TO-LAB model: Real time Touchless Lung Abnormality detection model using USRP based machine learning algorithm

Abstract

BACKGROUND:

Due to the increasing prevalence of respiratory diseases and the importance of early diagnosis. The need for non-invasive and touchless medical diagnostic solutions has become increasingly crucial in modern healthcare to detect lung abnormalities.

OBJECTIVE:

Existing methods for lung abnormality detection often rely on invasive and time-consuming procedures limiting their effectiveness in real-time diagnosis. This work introduces a novel Touchless Lung Abnormality (TO-LAB) detection model utilizing universal software radio peripherals (USRP) and machine learning algorithms.

METHODS:

The TO-LAB model integrates a blood pressure meter and an RGB-D depth-sensing camera to gather individual data without physical contact. Heart rate (HR) is analyzed through image conversion to IPPG signals, while blood pressure (BP) is obtained via analog conversion from the blood pressure meter. This touchless imaging setup facilitates the extraction of essential signal features crucial for respiratory pattern analysis. Advanced computer vision algorithms like Mel-frequency cepstral coefficients (MFCC) and Principal Component Analysis (PCA) process the acquired data to focus on breathing abnormalities. These features are then combined and inputted into a machine learning-based Multi-class SVM for breathing activity analysis. The Multi-class SVM categorizes breathing abnormalities as normal, shallow, or elevated based on the fused features. The efficiency of this TO-LAB model is evaluated with the simulated and real-time data.

RESULTS:

According to the findings, the proposed TO-LAB model attains the maximum accuracy of 96.15% for real time data; however, the accuracy increases to 99.54% for simulated data for the efficient classification of breathing abnormalities.

CONCLUSION:

From this analysis, our model attains better results in simulated data but it declines the accuracy while processing with real-time data. Moreover, this work has a significant medical impact since it presents a solution to the problem of gathering enough data during the epidemic to create a realistic model with a large dataset.

Keywords

Breathing abnormality touchless detection universal software radio peripherals machine learning multi-class SVM

1. Introduction

Lung disease endures to be a vital global health concern, as evidenced by current statistics and projections through 2023. Health-related respiratory issues play a critical role in overall well-being, and early detection of lung abnormalities is imperative for effective medical treatment [1]. According to the World Health Organization (WHO) [2], the global rates of morbidity and mortality are significantly affected by respiratory infections and chronic obstructive pulmonary disease (COPD). Approximately 10.8 million people will die from lung diseases worldwide by 2020 [3].

Traditionally, the diagnosis of lung disease has relied on various imaging modalities like X-rays, CT scans, and bronchoscopy, coupled with clinical assessments and pulmonary function tests [5, 6, 7]. While these techniques are valuable, they may have limitations in terms of accuracy, accessibility, and cost-effectiveness. Since, finding the breathing rate and breathing activity of the patients is crucial for COVID-19 scenarios since irregular breathing readings could signify a decline in the patient’s health [8]. Many health problems including latent injury or metabolic ailments could result in erratic respiration. Shallow, deep, or fast breathing are the sign of respiratory disorders [9].

Non-contact monitoring is one effective method of stopping the spread of contagious diseases during a pandemic [10, 11]. In recent years, advancements in medical imaging [12], FPGA [13, 14] artificial intelligence [15], and touchless technologies have opened new avenues for enhancing the accuracy and efficiency of lung abnormality detection. Nowadays, machine learning (ML) techniques have gained increasing attention for their potential to enhance the detection and diagnosis of lung disease. To construct the ML model [16], the breathing patterns of test individuals must be examined. By leveraging large datasets of medical data, ML algorithms can learn complex patterns and relationships associated with different lung conditions. In pandemics, this method of recording breathing patterns yields very little data, which makes it impossible to build a reliable detection model [17, 18]. The need for non-invasive and touchless medical diagnostic solutions has become increasingly crucial in modern healthcare to detect lung abnormalities. In this work, a novel Touchless Lung Abnormality (TO-LAB) detection model has been proposed that includes universal software radio peripherals (USRP) and machine learning algorithms. The key contribution of this research work is pointed below:

The ML based monitoring model is presented in this paper for efficient detection of breathing abnormalities of an individual in non-contact manner.

Initially, the digital-to-analog conversion is applied BP data, while HR data is captured using a camera and subsequently transformed into frames that are converted into IPPG signals.

Figure 1.

Sample image illustrating the different lung diseases [4].

The feature extraction process is employed with MFCC and PCA techniques for retrieving the most important features and fused them using multiplier for better classification.

The ML-based Multi-class SVM classifies the breathing abnormalities into normal, shallow and elevated patterns based on fused features of the input data.

The experimental results are verified and compared with different ML algorithms based on accuracy with the simulated and real-time data.

The rest of the paper is pre-arranged as follows Section-2 precises current research works in non-contact disease recognition, Section-3 comprises the comprehensive description of proposed TO-LAB model, Section-4 includes with results and discussion and Section-5 encompasses with conclusion and future work.

2. Literature review

The literature presents a number of contact-based and non-contact breath monitoring methodologies. The contactless methodologies include constant monitoring at homes and even in sleeping using camera-based imaging or radio-frequency technologies. On the other hand, Contact-based methodologies includes wearable sensors and Smart watches, etc. Further, few of them were described in this review section.

In 2020 Al-hababi et al. [19] designed contactless sensing testbed with artificial intelligence to classify post-surgery behaviors. In spinal cord surgery, patients are equipped with a USRP device to monitor their movements while lifting weights. The orthogonal frequency division multiplexing (OFDM) technology was used to retrieve wireless channel state information (WCSI). For experimental research, machine learning is used to categorize the proper and incorrect methods of lifting weights. According to the proposed testbed, the fine K-nearest neighbor algorithm (FKNN) is 99.6% accurate.

In 2021 Rahman et al. [20] had introduced a QUCoughScope framework to detect COVID-19 patients with cough and respiration rate. This project implements a web application and a deep learning-based backend server to collect coughing and breathing data from users at home and screen them for COVID-19 infection. The stacking algorithm performed better than the other CNN networks for binary categorization with 96.5% of accuracy.

In 2021 Khriji et al. [21] had developed a deep learning algorithm to analyze COVID-19 cough and breath data gathered on wearable sensors or mobile phones. This technique can distinguish between positive and negative COVID-19 cases. Initially preprocessing was used to eliminate noise from audio signals and noise from coughing and breathing. After deep LSTM was used to extract deep features, the recognition phase was completed using the audio elements recovered for classification.

In 2021, Purnomo et al. [22] suggested using frequency-modulated continuous wave (FMCW) to create a non-contact medical gadget that tracks and categorizes breathing patterns in real time. The peculiar breathing characteristics of patients with breathing disorders cannot be modeled by breathing rate. A classification model called Xtreme Gradient Boosting (XGBoost) was created to classify breathing pattern decisions. It also used Mel-frequency Cepstral Coefficient (MFCC) feature extraction.

