Information Fusion 103 (2024) 102106

Contents lists available at ScienceDirect

Information Fusion
journal homepage: www.elsevier.com/locate/inffus

Software defined radio frequency sensing framework for Internet of Medical

Things
Najah AbuAli a ,∗, Mohammad Bilal Khan a,b , Farman Ullah a , Mohammad Hayajneh a ,
Hikmat Ullah a , Shahid Mumtaz c
a
College of Information Technology, United Arab Emirates University, P.O. Box 15551, Al Ain, Abu Dhabi, United Arab Emirates
b Department of Electrical and Computer Engineering, COMSATS University Islamabad, Attock Campus, Kamra Road, Attock, 43600, Punjab, Pakistan
c Nottingham Trent University, 50 Shakespeare Street, Nottingham, NG14FQ, UK

ARTICLE INFO ABSTRACT

Keywords: The escalating demand for biomedical systems that can precisely diagnose and manage critical diseases
Deep learning underscores the need for innovative solutions. A non-invasive and intelligent Internet of Medical Things
IoMT (IoMT) system emerges as a promising technology, potentially enabling physicians to assess patients with
Respiratory abnormalities
reduced health risks. The respiratory rate is a pivotal vital sign among the primary clinical assessments.
RF sensing
The allure of Radio Frequency (RF) sensing lies in its ability to monitor respiratory patterns without
SDR
direct contact. However, the practical implementation of such systems often necessitates supplementary
hardware to manage the extensive data and radio functionalities, leading to concerns related to cost and
feasibility. Software-Defined Radio (SDR) technology presents itself as a viable solution to these challenges.
This research introduces a comprehensive framework for the IoMT system, aiming to diagnose respiratory
abnormalities early through RF sensing and SDR technology. We employ a deep learning framework and
compare its performance with traditional machine learning models to ensure reliable and precise classification
of respiratory abnormalities. The achieved results underscore the superiority of deep learning frameworks
over conventional machine learning models in classifying respiratory anomalies. Specifically, the deep learning
framework exhibits exceptional performance in discerning the temporal dependencies and patterns inherent
in respiratory abnormalities, achieving an average accuracy exceeding 98% for each respiratory abnormality
classification.

1. Introduction characterized by intermittent cessation of breathing during sleep [5,6].

Environmental exposure to smoke, pollution, and allergens can also
Within the vast expanse of Internet of Things (IoT) applications, lead to abnormal respiratory patterns. Early detection and monitoring
the Internet of Medical Things (IoMT) emerges as a particularly trans- of these patterns are crucial not only for effective treatment but also
formative force in healthcare [1,2]. By integrating sensors, devices, for enhancing the quality of life for affected individuals. Existing tech-
and data analytics, the IoMT promises to revolutionize patient care, nologies, including pulse oximeters, spirometers, and peak flow meters,
streamline healthcare operations, and elevate the overall quality of have undeniably advanced the field [7–9]. However, they come with
medical services. Central to the success of IoMT is its capability for inherent challenges. Issues related to accuracy, patient discomfort due
information fusion, which amalgamates data from diverse sources to to invasive methods, limited mobility due to tethered devices, and the
ensure comprehensive, real-time patient monitoring [3,4]. This fusion constraints of remote monitoring are some of the pressing concerns that
of information is not just a technological advancement; it represents the current landscape faces [10]. Addressing these challenges requires
a paradigm shift in how healthcare can be delivered, monitored, and a novel approach that combines cutting-edge technology with the
optimized. Respiratory health, a critical part of overall well-being, unique requirements of respiratory health monitoring. In this research,
offers a window into the broader implications of such advancements. we present such an approach by integrating software-defined radio
Alterations in respiratory patterns can be indicative of a wide range (SDR) technology with RF sensing within the IoMT framework. This
of health issues, from mild conditions like allergies to severe disor-
integration is not merely a technological amalgamation, but represents
ders such as asthma, COPD, lung cancer, and sleep apnea, which is

∗ Corresponding author.
E-mail address: [email protected] (N. AbuAli).

https://doi.org/10.1016/j.inffus.2023.102106
Received 15 June 2023; Received in revised form 7 October 2023; Accepted 30 October 2023
Available online 7 November 2023
1566-2535/© 2023 Elsevier B.V. All rights reserved.
N. AbuAli et al. Information Fusion 103 (2024) 102106

