Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

An Innovative Machine Learning Approach for Classifying ECG Signals in Healthcare Devices

An Innovative Machine Learning Approach for Classifying ECG Signals in Healthcare Devices Hindawi Journal of Healthcare Engineering Volume 2022, Article ID 7194419, 16 pages https://doi.org/10.1155/2022/7194419 Research Article An Innovative Machine Learning Approach for Classifying ECG Signals in Healthcare Devices 1 2 3 4 Kishore B, A. Nanda Gopal Reddy , Anila Kumar Chillara, Wesam Atef Hatamleh, 4 5 6 7 Kamel Dine Haouam, Rohit Verma, B. Lakshmi Dhevi, and Henry Kwame Atiglah Department of Computer Science and Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka 576104, India Mahaveer Institute of Science and Technology, Hyderabad, India Birla Institute of Technology & Science, Hyderabad, India Department of Computer Science, College of Computer and Information Sciences, King Saud University, P. O. Box 51178, Riyadh 11543, Saudi Arabia School of Electronics, Dublin City University, Dublin, Ireland Institute of Artificial Intelligence and Machine Learning, Saveetha School of Engineering, Saveetha Institute of Medical and Technical Sciences, Chennai, India Department of Electrical & Electronics Engineering, Tamale Technical University, Tamale, Ghana Correspondence should be addressed to Henry Kwame Atiglah; hkatiglah@tatu.edu.gh Received 22 January 2022; Revised 20 February 2022; Accepted 23 February 2022; Published 13 April 2022 Academic Editor: Bhagyaveni M. A Copyright © 2022 Kishore B et al. (is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. An ECG is a diagnostic technique that examines and records the heart’s electrical impulses. It is easy to categorise and prevent computational abstractions in the ECG signal using the conventional method for obtaining ECG features. It is a significant issue, but it is also a difficult and time-consuming chore for cardiologists and medical professionals. (e proposed classifier eliminates all of the following limitations. Machine learning in healthcare equipment reduces moral transgressions. (is study’s primary purpose is to calculate the R-R interval and analyze the blockage utilising simple algorithms and approaches that give high accuracy. (e MIT-BIH dataset may be used to rebuild the data. (e acquired data may include both normal and abnormal ECGs. A Gabor filter is employed to generate a noiseless signal, and DCT-DOST is used to calculate the signal’s amplitude. (e amplitude is computed to detect any cardiac anomalies. A genetic algorithm derives the main highlights from the R peak and cycle segment length underlying the ECG signal. So, combining data with specific qualities maximises identification. (e genetic algorithm aids in hereditary computations, which aids in multitarget improvement. Finally, Radial Basis Function Neural Network (RBFNN) is presented as an example. An efficient feedforward neural network lowers the number of local minima in the signal. It shows progress in identifying both normal and abnormal ECG signals. abnormalities at regular interval. It is most helpful in di- 1. Introduction agnosing cardiac disorders such as myocardial infarction. Automatic electrocardiogram analysis is the best practice (e extended ischemia will continue till the cells start to die, utilized by clinicians for scrutinizing and recording the which is called myocardial infarction. In India, compiling functions of the heart by positioning the electrodes at the accurate data on sudden cardiac death, 5.5% of all-out external area of the skin membrane can be observed by mortality happened, and around 1-fifth of all the cardio- electrocardiogram device and a greater number of researches vascular passing and 6 lakh heart passing in the nation were are focused by scientists in recent years [1, 2]. Advancement suddenly occurred [3, 4]. A determination of myocardial in the technology provides enhancement in visualizing heart dead tissue is produced by incorporating the historical 2 Journal of Healthcare Engineering the first step in the process. Because of the limits of single- backdrop of the displaying ailment and physical investiga- tion with electrocardiogram discoveries. Free wall rupture is lead ECG analysis, which is lead dependent, a new multilead TWA detection is suggested to address these issues. For the a complicated one. In [5], it occurs in 1% of patients of acute myocardial infarction and it accounts for up to 7% of all purpose of translating the alternans-related information infarct-related deaths. Automatic ECG analysis works well in from the ST segment into a new signal known as the derived identification of cardiac-related problems in an enhanced lead, the Principal Component Analysis (PCA) approach is manner and for the better treatment. We mean to decide the utilized in this procedure. With the use of calibrated measure of heart tissue harm by multigoals examination of alternans records, the algorithm has been confirmed. Fig- electrocardiogram signals [6]. (e most critical component ure 1 represents the basic signal. In [11], the authors focus on determining the blockage of the ECG signal is the QRS complex, the pinnacle of which is indicated as R-peaks [7]. (e R-R intermission means the and R-R interval to achieve with good accuracy. Systematic finding of QRS complex is essential to extract the R-R in- time space between the two successive R tops. It is utilized to find the abnormalities in the heart normal operation called terval from the electrocardiogram recordings. To accurately analyze the cardiac rate variation, RR series plays a signif- arrhythmia. (e diagnosis involves an estimation of the size of infarct and to identify the acute complications [8]. In icant role and it is helpful to provide a quantitative evalu- ECG, Q and T waves play a major role in electrocardiogram ation of