In 2021, Rehman et al. [23] proposed a non-contact SDR platform for the recognition of breathing patterns and the categorization of COVID-19 and non-COVID-19 states. utilizing fine-grained OFDM symbols, different breathing patterns are detected utilizing differences in CSI brought on by human breathing. It has a 99.4% accuracy rate for detecting and categorizing respiratory patterns. Because this platform can be used practically in pandemic and non-pandemic conditions, it has a substantial clinical effect.

In 2021, Melek et al. [24] suggested an ML to focus on a single cough sound to identify COVID-19 patients from non-COVID-19 patients. The characteristics were collected from the cough noises using the MFCCs approach, and were then classified using seven different machine learning classifiers. The distinction between COVID-19 and non-COVID-19 patients could be made with 98.3% accuracy.

In 2021, Vrindavanam et al. [25] presented a different approach for categorizing COVID-19 cough samples. Patients with COVID-19 could quickly begin using contactless identification. The model was developed using fifteen dominant features. It used a technique for choosing features that involved ranking various scores obtained from the feature selection algorithms. It can be trained using different algorithms for samples of a forced cough and a natural cough.

In 2022, Nassif et al. [26] proposed that identifying COVID-19 could require speech signal and image pre-processing. There are many different kinds of queries using image-based, speech-based, and image and speech -based models. The patient’s breathing, voice, and cough were then categorized using both long- and short-term (LSTM) memory. Additionally, CNN models Densnet201, VGG16, Inceptionv3, Xception and ResNet50 for classifying chest X-ray images.

In 2022 Dong et al. [27] detects the respiratory rate, heart rate, muscle moment, sleep, and several other biological pointers with the aid of impulse-radio ultra-wideband radar. A reference group of healthy individuals was compared with 140 radar monitoring data from a hospital. XGBoost and logistic regression were then employed to divide the data into unhealthy and healthy groups. Compared to other single machine learning methods, the XGBoost with LR algorithm showed good precision of 92.5%

In 2022 Alkhodari and Khandoker [28] examined the viability of separating COVID-19, including asymptomatic, and healthy people using breathing sounds recorded from smartphones inside a structure of deep learning. This framework based on manually created features that were taken from the original recordings and the MFCC using CNN-BiLSTM. Additionally, it effectively identified COVID-19 individuals with a maximum sensitivity of 94.21%.

In 2023 Jamil and Roy [30] introduced efficient DL frameworks for detecting various VHDs, featuring simpler network architectures. Three frameworks were demonstrated, analyzing both 1D and 2D PCG signals. D-CNN extracts features for 2D PCG, while MFCC and LPCC were used for 1D PCG. Nature-inspired algorithms like particle swarm optimization (PSO) and genetic algorithm (GA) enable automatic feature selection from raw PCG data. They developed a vision transformer (ViT) model using self-attention on 2D PCG’s time-frequency representation, achieving an outstanding mean average accuracy of 99.90%.

Badnjevic et al. s[31] had designed an integrated software suite has been developed to aid physicians in analyzing and interpreting pulmonary function test results, enhancing asthma detection, diagnosis, and treatment. Utilizing fuzzy rules and a trained neural network, the software determines and classifies asthma status. Over one thousand report samples were utilized for training data. In testing with 289 patients previously diagnosed by physicians, the software achieved a classification accuracy of 97.22% for asthma and 98.61% for normal lung function, aligning closely with physicians’ reports.

In 2023 Singh et al. [32] developed “Tinku,” an affordable robot inspired by “WALL-E,” to assist children with special needs. It utilizes models like Yolo v3-tiny and SSD for tasks like obstacle avoidance and nonverbal communication. Developed under clinical supervision, Tinku undergoes thorough testing, including a survey on its appearance and interaction with children, particularly those with autism spectrum disorders (ASD). Tinku’s extensive features and meticulous management make it a valuable tool for teaching and supporting children with special needs in today’s world where ASD diagnoses are increasingly prevalent.

Badnjevic et al. [33] presented the outcomes of a software suite designed for asthma categorization, highlighting the significance of bronchodilator response (BDR) tests in evaluating variable airway obstruction, especially in patients with compromised lung function during visits. An intuitive integrated software solution was developed based on these aspects to assist physicians in interpreting and understanding data from pulmonary function tests, thus imprsoving the detection, diagnosis, and management of asthma. The software demonstrated a high level of agreement with physicians, accurately classifying asthma in 65 cases (90.27%).

Table 1
Tabulation analysis of the existing literature works

Authors	Methods	Advantage	Disadvantage
Al-hababi et al. (2020) [19]	Contactless sensing testbed to classify post-surgery behaviors	Tackles real-world challenges by reducing the cost.	Limited in accurately capturing subtle post-surgery behaviors without physical contact
Rahman et al. (2021) [20]	QUCoughScope framework to detect COVID-19 patients	Attains the binary categorization accuracy of 96.5%	Cough sounds may not always be reliable indicators of infection due to variability in cough patterns
Khriji et al. (2021) [21]	Deep learning-based LSTM to analyse COVID-19	Prediction of COVID-19 trends with high accuracy by capturing temporal dependencies	Overfitting due to limited and noisy data
Purnomo et al. (2021) [22]	Non-contact medical gadget that tracks the breathing patterns	Reducing discomfort and minimizing risk of infection transmission.	Limited accuracy in detecting subtle breathing irregularities
Rehman et al. (2021) [23]	SDR platform for the detection of COVID-19 based on breathing patterns	Achieves an accuracy of 99.4% for detecting and categorizing respiratory patterns	Inaccuracies due to variations in breathing patterns
Melek et al. (2021) [24]	ML algorithm to identify COVID-19 patients using cough sounds	Achieves an accuracy of 98.3%	Different cough characteristics leading to challenges in achieving consistent and accurate identification.
Vrindavanam et al. (2021) [25]	Design an approach for categorizing COVID-19 cough samples using fifteen dominant features	Trained using different algorithms for samples of a forced cough and a natural cough.	Low level of performance in the detection of COVID-19 cough samples
Nassif et al. (2022) [26]	LSTM to detect COVID-19 based on breathing, voice, and cough	LSTM beneficial for analysing temporal characteristics in COVID-19	The performance might be affected by noisy or inconsistent data,
Dong et al. (2022) [27]	XGBoost with LR algorithm	Attains high accuracy in detecting breathing activities due to its ensemble learning capabilities.	It requires significant computational resources and training time.
Alkhodari and Khandoker (2022) [28]	CNN-BiLSTM to classify COVID-19 based on breathing sounds	Achieves the maximum sensitivity of 94.21%.	Manually created features lowers the accuracy level.
Badnjević et al. (2013) [29]	Integrated software suite to precise analysis of ailments like asthma and COPD	Streamlined workflow and comprehensive data integration	Potential complexity and reliance on single software platform
Jamil and Roy (2024) [30]	Efficient DL frameworks for detecting various VHD	Enable automatic feature selection from raw PCG data	Detecting various VHD was limited interpretability of complex model outputs.
Badnjević et al. (2013) [31]	An integrated software suite	Achieves classification accuracy of 98.61% for normal lung function	This suite is potential dependency on a single vendor for updates and support.
Liu et al., (2023) [32]	RFID Technology-Based Smart Medicine Cabinet with Contactless Features	Enables easy inventory management and monitoring of medication usage.	Vulnerable to RFID signal interference or hacking.
Badnjević et al. (2023) [33]	Deep learning based artificial neural network	This model attained an accuracy rate of 82.4% in classification.	Classification performance level is low
Singh et al. (2023) [34]	“Tinku” an affordable robot inspired by WALL-E	Tinku supports children with special needs crucial in the rising prevalence of ASD diagnoses	Limited functionality and potential durability concerns compared to more expensive alternatives.
Badnjevic et al. (2013) [35]	A software suite designed for asthma categorization	Streamlined workflow and comprehensive data integration	Low level of accuracy
Zhu et al. (2022) [36]	Machine learning prediction model	The GBDT classifier achieved an accuracy of 0.931.	Performance level is low