a paradigm shift in the way respiratory health can be monitored. By of- were not limited to activity recognition. In [19], postures were detected
fering a contactless, precise, and intelligent monitoring system, we aim after spinal cord surgery, achieving a classification accuracy of 99.6%.
to overcome the limitations of existing technologies. The employment Although these studies undeniably advanced the field, a recur-
of deep learning techniques further accentuates the system’s diagnostic ring theme was the focus on binary classification problems, often
capabilities, offering a distinctive, data-driven perspective in the field distinguishing between ‘normal’and ‘abnormal’states [20]. Such an ap-
of respiratory health monitoring [11,12]. Our contributions to this field proach, although effective in controlled environments, may not be
are multifaceted. robust enough for real-world applications with multifaceted data.
The potential of SDR technology in respiratory health was explored
• At the forefront is the proposed innovative IoMT framework,
by [21], which achieved a classification accuracy of 99.4% for abnor-
which harnesses the power of RF sensing and SDR technology
mal respiratory patterns. Similarly, [22] introduced a system that could
to provide early and contactless diagnosis of respiratory abnor- detect Cheyne-Stokes respiration in patients with heart failure, achiev-
malities, including sleep apnea. This system is designed with ing a diagnostic accuracy of 97%. A significant challenge of the RF
the patient’s comfort in mind, eliminating the need for invasive detection system is the accurate monitoring of multiple persons in the
sensors or tethered devices. same environment. While [23] presented a system that could achieve
• To ensure the accuracy and reliability of the data, we employ a classification accuracy of 99.7% for a single person, the accuracy
advanced signal-processing algorithms. These algorithms refine dropped to 93.5% and 88.4% for two and three persons, respectively.
the raw data, eliminating noise and artifacts thereby enhancing This decline in accuracy underscores the complexities introduced when
the accuracy of respiratory rate measurements. multiple individuals are present in the same environment. Factors such
• The proposed approach is the implementation of a deep learn- as overlapping RF signals, interference, and unique respiratory patterns
ing framework, specifically designed for classifying respiratory of each individual can confound the system, making accurate detection
patterns. In comparative analyses, this framework consistently and classification more challenging. Despite these advances, a signif-
outperforms conventional machine learning models, highlighting icant limitation persisted: manual design of complex features to train
its effectiveness and potential to revolutionize respiratory health machine learning models [24]. This not only made model comparisons
monitoring. challenging due to the lack of standard datasets but also raised concerns
about the applicability of these models to large, real-world datasets.
The proposed research explores the combination of IoMT, SDR
technology, and deep learning, and its potential to transform health- Deep learning, with its ability to automatically extract features, has
recently emerged as a promising solution to these challenges. While
care [13,14]. By addressing gaps in the current literature and introduc-
Convolutional Neural Networks (CNNs) have been widely adopted for
ing a flexible, scalable, and efficient framework, we aim to establish
image-based medical diagnoses [25–29], their efficacy in time-series
a new benchmark for IoMT-based healthcare solutions, with a focus
data remains debated [30]. On the other hand, recurring neural net-
on respiratory health. This synergy has the potential to improve pa-
works (RNNs) have shown promise in learning sequential and temporal
tient outcomes and contribute to the advancement of the healthcare
features, as evidenced by their applications in natural language pro-
industry.
cessing (NLP), audio, and speech processing [31,32]. Hybrid models,
The paper is organized into five sections. Section 2 discusses the
combining CNNs and RNNs, have also been explored, especially for
available techniques for the intelligent diagnosis of health disorders
video data classification, demonstrating success in learning both spatial
using SDR technology and RF sensing, and deep learning for building
and temporal features [33].
a platform. In Section 3, the methods and materials are presented to
Recent studies have also explored deep learning for biomedical
develop an intelligent and contactless diagnosis of respiratory abnor-
diagnosis based on audio data. For example, [34] used a chest mounted
malities. Section 4 involves the analysis of the results and a discussion
sensor to collect audio data from human speech and coughing, which
on the performance evaluation of the classification accuracy of respira-
was then classified using a CNN-RNN hybrid model. Similarly, [35,36]
tory abnormalities using deep learning and machine learning models.
employed deep learning models to classify sleep apnea disorders and
Finally, Section 5 concludes the paper and presents future recommen-
heart sounds, respectively. In [37], proposes an innovative solution that
dations for the developed system, along with suggestions to improve it
uses long short-term memory (LSTM) networks to differentiate between
in the future.
apnea and hypopnea episodes. The research demonstrates the model’s
effectiveness in classifying episodes and estimating the respiratory
2. Related work event index. This offers a promising alternative to traditional diagnostic
methods. The deep learning model’s performance is significant and has
The significance of diagnosing respiratory abnormalities is shown valuable contributions to medical applications.
paramount, given the prevalence and potential severity of associated In summary, while significant advances have been made in the
conditions. Extensive research has been dedicated to using RF signals domain of RF signal-based sensing and SDR technology for the clas-
and SDR technology for the intelligent classification of health disorders. sification of health disorders, gaps persist. The prevalent focus on
In [15], an SDR-based system was introduced for human activity binary classification and the manual feature crafting in machine learn-
recognition (HAR). This system, marked by its flexibility and scala- ing models highlight the need for more robust, automated solutions.
bility, showcased the potential of SDR technology in sensing human Our research, as delineated in the introduction, aims to bridge these
health disorders. Furthermore, [16] demonstrated the detection of hand gaps, offering our contribution by integrating IoMT, SDR technology,
movements using RF sensing and SDR technology, while [17] expanded and deep learning for a comprehensive solution to respiratory health
the scope to detect both small-scale movements (such as breathing and monitoring.
coughing) and large-scale movements (hand movements).
The versatility of SDR technology was further highlighted by [18], 3. Materials and methods
which introduced a contactless through-the-walls (TTW) smart sensing
system to monitor human physical activity during isolation. Their ap- The IoMT framework uses RF sensing and SDR technology to detect
proach, which utilized fine-grained wireless channel state information and classify respiratory abnormalities through four processing units:
data, achieved a classification accuracy of 99.7% for various physical respiration abnormalities time series data collection, signal processing,
activities using machine learning models. However, such advances classification, and internet gateway.

2
N. AbuAli et al. Information Fusion 103 (2024) 102106

Fig. 1. Data collection using SDR technology-based RF sensing.

3.1. Data collection Table 1

Data collection information on respiratory abnormalities.

The RF sensing platform, leveraging SDR technology, facilitates Sr. No Information Quantity/Operating

non-invasive data acquisition pertaining to respiratory anomalies, as 1 USRP Devices 1

2 Antennas 2
depicted in Fig. 1. This platform combines a universal software radio
3 Frequency in GHz 4.8
peripheral device (USRP), a computer laptop, and a pair of directional 4 Distance between Antennas in Feets 4
antennas. The USRP device is responsible for delineating radio func- 5 Height of the Antennas in Feets 1.5
tionalities, which are orchestrated via a proprietary code developed 6 Sampling Rate (S/s) 200
within the LabVIEW software environment housed on the laptop. Pa- 7 Samples 12 000
8 Laptops 1
rameters such as operating frequency, sampling rate, and antenna gain
9 Subjects 25
are meticulously defined within this software. Directional antennas 10 Total Experiments 100
play a pivotal role in both the transmission and reception of elec- 11 Total subcarriers 256
tromagnetic (EM) waves, thus evaluating the wireless channel state 12 Selected subcarriers 146
13 Activity duration in seconds 60
information (WCSI) at the receiving end. WCSI provides invaluable
14 Number of activities 4
information on the wireless channel. The presence of a human within 15 Each activity records 3650
this channel invariably results in a distinctive channel frequency re- 16 Total number of records 14 600
sponse (CFR) due to macro-movements like hand and leg motions,
and micro-movements such as those of the chest and abdomen. For
the purposes of this research, the focus is predominantly on micro-
movements associated with respiratory anomalies, all of which are 3.2.1. Denoising
meticulously extracted within a controlled laboratory setting. Denoising is executed using the discrete wavelet transform (DWT).
The study encompasses data from four distinct respiratory patterns: The data received from the RF sensing system are often marred by
Eupnea (standard relaxed breathing), Bradypnea (decelerated breathing high-level impulses and extraneous signal bursts. Conventional denois-
rate), Tachypnea (accelerated breathing rate), and Sleep Apnea (a sleep ing methodologies, such as low-pass filters (LPF) and median filters
disorder characterized by intermittent cessation of breathing). The data (MF), exhibit suboptimal performance for intricate applications [38].
set is enriched by incorporating subjects from diverse demographics While, in theory, an LPF or MF should effectively filter out these
of age. Before data acquisition, each participant undergoes a compre- superfluous signals, residual noise often persists, thereby altering the
hensive orientation through a video demonstration. It is imperative to signal’s intrinsic information. Thus, direct application of these filters
note that the data acquisition platform has no adverse implications on for denoising is inadvisable. Advanced filtering paradigms, especially
human health. All experimental procedures are initiated after securing those based on wavelet transform, are adept at meticulously filter-
both verbal and written consent from the participants. A detailed ing out undesired signals from RF sensing data, particularly given
description of the real-time experimental setup for data acquisition is their prowess in discerning human respiratory patterns. When config-
available in Fig. 2. The complete details of the data acquisition are ured with optimal wavelet transform parameters, these filters effec-
tabulated in Table 1. tively eliminate residual noise, yielding a lucid respiratory pattern,
thereby surpassing conventional denoising techniques. The wavelet
3.2. Data prepossessing transform (WT) has garnered significant attention in signal processing,
with numerous researchers leveraging it to discern valuable frequency
Data preprocessing is a pivotal phase in extracting subjective in- information [39].
formation from the data set. Raw data obtained from the RF sens-
ing platform are loaded with extraneous channel information, which 3.2.2. Sub-carrier selection
can obscure the desired trends of respiratory abnormalities, poten- The orthogonal frequency division multiplexing (OFDM) is used
tially leading to erroneous diagnostic information. Hence, rigorous to accurately extract granular information during transmission and
data preprocessing is indispensable for extracting salient information reception. 256 sub-carriers are concurrently received from each CFR.
from the dataset. This study employs a comprehensive preprocessing As illustrated in Fig. 3, the amplitude variations of different sub-carriers
regimen that includes cleaning, selection of subcarriers, smoothing, and exhibit differential sensitivities to respiratory behaviors attributable to
normalization, as shown in Fig. 3. their distinct sub-carrier frequencies. Notably, nulls and DC sub-carriers

3
N. AbuAli et al. Information Fusion 103 (2024) 102106

Fig. 2. Real-time experimental setup.

