal processing, and quantum information theory.
He has published over 400 publications in these areas and holds four
U.S. patents. He is a fellow of the Higher Education Academy (FHEA).
He was the Technical Co-Chair of the 17th IEEE International workshop on
Signal Processing advances in Wireless Communications, Edinburgh, U.K.,
in July 2016. He was an Associate Editor of IEEE TRANSACTIONS ON SIGNAL
PROCESSING, from 2015 to 2017.

103568 VOLUME 11, 2023