The technical advantages and disadvantages of the existing works are summarized in Table 1. The devices employed in contact-based methodologies are often problematic, heavy, and expensive for patients. Camera-based imaging breath monitoring requires a thermal imaging camera or depth camera that can be computationally demanding and expensive [37]. Traditional diagnostic methods, often reliant on invasive procedures or extensive physical contact, may present challenges in terms of patient comfort, hygiene, and accessibility. In response to these challenges, a touchless approach was introduced that harness the power of USRP and ML have emerged to detect lung abnormality. The current application uses an MFCC and PCA feature extraction block and a recognition engine to categorize touchless infection. The feature computing algorithm is implemented in FPGAs for real-time lung abnormality categorization of breathing using a system-level design approach.

3. Proposed to-lab methodology

In this section, a novel TO-LAB model was described foe effective identification of breathing abnormalities such as normal, shallow and elevated breathing. Figure 2 shows the systematic model for our proposed machine learning-enabled respiratory abnormality detection system.

Figure 2.

Schematic flow of the proposed TO-LAB model.

3.1 Digital-to-Analog converter

Digital-to-Analog Converter (DAC) plays a vital role in converting digital representations of blood pressure patterns into analog voltage signals. In this application, digital data derived from monitoring a person’s breathing, often obtained through sensors, is processed and converted by the DAC into a continuous analog voltage signal. This reconstructed analog signal mirrors the variations in chest movement during inhalation and exhalation. The precision and accuracy of the DAC are crucial in faithfully reproducing the digital breathing data as an analog signal. This analog output can then be utilized for applications such as interfacing with external devices or systems that respond to analog signals, allowing for real-time feedback or integration into various healthcare and physiological monitoring setups.

3.2 Mel-frequency cepstral coefficients

In this section, the relevant features are extracted from the breathing patterns using MFCC to detect breathing abnormalities in the lungs. The MFCC considers breathing patterns to be frequencies. The last step in MFCC is Discrete Cosine Transform (DCT) which measures the similarity of the BP signals.

Figure 3.

Representation of MFCC for the input BP signal.

Figure 3 displays the process of feature extraction in BP signal with the MFCC technique. The MFCC approach involves five important steps as shown in Fig. 2 to extract cepstral features that is represented as (A). The BP rate for each time window was derived through the analysis of the frequency content of the extracted respiratory signals. Thus, a high frequency filter produces a wider bandwidth. For particular patterns of time, the coefficients serve as a good representation of the key elements in the signal information. The mathematical definition of MFCC is:

$\displaystyle d\left(m\right)=\textit{DCT}\left({\textit{log}\left({\left|{% \textit{FFT}\left({c\left(m\right)}\right)}\right|}\right)}\right)$ (1)

An FPGA was used to implement the MFCC calculation algorithm utilizing a system level approach with Xilinx System Generator (XSG) tool. To fully capture the temporal properties and variations of the input signal, it is first divided into time frames, which typically overlap significantly.

A high-frequency component known as the leakage effect will be produced at each end of a signal block. The initial and last points in the frame are kept continuous while the leaking impact is reduced through the employment of a Hamming window. After Hamming windowing, the signal is indicated by c(m)*x(m) if the signal in a frame is denoted by c(n), where m=0….M-1 and x(m) defines the Hamming window as,

$\displaystyle x\left({m,b}\right)=\left({1-b}\right)-\textit{bcos}\left({2\pi m% /\left({M-1}\right)}\right)$ (2)

where $0\leqslant m\leqslant M-1$ . Second, a fast Fourier transform (FFT) has been used to transform the windowed signal in the correct manner. A specific signal’s discrete Fourier transform (DFT) and its inverse was computed rapidly and effectively using the FFT. Based on the concept that any real signal is a linear blend of several sine waves at various frequencies, a discrete Fourier transform essentially does nothing more than separate out the component frequencies from a signal. The sequence of M complex numbers $y_{0},y_{1}$ , $\ldots$ $y_{M-1}$ is converted into the sequence of M complex numbers $Y_{0},Y_{1}$ , $\ldots$ $Y_{M-1}$ . In this transformation,

$\displaystyle Y_{n}=\mathop{\sum}\nolimits^{x}_{m}e^{-\frac{2\pi i}{M}};n=0,1,% \ldots M-1$ (3)

This method produces a breakdown of the frequency elements that made up the original signal. Third, the bandpass filters integrated into the Mel-frequency scale are used to pass these frequency-separated signals. The MFCC technique offers two essential pieces of information, one of which is the generalized frequency response. Cepstral coefficients for the Mel frequency are the second important piece of information. The coefficients of the band-passed frequency response are extracted from the log by DCT of each intermediate signal. The DCT transform typically produces the MFCC of the audio signal. The following equation describes a collection of cepstral coefficients ( $C_{j}$ ) produced by the DCT transform:

$\displaystyle C_{j}=\mathop{\sum}\nolimits^{n}_{i}\cos\left[{\frac{j\pi}{N}% \left({-0.5}\right)}\right]$ (4) $\displaystyle n_{i}=\log_{e}\left({Z_{i}}\right)$ (5)

$Z_{i}$ is the output signal amplitude with the $i^{\text{th}}$ Mel-scale filter and N the number of Mel-scale filters employed in the filter bank analysis. During the processing stage, the DCT coefficients are kept to form a sequence of signal vectors known as the cepstral coefficient of the Mel frequency. The extracted cepstral Mel-frequency coefficients (A) are fused in the multiplier.

3.3 Principle component analysis

The I/O protocols and hardware that are used to implement the PCA method. MicroBlaze controls direct memory access (DMA) using a prefetching-based DRAM and DDR3 SDRAM for data input. Programmable components are used in the implementation of the PCA method. Debugging is also carried out using a RS232 controller (during development). MicroBlaze and its memory, PLB bus, DMA, and RS232 controllers are fixed components, while the reconfigurable unit is a reconfigurable component, which allows us to co-design hardware and software. An algorithm is implemented entirely on this reconfigurable unit or just its most suitable components (those that can be executed on hardware). The process converting the image to IPPG signal is displayed in Fig. 4.