Fig. 3. Signal processing of primary data.

4
N. AbuAli et al. Information Fusion 103 (2024) 102106

Fig. 4. Classification of using deep learning models.

manifest minimal amplitude variations, rendering them unsuitable for Table 2

Deep learning models parameters tuning.
respiratory data analysis, necessitating their exclusion from the dataset.
Sub-carriers with diminished sensitivity can also impede accurate res- Sr.No Parameter Tunning

piration detection and are thus prudently 110 subcarriers were re- 1 Numbers of Layers 1
2 Numbers of Hidden Units 16, 32, 64
moved. This study adopts the higher-variance scoring method to select
3 Learning Rate 0.2
subcarriers, given their heightened sensitivity to human respiration. 4 Dropout Rate 0.2
5 Batch size 32, 64, 128
3.2.3. Smoothing 6 Dense Layer Activation Function Softmax
For signal refinement, a moving average filter is deployed. This 7 Epochs 50
8 Optimizer Adam
finite impulse response filter is instrumental in mitigating short-term
surges or random noise perturbations, ensuring the preservation of the
right respiratory pattern. A linear moving average is applied, with a
window size of 256, to the input time series data to enhance the clarity This study evaluates the performance of both deep and conventional
of respiratory patterns. machine learning models in terms of classification accuracy for the
respiratory abnormalities dataset. The deep learning frameworks un-
3.2.4. Normalization der scrutiny include recurrent neural networks (RNN), long-short-term
The refined respiratory patterns undergo normalization using the memory (LSTM), bidirectional LSTM (Bi-LSTM), gated recurrent units
statistical profile (𝜇, 𝜎 𝑠 ) of the time series data, where 𝜇 and 𝜎𝑠 denote (GRU), and bidirectional GRU (Bi-GRU). The hyperparameter optimiza-
the mean and standard deviation, respectively. The objective is to tion process is carried out to improve model performance, with the
achieve a normalized respiratory pattern with a statistical profile of specific tuning parameters defined in Table 2.
(0, 1). Here, 𝑌 represents the output time series data of the normalized
respiratory pattern, 𝑋 represents the input respiratory pattern of length
𝑥, and 𝑥𝑖 is the 𝑖th sample of 𝑋. Normalization is pivotal for enhancing 3.3.1. RNN model
the efficacy of certain AI models, especially those reliant on distance- The Recurrent Neural Network (RNN) is a neural architecture adept
based metrics or specific data distributions. By judiciously scaling at processing sequential data by leveraging historical inputs and their
the features, the algorithm can more precisely discern inter-variable associated outputs. Its inherent feedback connections facilitate the
relationships, culminating in better classification outcomes. retention of prior inputs, enhancing its predictive capabilities for sub-
sequent data points. Given its proficiency in discerning temporal de-
3.3. Classification pendencies, the RNN model is particularly suited for time series data.
For time series data for respiratory abnormalities, the RNN model
For the classification of time series data of respiratory abnormal- assimilates antecedent respiratory patterns to categorize the present
ities using a deep learning framework, the initial step involves la- respiratory pattern. Recurrent linkage in the RNN model ensures seam-
beling the data of the processed time series of respiratory patterns less information transition from one temporal step to the next within
with the corresponding breathing abnormality, as illustrated in Fig. 4. the dataset.

5
N. AbuAli et al. Information Fusion 103 (2024) 102106

Unlike conventional neural networks, the forward and backward and ℎ𝑡−1 through a sigmoid activation, an input gate vector 𝑖𝑡 is derived,
propagation processes of the RNN model are distinct due to their recur- as articulated in Eq. (5).
rent nature. At each temporal step 𝑡, the RNN model receives an input ( )
𝑖𝑡 = 𝜎 𝑊𝑖 ×𝑥𝑡 + 𝑈𝑖 ×ℎ𝑡−1 + 𝑏𝑖 (5)
vector 𝑥𝑡 , representing the current sample in the respiratory pattern
time series. This input, in conjunction with the preceding hidden state Eq. (5) clarifies that 𝑊𝑖 and 𝑈𝑖 are weight matrices, and 𝑏𝑖 is the bias
ℎ𝑡−1 , undergoes a series of weight multiplications to compute ℎ𝑡 . The vector, specific to 𝑖𝑡 and 𝐶̃𝑡 candidate update vector. The output 𝑜𝑡
resultant hidden state ℎ𝑡 is contingent on the RNN architecture, as ascertains which segments of 𝐶𝑡 should be relayed as ℎ𝑡 . By processing
articulated in Eq. (1). the 𝑥𝑡 and ℎ𝑡−1 through a sigmoid activation, an output 𝑜𝑡 is derived,
( ) as depicted in Eq. (6).
ℎ𝑡 = 𝜎ℎ 𝑈ℎ × 𝑥𝑡 + 𝑉ℎ × ℎ𝑡−1 + 𝑏ℎ (1)
( )
𝑜𝑡 = 𝜎 𝑊𝑜 × 𝑥𝑡 + 𝑈𝑜 × ℎ𝑡−1 + 𝑏𝑜 (6)
In Eq. (1), matrices 𝑈ℎ and 𝑉ℎ represent weights, 𝑏ℎ signifies the bias
vector, and 𝜎ℎ denotes the activation function applied on a per-sample Eq. (6) explicates that 𝑊𝑜 and 𝑈𝑜 are weight matrices, and 𝑏𝑜 is the bias
basis. Subsequently, the hidden state ℎ𝑡 is utilized to compute the vector, specific to 𝑜𝑡 . Refreshed 𝐶𝑡 is derived by combining 𝑓𝑡 , 𝑖𝑡 , and
output 𝑜𝑡 for the current temporal step, as delineated in Eq. (2): 𝑜𝑡 , as illustrated in Eq. (7).
( )
𝑜𝑡 = 𝜎𝑜 𝑊𝑜 × ℎ𝑡 + 𝑏𝑜 (2) 𝐶𝑡 = 𝑓𝑡 ×𝐶 𝑡−1 + 𝑖𝑡 × 𝑔𝑡 (7)

Eq. (2) specifies that 𝑊𝑜 represents the weight matrix, 𝑏𝑜 is the bias The hidden state ℎ𝑡 is deduced by multiplying the output gate with
vector, and 𝜎𝑜 is the sample-wise applied activation function. This the updated 𝐶𝑡 , post the application of the 𝑡𝑎𝑛ℎ activation function, as
procedure is reiterated for each sample in the time series data, with presented in Eq. (8).
each iteration updating ℎ𝑡 and generating an 𝑜𝑡 for each 𝑡. Subsequently,
ℎ𝑡 = 𝑜𝑡 × 𝑡𝑎𝑛ℎ(𝐶𝑡 ) (8)
a loss function is formulated by evaluating the predicted 𝑜𝑡 with the
target 𝑜𝑡 for that specific 𝑡. Depending on the task at hand, various The resultant ℎ𝑡 can be harnessed for diverse applications, such as
loss functions, such as the mean squared error (MSE) for classification, prognostications or generating subsequent samples in the time series.
can be employed. Gradients of this loss, relative to the RNN model’s A loss function is formulated by evaluating the predicted output or ℎ𝑡
parameters (weights and biases), are then computed. This computation in each 𝑡 with the target output or ℎ𝑡 . Gradients of this loss, relative to
is executed independently for each 𝑡, factoring in all antecedent 𝑡 the LSTM’s parameters (weights and biases), are then computed. This
contributions. The chain rule facilitates gradient calculation. These intricate process entails gradient computation for each 𝑡 independently
gradients are then harnessed to refine the RNN model’s weight matrix and considering all subsequent 𝑡 contributions. These gradients are
and bias vector via the Adam optimizer. The overarching objective retroactively propagated, factoring in the recurrent connections and the
is to minimize the cumulative loss across all temporal steps. During gating mechanisms intrinsic to the LSTM architecture. The gradients
backpropagation, gradients for each 𝑡 are aggregated and retroactively for each temporal step are aggregated and backpropagated to refine
propagated to update the RNN model’s weight matrices and bias vec- the LSTM’s weights and biases. The LSTM’s gating mechanisms ensure
tors. It is imperative to note that traditional RNNs are susceptible that gradients traverse temporally without vanishing or exploding. The
to the vanishing gradient problem (VGP). This implies that during gates 𝑓𝑡 , 𝑖𝑡 , and 𝑜𝑡 modulate information and gradient flow, empowering
extensive backpropagation through numerous temporal steps, gradients the LSTM to discern and retain enduring dependencies in time series
can either attenuate or amplify. Advanced RNN variants like LSTM and data.
GRU were conceptualized to circumvent this limitation, which integrate
specialized gating mechanisms to capture long-standing dependencies 3.3.3. Bi-LSTM model
adeptly [40]. The Bidirectional LSTM (Bi-LSTM) is an enhancement of the LSTM
architecture, designed to assimilate information from both the an-
3.3.2. LSTM model tecedent and subsequent temporal steps. This dual-layer architecture
The Long Short-Term Memory (LSTM) model is an evolved RNN ar- comprises two LSTM layers: one processing the time series data in
chitecture methodically crafted to counteract the VGP and adeptly cap- a forward trajectory and the other in a reverse trajectory. At each
ture enduring dependencies in sequential data. By introducing intricate temporal step 𝑡, the Bi-LSTM receives an input 𝑥𝑡 , emblematic of the
current sample in the time series. This input is concurrently channeled
gating mechanisms and memory cells, the LSTM model orchestrates
into both the forward and reverse LSTM layers. The forward LSTM layer
the flow of information within the network. Analogous to other RNN
processes the input from the beginning to the conclusion of the time
architectures, at each temporal step 𝑡, the input 𝑥𝑡 is channeled into
series, while the reverse LSTM layer operates in the opposite direction.
the LSTM model, representing the current sample in the time series.
At each temporal step, the forward and reverse hidden states, ℎ𝑓 𝑡 and
Central to the LSTM model is a memory cell, tasked with preserving
ℎ𝑏𝑡 , are combined to form a unified hidden state. This combined state
and updating information temporally. The preceding cell state, 𝐶𝑡−1 ,
can be employed for diverse tasks, such as predictions or generating
combined with the current 𝑥𝑡 , yields a new candidate cell state 𝐶̃𝑡 , as
subsequent samples in the time series. A loss function is formulated by
expressed in Eq. (3).
evaluating the predicted output or ℎ𝑡 in each 𝑡 with the target output or
( )
𝐶̃𝑡 = 𝑡𝑎𝑛ℎ 𝑊𝑐 × 𝑥𝑡 + 𝑈𝑐 ×𝐶 𝑡−1 + 𝑏𝑐 (3) ℎ𝑡 . Gradients of this loss, relative to the Bi-LSTM’s parameters (weights
and biases), are then computed. This intricate process entails gradient
Eq. (3) elucidates that 𝑊𝑐 and 𝑈𝑐 are weight matrices, 𝑏𝑐 is the bias computation for each 𝑡 independently, while also considering contribu-
vector, and the activation is typically the hyperbolic tangent (𝑡𝑎𝑛ℎ) tions from both forward and reverse trajectories. These gradients are
function. The forget gate, a pivotal component, arbitrates the retention retroactively propagated, factoring in recurrent connections in both the
or discarding of information from 𝐶𝑡−1 . By processing the 𝑥𝑡 and ℎ𝑡−1 forward and reverse LSTM layers. The gradients for each temporal step
through a sigmoid activation, a forget gate vector 𝑓𝑡 is derived, as are aggregated and back-propagated to refine the Bi-LSTM’s weights
detailed in Eq. (4). and biases. By combining gradients from both trajectories at each
( ) temporal step, the Bi-LSTM ensures bidirectional information flow,
𝑓𝑡 = 𝜎 𝑊𝑓 × 𝑥𝑡 + 𝑈𝑓 × ℎ𝑡−1 + 𝑏𝑓 (4)
enhancing the learning process. This bidirectional approach empowers
Eq. (4) specifies that 𝑊𝑓 and 𝑈𝑓 are weight matrices, and 𝑏𝑓 is the bias the Bi-LSTM to discern dependencies from both antecedent and sub-
vector, with 𝜎 being the sigmoid activation function. The input gate sequent contexts, making it crucial for tasks necessitating a holistic
discerns the new information to be stored in 𝐶𝑡 . By processing the 𝑥𝑡 comprehension of the input time series data.