Figure 4.

Representation of image to IPPG signal conversion.

A MicroBlaze has the main advantage of allowing easy reconfiguration and adaptation for various inputs and outputs (SD Card, Ethernet, PCI bus, etc.). The covariance matrix was not computed using a hardware implementation. Due to the penalty caused by I/O communications, this option is inferior to computing them using software. Consequently, these variations are computed by the MicroBlaze CPU. The various computations are combined between software and hardware in this way. The concept is that the hardware will perform operations on the covariance matrix elements while the software calculates the variances. PCA method was constructed by utilizing the hardware and bus interactions, which use two FIFOs. The Jacobi module is applied to compute the frequency and amplitude of the signal in the new base are retrieved by the projection modules.

In order to detect heart rate from a video sequence, the iPPG signal is transformed from the time-domain to the frequency domain as shown in Fig. 3. The frequency and time domains of an iPPG signal captured during an 11-second user face capture using a standard laptop camera recording at 30 frames per second with a 640 $\times$ 480 resolution. The real-time processing of each frame will cause 21 frames to be dropped despite 30 frames being displayed per second. Therefore, it will take 11 seconds to process 99 frames.

The primary and mean components of the data are determined through PCA. It is frequently applied as a method of feature extraction. The technique is typically used to boost variance and capture potent feature patterns. The feature extraction from respiratory patterns improves PCA performance. A lower frequency subspace is created by the various operations on the feature matrix. then, from the lower dimension matrix, determine the covariance matrix. The covariance matrix illustrates the relative variance between signals. Principal components are regarded as being the frequency with the highest amplitude are defined as the principal coefficients features (B) from this feature extraction modules and this features also fused in the fusion module.

3.4 Fusion module

In the fusion module design, two feature vectors are dot-product using the architecture of the multiplier. The first vectors are extracted from the BP i.e., represented as $A$ from MFCC and second vectors are extracted from the HR i.e., represented as $B$ from PCA. The fusion of feature vectors is represented in Fig. 5. It includes with a row of multipliers and a tree of adders so that for two vectors $A=\left\{{a_{0},a_{1},a_{2},\ldots,a_{n}}\right\}$ and $B=\left\{{b_{0},b_{1},b_{2},\ldots,b_{n}}\right\}$ .

Figure 5.

Hardware architecture to fuse the two feature vectors.

The product element by element resulting $\left\{{a_{0}\times b_{0},a_{1}\times b_{1},\ldots,a_{n}\times b_{n}}\right\}$ . Then, each of these multiplications is fused together 2 to 2, resulting $\{a_{0}\times b_{0}+a_{1}\times b_{1},a_{2}\times b_{2}+a_{3}\times b_{3},% \ldots,a_{n-1}\times b_{n-1}+$ $a_{n}\times b_{n}\}$ at the first level. Through many stages, the addition of 2 to 2 is continued until the dot-product is reached. A pipeline structure was utilized, which is able to multiply two vectors concurrently, cycle by cycle. As a result, the vectors in the FIFO memory are multiplied by the feature-vectors acquired in the Jacobi module for further processing of fused features.

3.5 Classification

Multi-class SVM is a classification technique used to classify data into more than two classes. This approach was used to classify breathing patterns based on the extracted features. Multi-class optimization problems are usually decomposed into several binary classification problems in this method instead of solving them directly. In order to classify breathing abnormality, a three-fold random cross-validation procedure was used. In the preceding module, the multi-class SVM was used to categorize breathing abnormalities. Multi-class SVM was designed with the help of statistical learning theory and structural risk minimization. A multiclass SVM is usually composed of multiple binary SVMs. Data points belonging to a particular class are trained to distinguish them from those belonging to all other classes during training. In the case of a multiclass classifier, the output class is predicted based on the class with the highest confidence value among all classes. The decision function in a linear SVM is mathematically derived as,

$\displaystyle f\left(\textit{AB}\right)=w^{T}\textit{AB}+b$ (6)

Where $f\left({AB}\right)$ is the decision function, $w$ is the weight vector, AB are the fused feature vectors from the multiplier, and $b$ is the bias term. The class prediction is then determined by,

$\displaystyle\hat{Z}=\textit{argmax}_{k}(w^{T}_{k}\textit{AB}+b_{k})$ (7)

Where $\hat{Z}$ is the predicted class, $k$ represents each class, $w_{k}$ and $b_{k}$ are the weights and bias for class $k$ . In multi-class SVM, the decision boundaries are determined to maximize the margin between classes while minimizing classification error. Finally, the Multiclass SVMs are robust and can handle complex decision boundaries for classifying three different types of breathing abnormalities.

Table 2

Hyperparameter tuning of Multi-class SVM

Parameters	Values
Kernel type	rbf
Kernel degree	3
Regularization parameter	C
Class weight	None
Random state	None
Decision function shape	One-vs-rest
Gamma	Scale
Cache size	200
Verbose	False
Epsilon loss	0.1

Table 2 illustrates the hyperparameters of the multi-class SVM. In the multiclass classification, the SVM is tuned with the radial basis function (RBF) kernel that allows the model to create nonlinear decision boundaries by mapping the input data into a high-dimensional feature space. Selecting an appropriate value for gamma is scaled (0.0) for achieving good generalization performance. Particularly in the selection of RBF kernel, the parameter gamma reaches the high values in classification. The regularization parameter is C which specified as 0.1 to control the trade-off between maximizing the margin and minimizing the classification error. Epsilon loss specifies the variance among the predicted value and the actual target value falls within this margin of tolerance for errors.

4. Experimental results

This section presents the experimental arrangement and discusses the findings. To evaluate the proposed touchless detection of lung abnormality, parallel architecture of breathing patterns for three classes of respiration (normal, shallow and elevated) are utilized for training and testing samples. The same categorization results are obtained using MATLAB software and XSG tool. The overall experimental setup was displayed in the Fig. 6.

A framework for FPGA-based touchless detection accepting a sequence as input is designed for the experiments. The different breathing pattern calculated and then determines the video sequences for the classification of breathing abnormalities. The signal vectors are based on FPGA technology and reach a frequency of 100 KHz. The selected sample frequency allows the system to categorize vectors every 3 seconds after receiving 300 vectors. The proposed TO-LAB method consists of two phases: an online categorization stage and an offline training phase.

Table 3
Calculation of classifying the different breathing abnormalities

Condition	Heart rate	Blood pressure rate	Classified result
Normal breath $+$ normal blood pressure	72 bpm	$>$ 120 mmHg	Normal breathing
Normal breath $+$ abnormal blood pressure	80 to 90 bpm	120 to 129 mmHg	Elevated breathing
Abnormal breath $+$ Abnormal blood pressure	$<$ 90 bpm	$<$ 129 mmHg	Shallow breathing

Table 4

Dataset description of the proposed model

Breathing patterns	MIT-BIH arrhythmia dataset	Simulated dataset	Total
Normal	1363	4089	5452
Shallow	1090	3270	4360
Elevated	1293	3879	5172
Total	3,650	11,238	14,888

Figure 6.