6
N. AbuAli et al. Information Fusion 103 (2024) 102106

3.3.4. GRU model of this loss, relative to the Bi-GRU’s parameters (weights and biases),
The Gated Recurrent Unit (GRU) is a refined variant of RNN created are then computed. This intricate process entails gradient computation
to counteract VGP and capture enduring dependencies in sequential for each 𝑡 independently, while also considering contributions from
data. By merging the forget and input gates into a single update gate, both forward and reverse trajectories. These gradients are retroactively
the GRU streamlines the LSTM architecture. At each temporal step 𝑡, propagated, factoring in recurrent connections in both the forward
the GRU receives an input 𝑥𝑡 , representative of the current sample in and reverse GRU layers. The gradients for each temporal step are
the time series. This input is channeled into the GRU network, where aggregated and back-propagated to refine the Bi-GRU’s weights and
the update gate arbitrates the retention of ℎ𝑡−1 and the incorporation biases. By combining gradients from both trajectories at each temporal
of new information. This gate processes 𝑥𝑡 and ℎ𝑡−1 through a sigmoid step, the Bi-GRU ensures bidirectional information flow, enhancing the
activation, resulting in an update gate vector 𝑧𝑡 , as shown in Eq. (9). learning process. This bidirectional approach empowers the Bi-GRU to
( ) discern dependencies from both antecedent and subsequent contexts,
𝑧𝑡 = 𝜎 𝑊𝑧 ×𝑥𝑡 + 𝑈𝑧 ×ℎ𝑡−1 + 𝑏𝑧 (9) making it crucial for tasks necessitating a holistic comprehension of
Eq. (9) clarifies that 𝑊𝑧 and 𝑈𝑧 are weight matrices, and 𝑏𝑧 is the the input time series data.
bias vector. The reset gate, another pivotal component, determines These deep learning architectures are used to classify time series
which segments of ℎ𝑡−1 should be discarded. By processing the 𝑥𝑡 and data related to respiratory abnormalities. By discerning temporal de-
ℎ𝑡−1 through a sigmoid activation, a reset gate vector 𝑟𝑡 is derived, as pendencies and identifying indicative patterns, they can detect anoma-
detailed in Eq. (10). lies. The choice of a specific model is contingent upon the data’s
( ) characteristics and the problem’s distinctions, with parameter tuning
𝑟𝑡 = 𝜎 𝑊𝑟 ×𝑥𝑡 + 𝑈𝑟 ×ℎ𝑡−1 + 𝑏𝑟 (10) further optimizing performance.
Eq. (10) specifies that 𝑊𝑟 and 𝑈𝑟 are weight matrices, and 𝑏𝑟 is the bias
3.4. Internet gateway
vector. The reset gate 𝑟𝑡 is then used to calculate a candidate hidden
state ̃
ℎ𝑡 , which represents the new information to be incorporated into
The gateway, fortified with a deep learning model for the clas-
ℎ𝑡 . This computation, which processes the 𝑥𝑡 and ℎ𝑡−1 through the
sification of respiratory abnormalities, harnesses RF detection based
hyperbolic tangent activation function 𝑡𝑎𝑛ℎ, yields ̃ ℎ𝑡 , as presented
on the SDR technology dataset, as shown in Fig. 5, collects classi-
in Eq. (11).
fied respiratory abnormality data, and performs preliminary processing
( )
̃
ℎ𝑡 = 𝑡𝑎𝑛ℎ 𝑊 (𝑟𝑡 × ℎ𝑡−1 )+𝑈 × 𝑥𝑡 + 𝑏 (11) tasks. Given the confidential nature of medical data, the gateway is
presumed to be stimulated with robust security protocols. Furthermore,
Eq. (11) explicates that 𝑊 and 𝑈 are weight matrices, and 𝑏 is the bias
it adheres to stringent healthcare data regulations, such as the Health
vector. The hidden state ℎ𝑡 is refreshed by integrating ℎ𝑡−1 and ̃
ℎ𝑡 using
Insurance Portability and Accountability Act (HIPAA) or the General
the update gate 𝑧𝑡 , as illustrated in Eq. (12).
Data Protection Regulation (GDPR) [41].
( )
ℎ𝑡 = 1 − 𝑧𝑡 × ℎ𝑡−1 + 𝑧𝑡 × ̃ ℎ𝑡 (12)
4. Results and discussion
A loss function is formulated by computing the predicted output or ℎ𝑡
at each 𝑡 with the target output or ℎ𝑡 . Gradients of this loss, relative to This section presents the results of the analysis of the respiratory ab-
the GRU’s parameters (weights and biases), are then computed. This normalities time series dataset using both deep learning and traditional
elaborate process involves the calculation of the gradient for each 𝑡 machine learning algorithms. The dataset, which encompasses four
independently, while also considering all subsequent 𝑡 contributions. distinct breathing patterns, was subjected to classification to identify
These gradients are retroactively propagated, incorporating the recur- four types of respiratory abnormalities. A comparative assessment was
rent connections essential to the GRU architecture. The gradients for conducted to evaluate the performance of deep learning models against
each temporal step are aggregated and back-propagated to refine the their machine learning counterparts. For this comparative study, the
GRU’s weights and biases. During this backpropagation, the update and machine learning models employed include Ensemble Bagged Trees
reset gates modulate the gradient flow, enabling the GRU to discern (EBT), Quadratic Support Vector Machine (QSVM), Fine 𝐾 Nearest
which segments of the hidden state to retain or refresh. The update Neighbor (FKNN), Fine Tree (FT), and Kernel Näπve Bayes (KNB).
gate modulates the information transition from the antecedent hidden The evaluation methodology incorporated a five-fold cross-validation
state to the current one, while the reset gate arbitrates which segments (5-CV) technique.
of the antecedent hidden state to discard. Consequently, the GRU Convergence curves, accuracies, and confusion matrices are em-
captures enduring dependencies while maintaining a more streamlined ployed to provide a granular understanding of the classification results.
architecture than its LSTM counterpart. The accuracy and loss convergence curves for deep learning models
are evaluated by varying the epochs, batch sizes, and hidden units.
3.3.5. Bi-GRU model In Fig. 6, the accuracy and loss convergence of both the training and
The Bidirectional GRU (Bi-GRU) is an enhancement of the GRU validation curves are presented for deep learning models. Initially,
architecture, designed to assimilate information from both antecedent accuracy is not very high, but with the increasing number of epochs, the
and subsequent temporal steps. This dual-layered architecture com- accuracy is improved for all the deep learning models. The loss conver-
prises two GRU layers: one processing the time series data in a forward gence curves are also improved for both the training and validation by
trajectory and the other in a reverse trajectory. At each temporal step increasing the number of epochs. In Fig. 7, a performance comparison
𝑡, the Bi-GRU receives an input 𝑥𝑡 , representative of the current sample of the deep learning models using various batch sizes is presented. The
in the time series. This input is concurrently channeled into both the results conclude that the accuracy of the deep learning models follows
forward and reverse GRU layers. The forward GRU layer processes the trends by varying the batch size. Fig. 8 depicts the performance
the input from the beginning to the conclusion of the time series, comparison of deep learning models using various hidden units. The
while the reverse GRU layer operates in the opposite direction. At each results conclude that increasing the hidden units increases the accuracy
temporal step, the forward and reverse hidden states are combined of the deep learning models and this trend is almost present in all the
to form a unified hidden state. This combined state can be employed deep learning models used in this research. So, we use the 64 hidden
for diverse tasks, such as predicting or generating subsequent samples units for all deep learning for the rest of our implementation.
in the time series. A loss function is formulated by evaluating the Table 3 presents the confusion matrix for deep learning models
predicted output or ℎ𝑡 in each 𝑡 with the target output or ℎ𝑡 . Gradients applied to the time series data collected from the RF sensing system.

7
N. AbuAli et al. Information Fusion 103 (2024) 102106

Fig. 5. Information communicated to healthcare services through Internet.