Experimental arrangement of the proposed TO-LAB framework.

Table 3 shows the condition statements for classifying the breathing abnormalities such as normal, shallow breathing and elevated breathing. The normal breathing ranges from 80 to 120 mmHg with heart rate of 72 bpm; Elevated breathing ranges from 120 to 129 mmHg with heart rate of 90 bpm; and Shallow Breathing ranges greater than 129 mmHg with heart rate more than 90 bpm. The analysis is carried out based on this calculation for classifying the breathing abnormalities.

4.1 Dataset description

In this work, we have collected 3650 samples from a publicly available MIT-BIH arrhythmia database [38]. The details of data distributions are shown in Table 4. We collected the sample with different class targets omprising of three types of lung abnormalities.

The MIT-BIH Arrhythmia Database accessible on the Physionet website, comprises various segments categorized as atrial premature beats (AP), ventricular premature beats (VP), normal beats (N), right bundle branch block beats (RB), and left bundle branch blocks (LB).

The data was recorded using a sample frequency of 360 Hz with two leads. In practical applications, certain samples are captured using a real-time camera setup. Among six distinct breathing tasks, 3650 samples were collected, resulting in 14,888 samples for each lasting 30 seconds of all breathing activities. These data are then augmented are used for training or analysis of the ML algorithm.

4.2 Performance analysis

A real-time wireless monitoring and classification system for respiratory problems is based on the acquisition of fine-grained CSI. There were three distinct conditions examined in this section: elevated, shallow, and normal breathing.

Table 5
Coefficients values for simulated breathing abnormalities

Coefficients	Values	Breathing abnormalities
		Normal breathing	Shallow breathing	Elevated breathing
Amplitude	A1	0.541	0.408	0.405
	A2	0.342	0.305	0.325
	A3	0.062	0.442	3.854
	A4	0.074	0.175	1.832
	A5	0.221	0.108	3.645
Frequenc	B1	0.002	0.012	0.027
	B2	0.023	0.016	0.002
	B3	0.007	0.002	0.026
	B4	0.019	0.014	0.031
	B5	0.003	0.006	0.025

Three respiratory regimes were assessed using three devices (USRPs, RGB-D depth-sensing camera and BP meters) with University of Glasgow ethical approval. For ten seconds each, the participant was instructed to breathe shallowly and elevated. In addition, different data are collected from the USRP and the BP meter is attached to the arms of the participant at the same time. During the testing and data collection of our experiment, many factors including any physical items that might interfere with the wireless signal.

Table 5 illustrates the efficacy of the proposed TO-LAB model based on amplitude and frequency of the signals. The three different breathing patterns namely normal, shallow and elevated breathing of the gathered signals is illustrated in Table 5. The breathing patterns are varied from 0.25 to 3.85 for the amplitude variations and the breathing patterns are varied from 0.002 to 0.028 for the frequency variations.

Table 6

Performance assessment of breathing factor with proposed model

Breathing factor	Polynomial model	MSE	Correlation	Class
Fast inhalation	Normal	1.11e-05	0.945	Elevated
	Fast	4.28e-05	0.987
Normal inhalation	Normal	2.22e-06	0.995	Normal
	Fast	8.33e-06	0.956
Fast exhalation	Normal	4.55e-03	0.972	Shallow
	Fast	6.36e-03	0.984

Table 6 shows the efficiency assessment of random breathing component with selected model. MSE is a measure of the average squared difference between corresponding values of the estimated signal and the true signal. It is often used as a metric to assess the accuracy of a prediction or estimation algorithm. The correlation coefficient is a statistical measure that quantifies the degree of association or linear relation among two variables. In the context of signals, the correlation coefficient is often used to assess the similarity or resemblance between two signals. The experiment was conducted in order to classify respiratory components according to two distinct classes using the proposed model. To differentiate between different respiratory classes, the proposed framework uses breathing rate to detect and link the extracted breathing factors to the proposed TO-LAB model. The effectiveness of the proposed model was determined using the specific parameters viz., precision (P), recall (R), F1 score (F1), accuracy (A), and specificity (S).

$\displaystyle A=\frac{\left({\textit{TP}+\textit{FP}}\right)}{\left({\textit{% TP}+\textit{TN}+\textit{FN}+\textit{FP}}\right)}$ (8) $\displaystyle P=\frac{\textit{TP}}{\textit{TP}+\textit{FP}}$ (9) $\displaystyle S=\frac{\textit{TN}}{\textit{TN}+\textit{FP}}$ (10) $\displaystyle R=\frac{\textit{TP}}{\textit{TP}+\textit{FN}}$ (11) $\displaystyle\textit{F1}=2\left({\frac{P\ast R}{P+R}}\right)$ (12)

where TP and TN signifies true positives and negatives of the sample images, FP and FN indicates false positives and negatives of the input data samples.

The results attained by the multi-class SVM for the simulated and real-time datasets is illustrated in Table 7. The multi-class SVM reaches the average accuracy of 99.54% and 96.15% for simulated and real-time datasets. From this analysis, the multi-class SVM acquires the overall P, R, S and F1 of 97.34%, 95.37%, 93.64% and 91.42% respectively for simulated dataset.

Table 7

Performance analysis of the proposed model

Datasets	Breathing patterns	A	S	P	R	F1
Simulated datasets	Normal	99.76	97.18	97.21	97.14	98.62
	Shallow	99.34	96.23	96.18	97.32	98.25
	Elevated	99.51	97.02	97.22	98.08	98.43
Real-time datasets	Normal	95.12	96.14	97.42	98.15	98.52
	Shallow	96.08	95.06	96.25	96.42	98.28
	Elevated	97.25	98.17	98.11	98.14	98.02

Figure 7.

Validation curve of the proposed model for (a) simulated data and (b) real-time data.

Figure 7 displays the validation curve of the proposed approach for the simulated dataset and real-time dataset. The overall accuracy attained by the proposed approach is 99.54% as shown in Fig. 7(a) for the simulated data. The overall accuracy attained by the proposed approach is 96.15% as shown in Fig. 7(b) for the real-time data. A significant gap between the training and validation curves that indicate overfitting. Moreover, the loss rate of the proposed network is low that attains higher accuracy level.

Table 8

Quantitative evaluation based on MSE and CC parameters

Data sample	Mean square error	Correlation coefficient
1	0.015	0.978
2	0.023	0.945
3	0.007	0.964
4	0.005	0.958
5	0.017	0.973

Table 8 shows that the evaluation of the proposed approach in terms of the MSE and CC for classification. The correlation coefficient applied to assess the similarity between an original signal and a processed or distorted version of that signal. A lower MSE value indicates a smaller average error between the estimated and true signals, suggesting better accuracy of the estimation algorithm.

A confusion matrix (Fig. 8) used in classification to evaluate the efficacy of a classification algorithm. The confusion matrix is based on the comparison between the predicted labels and the true labels of a set of images. The Fig. 6(a) shows high classification accuracy for simulated data than the Fig. 6(b) for simulated data to classify the breathing abnormalities. This matrix is particularly useful for analysing the Multi-class SVM to classify signals into different abnormal categories in the context of signal classification.

4.3 Comparative analysis

In this section, a comparative examination of the proposed TO-LAB model with the existing machine learning models are discussed. To estimate that the fallouts of the proposed TO-LAB is more effective on the basis of A, S, P, F1, and R. Table 6 illustrates the comparison of different machine learning classifiers like Naïve Bayes (NB), Decision tree (DT), Random Forest (RF) and Support vector machine (SVM) with the proposed Multi-class SVM.

The effectiveness of each machine learning method was assessed based on its specific parameters, as shown in Table 8. Accuracy, prediction speed, and training duration were measured and observed per second respectively.

Table 9
Comparative analysis of existing machine learning models

Networks	A	S	P	R	F1
Naïve Bayes [39]	87.22	85.47	83.34	82.53	80.38
Decision tree [40]	89.41	88.56	85.27	84.63	86.75
Random forest [41]	93.67	93.64	92.47	91.78	90.87
SVM [42]	95.15	93.64	92.47	91.78	90.87
Multi-class SVM	99.54	97.34	95.37	93.64	91.42

Figure 8.

Confusion matrix of Multi-class SVM (a) for simulated data (b) for real-time data.

Four different machine learning methods are employed as classifiers. As shown in Table 9, the efficacy of each ML model was assessed in terms of accuracy and other parameters. Among all ML algorithms, Multi-class SVM has the highest classification accuracy for the simulated data. The comparison of the existing ML algorithms for classification of breathing abnormalities based on the specific parameters. The Multi-class SVM achieved a high accuracy than the prior methods. From Table 9, it is clear that Multi-class SVM better than the other ML models like NB, DT, RF and SVM. Relatively, Multi-class SVM attains high accuracy of 99.54%, that was greater than the other ML classifiers in the categorization stage. From Table 8, the proposed Multi-class SVM increases the overall accuracy by 12.07%, 10.17%, 5.89%, and 4.41% better than NB, DT, RF and SVM respectively.

Table 10

Accuracy comparison – Existing models vs TO-LAB model

Authors	Methods	Accuracy
Rehman et al. [22]	Non-contact SDR platform	99.04%
Melek et al. [24]	ML-MFCC	98.03%
Al-hababi et al. [19]	Non-contact sensing testbed	99.06%
Rahman et al. [20]	QUCoughScope	96.50%
Proposed	TO-LAB model	99.54%

Table 10 illustrates the comparison of existing techniques and proposed touchless detection model based on the simulated data. The proposed TO-LAB model improves the overall accuracy range of 0.50%, 1.01%, 0.48% and 3.065% better than [22, 24, 19, 20] respectively based on the simulated dataset. According to Table 10, the Proposed machine learning-based TO-LAB model yields the accuracy of 99.54% which is comparatively higher than existing works. Yet, the existing techniques not performs well when contrasted to the proposed TO-LAB framework. Thus, it is clearly seen that the proposed TO-LAB model has good competence than other existing techniques for the categorization of breathing patterns.

4.4 Ablation study

An experiment was conducted to assess the effectiveness of the proposed TO-LAB model using an ablation study on the gathered simulated dataset. Table 10 shows the comparison of removing the PCA and MFCC feature extraction techniques for the detection of lung abnormalities.

Table 11
Ablation analysis – with and without feature extraction techniques

Parameters	With PCA and without MFCC	Without PCA and with MFCC	With PCA and with MFCC
Accuracy	95.42%	97.08%	99.54%
Precision	93.21%	95.12%	98.21%
Recall	93.02%	95.87%	98.04%

This section extensively assessed the efficacy of the proposed TO-LAB model by various performance metrics, employing either a single feature extraction technique PCA or MFCC. From Table 11 MFCC is found to be more suitable for assessing breathing conditions and detecting subtle changes with high accuracy. Detailed breathing patterns allowed for the thorough evaluation of any anomalies or damage attributed to lung abnormalities. In this ablation analysis, the proposed TO-LAB model demonstrated superior accuracy (97.08%%) when utilizing MFCC, which outperformed than PCA (95.42%). Additionally, all other evaluated metrics exhibited variability alongside accuracy. Interestingly, combining both PCA and MFCC resulted in a notably higher level of recognition accuracy (99.54%). Through ablation analysis, it was determined that the proposed TO-LAB model utilizing both PCA and MFCC achieved high accuracy in categorizing breathing patterns into different classes.

5. Discussion

In this work, the goal is to detect lung disease by classifying the breathing abnormalities using ML algorithms with low complexity and high accuracy. For the experimental analysis, we used both the simulated data and real-time data for the detection of lung abnormalities. the results were assessed using multiple assessment parameters with MSE, correlation, A, P, R, F1, and S. Table 4 presents the efficiency evaluation of the random breathing component using the chosen model. The experiment aimed to classify respiratory components into two distinct classes utilizing the proposed model. To distinguish between different respiratory classes, the framework employs breathing rate to detect and associate the extracted breathing factors with the proposed TO-LAB model. Moreover, the efficiency of the proposed Multiclass SVM with A, P, R, F1, and S. Table 6 shows the performance analysis based on these parameters. The multi-class SVM reaches the average accuracy of 99.54% and 96.15% for simulated and real-time datasets. Additionally, the multi-class SVM reaches the overall P, R, S and F1 of 96.87%, 97.51%, 96.81% & 98.43% and 97.26%, 97.57%, 98.27% & 96.45 respectively for simulated and real-time datasets.

The Multi-class SVM was compared with the existing ML algorithms based on the forementioned parameters. Multi-class SVM yields maximum accuracy level of 99.54%, that was greater than the other ML classifiers in the categorization stage. It is clear that Multi-class SVM performs better than the other ML models like NB, DT, RF and SVM based on the simulated data as illustrated in Table 8. The training and prediction time complexity of the Multi-class SVM is specified as $O\left({k\cdot n2\cdot d}\right)$ and $O\left(k\right)$ . Moreover, the space complexity is $O\left({k\cdot n\cdot d}\right)$ , where $k$ is the number of classes, $n$ is the number of samples, and $d$ is the number of features. Multi-class SVM demonstrates competitive training and prediction time complexities while offering consistent performance across different dataset sizes. Additionally, its space complexity remains reasonable, making it a suitable choice for various machine learning tasks, especially when dealing with multi-class classification problems.

Moreover, the efficacy of the proposed TO-LAB model was assessed with the different prevailing models as depicted in Table 9. From this scrutiny, the proposed TO-LAB model improves the overall accuracy range of 0.50%, 1.01%, 0.48% and 3.065% better than Non-contact SDR platform, ML-MFCC, Non-contact sensing testbed and QUCoughScope respectively based on the simulated dataset.Hence, it is evident that the proposed TO-LAB model exhibits superior performance compared to other existing models in categorizing breathing patterns. Further, the ablation study thoroughly assessed the efficacy of the proposed TO-LAB model with various performance metrics, utilizing either PCA or MFCC as a single feature extraction technique. Results indicate that MFCC is more effective in evaluating breathing conditions and detecting subtle changes with greater accuracy, as observed from the Table 10. This ablation analysis revealed that the TO-LAB model, incorporating both PCA and MFCC attained impressive accuracy in classifying breathing patterns across various categories.