Table 3
Confusion matrix of respiratory abnormalities data using deep learning models.
Models Actual/Predicted Eupnea Bradypnea Tachypnea Sleep Apnea
Eupnea 3645 0 1 4
Bradypnea 18 3523 35 74
RNN
Tachypnea 0 1 3647 2
Sleep Apnea 13 48 20 3569
Eupnea 3650 0 0 0
Bradypnea 6 3593 7 44
LSTM
Tachypnea 0 0 3647 3
Sleep Apnea 6 53 5 3586
Eupnea 3649 0 0 1
Bradypnea 7 3602 10 31
Bi-LSTM
Tachypnea 0 0 3649 1
Sleep Apnea 1 52 7 3590
Eupnea 3648 1 0 1
Bradypnea 4 3590 7 49
GRU
Tachypnea 0 2 3647 1
Sleep Apnea 8 47 9 3586
Eupnea 3649 1 0 0
Bradypnea 4 3615 1 30
Bi-GRU
Tachypnea 0 4 3645 1
Sleep Apnea 2 43 10 3595

For instance, the RNN model accurately predicted 3645 Eupnea sam- were correctly predicted as Sleep Apnea and 48 misclassifications out
ples, while misclassifying 5. 3523 samples were correctly predicted of 3650 samples for Bradypnea.
as Bradypnea and 127 were misclassified. 3647 samples were cor- The GRU model correctly predicted 3648 samples as Eupnea, with
rectly predicted as Tachypnea, and 3 were misclassified. 3569 samples only two misclassifications. However, the model had 60 misclassi-
were correctly predicted as Sleep Apnea, and 81 were misclassified. fications for Bradypnea and 64 misclassifications for Sleep Apnea,
The LSTM, Bi-LSTM, GRU, and Bi-GRU models exhibited analogous out of 3650 samples. For Tachypnea, the model had 3647 correct
patterns. predictions and 3 misclassifications. The Bi-GRU model had similar
The LSTM model showed superior performance in predicting the performance in predicting Eupnea, with 3649 correct predictions and
Eupnea class, with all 3650 samples correctly predicted and without only one misclassification. However, the model had 55 misclassifica-
misclassifications. However, the model had 3 misclassifications out of tions for Sleep Apnea and 35 misclassifications for Bradypnea, and
3650 samples predicted as Tachypnea and 81 misclassifications out of 3595 and 3615 were correctly predicted, respectively. The model had
3650 samples predicted as Sleep Apnea. For Bradypnea, the model had 5 misclassifications out of 3650 samples predicted as Tachypnea.
3593 correct predictions and 57 misclassifications. Deep learning models generally showed high accuracy in predicting
The Bi-LSTM model showed similar performance in predicting Eup- the Eupnea class, with very few misclassifications. However, the models
nea, with 3649 correct predictions and only one misclassification. Sim- had varying degrees of success in predicting the other classes (Bradyp-
ilarly, only one sample was misclassified out of 3650 samples predicted nea, Tachypnea, and Sleep Apnea), with some models having more false
as Tachypnea. However, the model had 60 misclassifications, and 3590 positives or false negatives than others. These results suggest that deep

8
N. AbuAli et al. Information Fusion 103 (2024) 102106

Fig. 6. Accuracy and loss curves for all the deep learning models.

9
N. AbuAli et al. Information Fusion 103 (2024) 102106

Fig. 7. Performance comparison of deep learning models using various batch sizes.

Fig. 8. Performance comparison of deep learning models using various hidden units.

learning models may be useful in predicting respiratory disorders, but correctly predicted as Bradypnea, but there were 278 misclassified sam-
further research is needed to improve their accuracy and effectiveness. ples. The Tachypnea classification was correct for 3416 samples, but
Table 4 presents the confusion matrix for several machine-learning 234 samples were misclassified. Finally, the Sleep Apnea classification
models that were used to classify four respiratory patterns based on was correct for 3336 samples, but 314 were misclassified.
the collected time-series data from the RF sensing system. The EBT The FT model showed 3076 correct predictions for Eupnea, but 574
model correctly predicted 3476 samples as Eupnea, but 174 samples samples were misclassified. For Bradypnea, 2919 samples were cor-
were misclassified. Similarly, 3462 samples were correctly predicted as rectly predicted, but 731 samples were misclassified. There were 3066
Bradypnea, but there were 188 misclassified samples. The Tachypnea
correct predictions for Tachypnea, but 564 samples were misclassified.
classification was correct for 3514 samples, but 136 samples were
Finally, there were 2866 correct predictions for Sleep Apnea, but 784
misclassified. Finally, the Sleep Apnea classification was correct for
were misclassified.
3404 samples, but 246 were misclassified.
The KNB model showed that only 2443 samples were correctly
The Q-SVM model showed 3401 correct predictions for Eupnea, but
there were 249 misclassified samples. For Bradypnea, 3470 samples predicted as Eupnea, while 1217 samples were misclassified. Similarly,
were correctly classified, but there were 180 misclassified samples. The 2501 samples were correctly predicted as Bradypnea, but 1149 samples
Tachypnea classification was correct for 3443 samples, but there were were misclassified. There were 2141 correct predictions for Tachypnea,
207 misclassified samples. Finally, there were 3311 correct predictions but 1509 samples were misclassified. Finally, there were 2391 correct
for Sleep Apnea, but 339 were misclassified. predictions of Sleep Apnea, but 1259 samples were misclassified.
In the FK-NN model, 3332 samples were correctly predicted as Eup- The EBT, Q-SVM, FK-NN, FT, and KNB models demonstrated vary-
nea, but 318 samples were misclassified. Similarly, 3372 samples were ing degrees of classification accuracy. However, it is noteworthy that

10
N. AbuAli et al. Information Fusion 103 (2024) 102106

Table 4
Confusion matrix of respiratory abnormalities data using machine learning models.
Models Actual/Predicted Eupnea Bradypnea Tachypnea Sleep Apnea
Eupnea 3476 72 27 75
Bradypnea 74 3462 21 93
EBT
Tachypnea 69 9 3514 58
Sleep Apnea 63 160 23 3404
Eupnea 3401 72 48 129
Bradypnea 71 3470 8 101
Q-SVM
Tachypnea 76 7 3443 124
Sleep Apnea 151 139 49 3311
Eupnea 3332 120 96 102
Bradypnea 86 3372 32 160
FK-NN
Tachypnea 98 29 3416 107
Sleep Apnea 93 129 92 3336
Eupnea 3076 194 197 183
Bradypnea 425 2919 140 166
FT
Tachypnea 217 104 3066 263
Sleep Apnea 260 388 136 2866
Eupnea 2433 559 158 500
Bradypnea 371 2501 230 548
KNB
Tachypnea 582 300 2141 627
Sleep Apnea 289 695 275 2391

5. Conclusion

Table 5
This paper introduced an intelligent, contactless SDRF sensing
Performance analysis of machine and deep learning models.
framework tailored for IoMT. Empirical evaluations underscored the
Machine learning Deep learning
framework’s precision and reliability, positioning it as a valuable tool
Model Accuracy Model Accuracy
for contemporary healthcare services. Its potential to non-invasively
EBT 94.91% RNN 98.52% diagnose respiratory ailments augments safety protocols, safeguard-
Q-SVM 93.32% LSTM 99.16%
FK-NN 92.27% Bi-LSTM 99.24%
ing medical professionals from direct patient contact. The system
FT 81.72% GRU 99.21% accurately classified four distinct respiratory patterns: Eupnea, Bradyp-
KNB 64.89% Bi-GRU 99.34% nea, Tachypnea, and Sleep Apnea, leveraging advanced deep-learning
models. Comprehensive evaluations revealed that the BiGRU deep
learning model outperformed its counterparts, achieving a classification
the KNB model exhibited a much higher misclassification rate as com- accuracy of 98.52%. Furthermore, a comparative analysis between
pared to all other machine learning models. deep learning and traditional machine learning models was conducted,
Table 5 offers an evaluation of the classification accuracies between highlighting the superior performance of the former in terms of clas-
deep learning and machine learning models. The findings underscore sification accuracy. Such findings accentuate the merits of integrating
that deep learning models, including RNN, LSTM, Bi-LSTM, GRU, and deep learning algorithms within health diagnostic systems.
Bi-GRU, outperform their machine learning counterparts in terms of While the primary focus of this study was respiratory pattern clas-
accuracy, achieving metrics ranging between 98.18% and 98.52%. sification, prospective research endeavors could broaden the system’s
In contrast, the machine learning models’ accuracy spanned between diagnostic capabilities to encompass a myriad of health conditions. This
64.89% and 94.91%. expansion might necessitate the integration of multifaceted sensors or
The results derived from deep learning models underscore their ro- metrics to encapsulate a comprehensive spectrum of health indicators.
bustness in the classification of respiratory abnormalities. These models It is imperative to rigorously evaluate the system’s robustness and
consistently demonstrated high accuracy in various respiratory pat- consistency across diverse scenarios and ambiances. Such evaluations
terns, emphasizing their outperformance accuracy level. Such findings could entail system trials on larger and more heterogeneous datasets,
suggest that the intricate architectures of time series deep learning encompassing a diverse demographic and health spectrum. Further
models, including RNN, LSTM, BiLSTM, GRU, and BiGRU, are more inquiries should also delve into the pragmatic implementation of this
suitable for this classification task compared to conventional machine framework within real-time medical ecosystems, addressing potential
learning models. The higher classification accuracy of deep learning challenges and ensuring its harmonious integration within existing
models can be attributed to their intrinsic ability to autonomously healthcare infrastructures. By regarding these future recommendations,
discern and extract salient features from the raw input data. These this research can catalyze the evolution and refinement of intelli-
architectures possess the ability to internalize complex patterns and gent, non-intrusive health diagnostic frameworks, thereby fortifying the
relationships within the data, facilitating more precise predictions. growing domain of IoMT.
On the contrary, traditional machine learning models often require
manual feature engineering, potentially constraining their ability to CRediT authorship contribution statement
detect subtle variations in respiratory abnormalities.
In summation, the empirical findings of this study demonstrate the Najah AbuAli: Investigation, Conceptualization, Methodology,
competence of deep learning models in the precise classification of Writing – review & editing, Funding acquisition, Research admin-
respiratory abnormalities. These findings are of significant relevance to istration. Mohammad Bilal Khan: Conceptualization, Methodology,
healthcare practitioners and researchers in the domain of the diagnosis Data curation, Writing – original draft. Farman Ullah: Software,
and monitoring of respiratory ailments. Leveraging deep learning mod- Formal analysis, Validation, Writing – review & editing. Mohammad
els can potentially enhance the accuracy and efficiency of respiratory Hayajneh: Conceptualization, Validation, Writing – review &
abnormality detection, paving the way for enhanced patient care and editing. Hikmat Ullah: Data curation, Software. Shahid Mumtaz:
timely interventions. Writing – review & editing.