6. Conclusion

This work introduced a novel TO-LAB model for touchless detection of lung disease by classifying the breathing abnormalities as normal, shallow and elevated patterns. The touchless nature of the proposed TO-LAB model enhances patient comfort, reduces the risk of cross-contamination, and facilitates remote monitoring. The efficiency of this TO-LAB model was evaluated with the simulated and real-time data. The specific parameters such as accuracy, specificity, precision, recall and f1 score were calculated for evaluating the performance of Multi-class SVM. The proposed Multi-class SVM increases the overall accuracy by 12.07%, 10.17%, 5.89%, and 4.41% better than NB, DT, RF and SVM respectively. According to the findings, the proposed TO-LAB model attains the maximum accuracy of 96.15% for real time data; however, the accuracy increases to 99.54% for simulated data for the efficient classification of breathing abnormalities. The proposed TO-LAB model improves the overall accuracy range of 0.50%, 1.01%, and 0.48% better than Non-contact SDR platform, ML-MFCC, and Non-contact sensing test bed respectively based on the simulated dataset. From this analysis, our model attains better results in simulated data but it declines the accuracy while processing with real-time data. Moreover, this work has a significant medical impact since it presents a solution to the problem of gathering enough data during the epidemic to create a realistic model with a large dataset. In the future, we aim to enhance the data collection by augmenting and concurrently monitoring multiple individuals and incorporating additional environmental factors. Additionally, we plan to fine-tune advanced deep learning models to improve the performance of our model on the specific dataset.

Ethical approval

My research guide reviewed and ethically approved this manuscript for publishing in this Journal.

Author contributions

The authors confirm contribution to the paper as follows: Study conception and design: C. Rajeshkumar, K. Ruba Soundar; Data collection: C. Rajeshkumar; Analysis and interpretation of results: K. Ruba Soundar; Draft manuscript preparation: C. Rajeshkumar, K. Ruba Soundar. All authors reviewed the results and approved the final version of the manuscript.

Funding

None.

Availability of data and material

Data sharing is not applicable to this article as no new data were created or analyzed in this Research.

Human and animal rights

This article does not contain any studies with human or animal subjects performed by any of the authors.

Informed consent

I certify that I have explained the nature and purpose of this study to the above-named individual, and I have discussed the potential benefits of this study participation. The questions the individual had about this study have been answered, and we will always be available to address future questions.

Footnotes

Acknowledgments

The author would like to express his heartfelt gratitude to the supervisor for his guidance and unwavering support during this research for his guidance and support.

Conflict of interest

This paper has no conflict of interest for publishing.

References

Hemamalini

And

Nachiyappan

Geeitha

Motupalli

Kumar

Ahilan

Rajesh

. Integrating bio medical sensors in detecting hidden signatures of COVID-19 with Artificial intelligence. Measurement. 2022; 194: 111054.

World Health Organization. The top 10 causes of death. Retrieved from https//www.who.int/news-room/fact-sheets/detail/the-top-10-causes-of-death: 2020.

GBD 2019 Diseases and Injuries Collaborators. Global burden of 369 diseases and injuries in 204 countries and territories, 1990–2019: a systematic analysis for the Global Burden of Disease Study 2019. The Lancet. 2020; 396(10258): 1204-1222.

Tang

Zaidi

Lim

Govindasamy

Then

Das

Cheong

. Headway and the remaining hurdles of mesenchymal stem cells therapy for bronchopulmonary dysplasia. The Clinical Respiratory Journal. 2022; 16(10): 629-645.

Clark

Kochovska

Currow

. Burden of respiratory problems in low-income and middle-income countries. Current Opinion in Supportive and Palliative Care. 2022; 16(4): 210-215.

Bernardi

Gabutti

Porta

Spicuzza

. Slow breathing reduces chemoreflex response to hypoxia and hypercapnia, and increases baroreflex sensitivity. Journal of hypertension. 2001; 19(12): 2221-2229.

Spiesshoefer

Becker

Tuleta

Mohr

Diller

Emdin

Florian

Yilmaz

Boentert

Giannoni

. Impact of simulated hyperventilation and periodic breathing on sympatho-vagal balance and hemodynamics in patients with and without heart failure. Respiration. 2019; 98(6): 482-494.

Kumar

Ahilan

Fred

Kumar

. ROI extraction in CT lung images of COVID-19 using Fast Fuzzy C means clustering, In Biomedical Engineering Tools for Management for Patients with COVID-19. 2021; 103-119.

Kumar

Ritam

Han

Guo

Chandra

. Deep learning for predicting respiratory rate from biosignals. Computers in Biology and Medicine. 2022; 144: 105338.

10.

Zhang

Zhai

Zhang

. Recent Progress of Optical Imaging Approaches for Noncontact Physiological Signal Measurement: A Review. Advanced Intelligent Systems. 2023; 2200345.

11.

Des Jardins

Burton

. Clinical Manifestations & Assessment of Respiratory Disease-E-Book, Elsevier Health Sciences, 2023.

12.

Gnana Sophia

Thanammal

Sujatha

. Secure storage of lung brain multi-modal medical images using DNA homomorphic encryption. International Journal of Current Bio-Medical Engineering. 2023; 1(1): 16-22.

13.

Senthilkumar

Thenmozhi

Kumudavalli

Yedukondalu

. Hybrid statistical and recurrent neural network architecture implementation in FPGA device used for severe acute respiratory syndrome coronavirus detector. Journal of Intelligent & Fuzzy Systems, (Preprint), 2023; pp. 1-14.

14.

Appathurai

Manogaran

Chilamkurti

. Trusted FPGA-based transport traffic inject, impersonate (I2) attacks beaconing in the Internet of Vehicles. IET Networks. 2019; 8(3): 169-178.

15.

Anusha

Lokesh

Benila Christabel

. Deep reinforcement learning based real time violence detection. International Journal of Data Science and Artificial Intelligence. 2023; 1(1): 33-40.

16.

Sundarasekar

Appathurai

. Automatic Brain Tumor Detection and Classification Based on IoT and Machine Learning Techniques. Fluctuation and Noise Letters. 2022; 21(3): 2250030.

17.

Surendiran

Aarthi

Thangamani

Sugavanam

Sarumathy

. A Systematic Review Using Machine Learning Algorithms for Predicting Preterm Birth. International Journal of Engineering Trends and Technology. 2022; 70(5): 46-59.

18.

Ahilan

Angel Sajani

Jasmine Gnana Malar

Muthu Kumar

. Machine Learning-Based Brain Disease Classification Using EEG and MEG Signals, In International Conference on Frontiers of Intelligent Computing: Theory and Applications. 2023; pp. 487-498.