11
N. AbuAli et al. Information Fusion 103 (2024) 102106

Declaration of competing interest [20] H. Chen, X. Yuan, Z. Pei, M. Li, J. Li, Triple-classification of respiratory sounds
using optimized s-transform and deep residual networks, IEEE Access 7 (2019)
The authors declare that they have no known competing financial 32845–32852.
[21] M. Rehman, R.A. Shah, M.B. Khan, N.A. AbuAli, S.A. Shah, X. Yang, A. Alo-
interests or personal relationships that could have appeared to
mainy, M.A. Imran, Q.H. Abbasi, Rf sensing based breathing patterns detection
influence the work reported in this paper. leveraging usrp devices, Sensors 21 (11) (2021) 3855.
[22] C. Yuan, M.B. Khan, X. Yang, F.H. Shah, Q.H. Abbasi, Cheyne-Stokes respiration
Data availability perception via machine learning algorithms, Electronics 11 (6) (2022) 958.
[23] M. Rehman, N.A.A. Ali, R.A. Shah, M.B. Khan, S.A. Shah, A. Alomainy, X. Yang,
M.A. Imran, Q.H. Abbasi, Development of an intelligent real-time multiperson
Data will be made available on request.
respiratory illnesses sensing system using SDR technology, IEEE Sens. J. 22 (19)
(2022) 18858–18869.
Acknowledgments [24] R.X.A. Pramono, S. Bowyer, E. Rodriguez-Villegas, Automatic adventitious
respiratory sound analysis: A systematic review, PLoS One 12 (5) (2017)
This work was supported by Zayed Health Center at UAE University e0177926.
under Fund code G00003476. [25] E. Hosseini-Asl, G. Gimel’farb, A. El-Baz, Alzheimer’s disease diagnostics by a
deeply supervised adaptable 3D convolutional network, 2016, arXiv preprint
arXiv:1607.00556.
References [26] M.J. Van Grinsven, B. van Ginneken, C.B. Hoyng, T. Theelen, C.I. Sánchez, Fast
convolutional neural network training using selective data sampling: Application
[1] F. Alshehri, G. Muhammad, A comprehensive survey of the Internet of Things to hemorrhage detection in color fundus images, IEEE Trans. Med. Imaging 35
(IoT) and AI-based smart healthcare, IEEE Access 9 (2020) 3660–3678. (5) (2016) 1273–1284.
[2] J. Ali, R.H. Jhaveri, M. Alswailim, B.-h. Roh, ESCALB: An effective slave con- [27] Y. Song, L. Zhang, S. Chen, D. Ni, B. Lei, T. Wang, Accurate segmentation of
troller allocation-based load balancing scheme for multi-domain SDN-enabled-IoT cervical cytoplasm and nuclei based on multiscale convolutional network and
networks, J. King Saud Univ.-Comput. Inf. Sci. 35 (6) (2023) 101566. graph partitioning, IEEE Trans. Biomed. Eng. 62 (10) (2015) 2421–2433.
[3] Y. Cheng, K. Wang, H. Xu, T. Li, Q. Jin, D. Cui, Recent developments in [28] O. Oktay, W. Bai, M. Lee, R. Guerrero, K. Kamnitsas, J. Caballero, A.
sensors for wearable device applications, Anal. Bioanal. Chem. 413 (24) (2021) de Marvao, S. Cook, D. O’Regan, D. Rueckert, Multi-input cardiac image super-
6037–6057.
resolution using convolutional neural networks, in: Medical Image Computing
[4] A. Awad, M.M. Fouda, M.M. Khashaba, E.R. Mohamed, K.M. Hosny, Utilization
and Computer-Assisted Intervention-MICCAI 2016: 19th International Confer-
of mobile edge computing on the Internet of Medical Things: A survey, ICT
ence, Athens, Greece, October 17-21, 2016, Proceedings, Part III, Vol. 19,
Express (2022).
Springer, 2016, pp. 246–254.
[5] C.K. Grissom, B.E. Jones, Respiratory health benefits and risks of living at
[29] P. Kisilev, E. Sason, E. Barkan, S. Hashoul, Medical image description using multi-
moderate altitude, High Altitude Med. Biol. 19 (2) (2018) 109–115.
task-loss CNN, in: Deep Learning and Data Labeling for Medical Applications:
[6] A.V. Devereaux, H. Backer, A. Salami, C. Wright, K. Christensen, K. Rice, C.
First International Workshop, LABELS 2016, and Second International Workshop,
Jakel-Smith, M. Metzner, J.K. Bains, K. Staats, et al., Oxygen and ventilator
logistics during California’s COVID-19 surge: when oxygen becomes a scarce DLMIA 2016, Held in Conjunction with MICCAI 2016, Athens, Greece, October
resource, Disaster Med. Public Health Preparedness 17 (2023) e33. 21, 2016, Proceedings. Vol. 1, Springer, 2016, pp. 121–129.
[7] A.M. Luks, E.R. Swenson, Pulse oximetry for monitoring patients with COVID-19 [30] Y. Bengio, P. Simard, P. Frasconi, Learning long-term dependencies with gradient
at home. Potential pitfalls and practical guidance, Ann. Am. Thoracic Soc. 17 descent is difficult, IEEE Trans. Neural Netw. 5 (2) (1994) 157–166.
(9) (2020) 1040–1046. [31] H. Salehinejad, S. Sankar, J. Barfett, E. Colak, S. Valaee, Recent advances in
[8] B. Theja, K. Shobha, C. Keshavamurthy, H. Chandrashekar, Spirometry, 2023. recurrent neural networks, 2017, arXiv preprint arXiv:1801.01078.
[9] A. Anisa, T. Hamzah, M.R. Mak’ruf, Peak flow meter with measurement analysis, [32] F. Ullah, M. Bilal, S.-K. Yoon, Intelligent time-series forecasting framework for
Indonesian J. Electron. Electromed. Eng. Med. Inf. 2 (3) (2020) 107–112. non-linear dynamic workload and resource prediction in cloud, Comput. Netw.
[10] C. Qiu, F. Wu, W. Han, M.R. Yuce, A wearable bioimpedance chest patch for 225 (2023) 109653.
real-time ambulatory respiratory monitoring, IEEE Trans. Biomed. Eng. 69 (9) [33] A. Ullah, J. Ahmad, K. Muhammad, M. Sajjad, S.W. Baik, Action recognition in
(2022) 2970–2981. video sequences using deep bi-directional LSTM with CNN features, IEEE Access
[11] A. Haleem, M. Javaid, R.P. Singh, R. Suman, Applications of Artificial Intelligence 6 (2017) 1155–1166.
(AI) for cardiology during COVID-19 pandemic, Sustain. Oper. Comput. 2 (2021) [34] J. Amoh, K. Odame, Deep neural networks for identifying cough sounds, IEEE
71–78. Trans. Biomed. Circuits Syst. 10 (5) (2016) 1003–1011.
[12] T. Shaik, X. Tao, N. Higgins, L. Li, R. Gururajan, X. Zhou, U.R. Acharya, Remote [35] H. Nakano, T. Furukawa, T. Tanigawa, Tracheal sound analysis using a deep
patient monitoring using artificial intelligence: Current state, applications, and neural network to detect sleep apnea, J. Clin. Sleep Med. 15 (8) (2019)
challenges, Wiley Interdiscip. Rev. Data Min. Knowl. Discov. (2023) e1485. 1125–1133.
[13] A.K. Kumar, M. Ritam, L. Han, S. Guo, R. Chandra, Deep learning for predicting [36] H. Ryu, J. Park, H. Shin, Classification of heart sound recordings using convolu-
respiratory rate from biosignals, Comput. Biol. Med. 144 (2022) 105338. tion neural network, in: 2016 Computing in Cardiology Conference, CinC, IEEE,
[14] M. Bahrami, M. Forouzanfar, Sleep apnea detection from single-lead ECG: A 2016, pp. 1153–1156.
comprehensive analysis of machine learning and deep learning algorithms, IEEE [37] J. Drzazga, B. Cyganek, An LSTM network for apnea and hypopnea episodes
Trans. Instrum. Meas. 71 (2022) 1–11.
detection in respiratory signals, Sensors 21 (17) (2021) 5858.
[15] D.M. Molla, H. Badis, L. George, M. Berbineau, Software defined radio platforms
[38] S. Li, Z. Li, J. Zhang, H. Zhang, A denoising method of diaphragm electromyo-
for wireless technologies, IEEE Access 10 (2022) 26203–26229.
gram signals based on dual-threshold filter, J. Mech. Med. Biol. 22 (03) (2022)
[16] Y. Ge, A. Taha, S.A. Shah, K. Dashtipour, S. Zhu, J. Cooper, Q.H. Abbasi,
2240009.
M.A. Imran, Contactless WiFi sensing and monitoring for future healthcare-
[39] H.Y. Mir, O. Singh, Adaptive data analysis methods for biomedical signal pro-
emerging trends, challenges, and opportunities, IEEE Rev. Biomed. Eng. 16
cessing applications, in: AI-Enabled Smart Healthcare using Biomedical Signals,
(2022) 171–191.
IGI Global, 2022, pp. 52–71.
[17] T. Bhowmik, R. Mojumder, D. Ghosh, I. Banerjee, An evaluative review on vari-
ous tele-health systems proposed in COVID phase, in: International Conference on [40] E. Bas, E. Egrioglu, E. Kolemen, Training simple recurrent deep artificial neural
Computational Intelligence in Pattern Recognition, Springer, 2022, pp. 201–210. network for forecasting using particle swarm optimization, Granular Comput. 7
[18] M.B. Khan, A. Mustafa, M. Rehman, N.A. AbuAli, C. Yuan, X. Yang, F.H. Shah, (2) (2022) 411–420.
Q.H. Abbasi, Non-contact smart sensing of physical activities during quarantine [41] L. Wang, J.P. Near, N. Somani, P. Gao, A. Low, D. Dao, D. Song, Data capsule:
period using SDR technology, Sensors 22 (4) (2022) 1348. A new paradigm for automatic compliance with data privacy regulations, in:
[19] M.A.M. Al-hababi, M.B. Khan, F. Al-Turjman, N. Zhao, X. Yang, Non-contact Heterogeneous Data Management, Polystores, and Analytics for Healthcare: VLDB
sensing testbed for post-surgery monitoring by exploiting artificial-intelligence, 2019 Workshops, Poly and DMAH, Los Angeles, CA, USA, August 30, 2019,
Appl. Sci. 10 (14) (2020) 4886. Revised Selected Papers, Vol. 5, Springer, 2019, pp. 3–23.

12