19.

Al-hababi

MAM

Khan

Al-Turjman

Zhao

Yang

. Non-contact sensing testbed for post-surgery monitoring by exploiting artificial-intelligence. Applied Sciences. 2020; 10(14): 4886.

20.

Rahman

Ibtehaz

Khandakar

Hossain

MSA

Mekki

YMS

Ezeddin

Bhuiyan

Ayari

Tahir

Qiblawey

Mahmud

. QUCoughScope: an intelligent application to detect COVID-19 patients using cough and breath sounds. Diagnostics. 2022; 12(4): 920.

21.

Khriji

Ammari

Messaoud

Bouaafia

Maraoui

Machhout

. COVID-19 recognition based on patient’s coughing and breathing patterns analysis: deep learning approach, In 2021 29th Conference of Open Innovations Association (FRUCT) IEEE. 2021; pp. 185-191.

22.

Purnomo

Lin

Adiprabowo

Hendria

. Non-contact monitoring and classification of breathing pattern for the supervision of people infected by COVID-19. Sensors. 2021; 21(9): 3172.

23.

Rehman

Shah

Khan

AbuAli

Shah

Yang

Alomainy

Imran

Abbasi

. Rf sensing based breathing patterns detection leveraging usrp devices. Sensors. 2021; 21(11): 3855.

24.

Melek

. Diagnosis of COVID-19 and non-COVID-19 patients by classifying only a single cough sound. Neural Computing and Applications. 2021; 33(24): 17621-17632.

25.

Vrindavanam

Srinath

Shankar

Nagesh

. Machine learning based COVID-19 cough classification models-a comparative analysis, In 2021 5th International Conference on Computing Methodologies and Communication (ICCMC) IEEE, 2021, pp. 420-426.

26.

Nassif

Shahin

Bader

Hassan

Werghi

. COVID-19 detection systems using deep-learning algorithms based on speech and image data. Mathematics. 2022; 10(4): 564.

27.

Dong

Qiao

Shang

Liao

Yuan

Cheng

Zhang

Wang

Chen

. Non-contact screening system based for COVID-19 on XGBoost and logistic regression. Computers in Biology and Medicine. 2022; 141: 105003.

28.

Alkhodari

Khandoker

. Detection of COVID-19 in smartphone-based breathing recordings: A pre-screening deep learning tool. PloS One. 2022; 17(1): 0262448.

29.

Badnjević

Cifrek

Koruga

. Integrated software suite for diagnosis of respiratory diseases. In Eurocon IEEE. 2013; 564-568.

30.

Jamil

Roy

. An efficient and robust phonocardiography (pcg)-based valvular heart diseases (vhd) detection framework using vision transformer (vit). Computers in Biology and Medicine. 2023; 158: 106734.

31.

Badnjevic

Cifrek

. Classification of asthma utilizing integrated software suite. In 6th European Conference of the International Federation for Medical and Biological Engineering: MBEC 2014, 7-11 September 2014, Dubrovnik. Croatia Springer International Publishing. 2015; 415-418.

32.

Liu

. Research on contactless intelligent medication pickup mode selection based on a hospital in China under COVID-19. Technology and Health Care, (Preprint), 2023; pp. 1-19.

33.

Badnjević

Pokvić

Smajlhodžić-Deljo

Spahić

Bego

Meseldžić

Prnjavorac

Bedak

. Application of artificial intelligence for the classification of the clinical outcome and therapy in patients with viral infections: The case of COVID-19. Technology and Health Care, (Preprint), 2023; pp. 1-12.

34.

Singh

Raj

Kumar

Verma

Roy

, Deep learning-based cost-effective and responsive robot for autism treatment. Drones. 2023; 7(2): 81.

35.

Badnjević

Koruga

Cifrek

Smith

Bego

. Interpretation of pulmonary function test results in relation to asthma classification using integrated software suite. In 2013 36th International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO), IEEE, 2013, pp. 140-144.

36.

Zhu

Zhang

Zhu

Gao

Liu

Zhou

Shan

, Severity detection of COVID-19 infection with machine learning of clinical records and CT images. Technology and Health Care, (Preprint). 2022; 1-16.

37.

Khan

Mustafa

Rehman

AbuAli

Yuan

Yang

Shah

Abbasi

. Non-contact smart sensing of physical activities during quarantine period using SDR technology. Sensors. 2022; 22(4): 1348.

38.

Yan

Zhang

. online automatic diagnosis system of cardiac arrhythmias based on MIT-BIH ECG database. Journal of Healthcare Engineering. 2021.

39.

Bhatia

Malhotra

. Naïve bayes classifier for predicting the novel coronavirus. In 2021 third international conference on intelligent communication technologies and virtual mobile networks (ICICV). IEEE, 2021, pp. 880-883.

40.

Ashleibta

Abbasi

Shah

Khalid

AbuAli

Imran

. Non-invasive RF sensing for detecting breathing abnormalities using software defined radios. IEEE Sensors Journal. 2020; 21(4): 5111-5118.

41.

Prabha

Yadav

Rani

Singh

. Respiratory effort signal-based sleep apnea detection system using improved random forest classifier. IETE Journal of Research. 2023; 69(9): 6326-6339.

42.

Bernard

Mounikaa

Nandhini

Mohanarathinam

. Breathe analysis for identification of respiratory abnormalities using svm. In 2020 6th International Conference on Advanced Computing and Communication Systems (ICACCS) IEEE. 2020, pp. 640-643.

43.

Muthukumar

Hevin Rajesh

Jenice Prabhu

. Reduancy aware dynamic routing protocol using salp swarm optimization algorithm. International Journal of System Design and Computing. 2023; 1(1): 35-42.

44.

Dakshina

Della Reasa

Bindhu

. Alzheimer disease detection via deep learning-based shuffle network. International Journal of Current Bio-Medical Engineering. 2023; 1(1): 9-15.

TO-LAB model: Real time Touchless Lung Abnormality detection model using USRP based machine learning algorithm

Abstract

BACKGROUND:

OBJECTIVE:

METHODS:

RESULTS:

CONCLUSION:

Keywords

1. Introduction

Table 1 Tabulation analysis of the existing literature works

3.2 Mel-frequency cepstral coefficients

Table 3 Calculation of classifying the different breathing abnormalities

4.2 Performance analysis

Table 5 Coefficients values for simulated breathing abnormalities

Table 9 Comparative analysis of existing machine learning models

Table 11 Ablation analysis – with and without feature extraction techniques

6. Conclusion

Ethical approval

Author contributions

Funding

Availability of data and material

Human and animal rights

Informed consent

Footnotes

Acknowledgments

Conflict of interest

References

Table 1
Tabulation analysis of the existing literature works

Table 3
Calculation of classifying the different breathing abnormalities

Table 5
Coefficients values for simulated breathing abnormalities

Table 9
Comparative analysis of existing machine learning models

Table 11
Ablation analysis – with and without feature extraction techniques