heart autonomic capacity in wellbeing and in signal. In P wave, if any problem occurs, it does not lead to sickness states [12]. In the past decades, wide collections of any complication. So, QRS detection is necessary to achieve algorithm and techniques were used in understanding au- our target. T-wave change occurs in larger area, and it tomatic regulation of heart beat. But the ECG recording may denotes ischemia and ST segment change occurs in lesser contain fictitious occurrences because of multiple disrup- tions like commotion interference in the signal, unexpected number of leads and it indicates the myocardial injury and Q-wave overlie and it denotes the main area of myocardial change in amplitude of QRS, and so on [13, 14]. Since there are so many methods for detecting a QRS signal as well as necrosis [9, 10]. Many researchers have worked in the area of medical field in performing analysis of cancer detection, preventing its propagation, it is important to pick a method that works in real time and can handle big datasets while electrocardiogram analysis, and so on. Sudden Cardiac Arrest (SCA) or Sudden Cardiac Death requiring little computational effort [15, 16]. In this study, (SCD) is one of the most common causes of cardiac mor- the noise interference that is present in the electrocardio- tality in the world, accounting for about one-third of all gram will be removed by handling preprocessing and then it cardiac deaths. If the danger of SCD can be detected at an is split up into samples by using the algorithm DCT-based early stage, it may be feasible to preserve the lives of patients DOST [17] and amplitude is computed in each interval. If by administering suitable treatment at the right time. (e there is any complication found in computing amplitude, it detects a block in such area. Initially it is of 100 hertz. It is risk of SCD may be detected by analyzing the conventional 12-lead Electrocardiogram (ECG) data, which are available split up into 5 intervals PQRST and amplitude is of 1 millivolt. Frequency is computed by f � 1/Tand the next step in most hospitals. Various studies have shown that differ- ences in the shape of the ECG, particularly in the ST-T wave involved in our work is feature extraction [18]. It helps to and QT segments, are directly associated with the risk of compute the mean and average of each interval and finally, SCD and Ventricular Arrhythmias (VA). Some of these Radial Basis Function Neural Network (RBFNN) is used to changes are so minute that they are not detectable just by analogize the trained and test data. (e data is collected from looking at the ECG for a short period of time. As a result, the MIT-BIH dataset. (e collected information has normal advanced-level computerised ECG algorithms are needed dataset and abnormal dataset [19]. (e trained and test for this new field of investigation. (e single and multilead dataset is analogized with the ratio 1 : 6 and the expected approaches used in this thesis are proved to be effective in accuracy is met. (e last objective of this work is to decide the perfect calculation for analogizing various classes of ECG the analysis of ST and QT segments. By using the multilead idea, the goal is to enhance the quantitative and qualitative oddities by quantitatively looking at the different QRS identification method to detect the blockage and R-R in- performances of the currently existing methodologies. T- Wave Alternans (TWA) and the QT interval, two nonin- terval and delineating their failure instance [20, 21]. By this vasive SCD indicators, are investigated in depth in this study. study, it achieves 98.5% accuracy. For the categorization of MI and healthy people, a novel Stationary Wavelet (SWT) method is proposed. Multilead 2. Proposed Works QT interval analysis is also carried out using three frank leads, designated as X, Y, and Z. Multilead and single-lead (ere are many databases accessible for public use, including approaches are used to analyze patients with a variety of the MIT-BIH arrhythmia database, which contains standard cardiac problems as well as healthy subjects. In addition, as a investigative material for the identification of cardiac ar- consequence of participation in the worldwide-level chal- rhythmias. It has been in use since 1980 for the purposes of lenge, the methods for measuring the foetal QRS and QT basic research and medical device development in the field of intervals were explained in the dissertation. PhysioNet/CinC cardiac rhythm and associated illnesses. 2013 is a collaborative effort. Brief description of the work (ere are many databases accessible for public use, in- given in the thesis is provided in the paragraphs following cluding the MIT-BIH arrhythmia database, which contains the thesis. (e creation of a multilead TWA detection idea is standard investigative material for the identification of Journal of Healthcare Engineering 3 T-OFFSET T-ONSET P-R SEGMENT S-T SEGMENT P-ONSET P-OFFSET P-R INTERVAL S-T INTERVAL QRS INTERVAL Figure 1: Structure of ECG signal. cardiac arrhythmias. It has been in use since 1980 for the Since the mid-1970s, our research group has inves- purposes of scientific research and medical device devel- tigated irregularities in heart rhythm (arrhythmias) as opment in the field of cardiac rhythm and associated ill- seen in long-term electrocardiograms (ECGs), as well as nesses. (e goal of creating the database is to develop automated approaches for detecting arrhythmias in real time. Other research groups in academia and business automated arrhythmia detectors that read the variety of the signal and, on the basis of that, can perform automated have pursued topics that are comparable to this one. Until 1980, anyone seeking to pursue such a career were re- cardiac diagnostics. (e many intricacies of the ECG, such as the variation of the waveform of the pulse and the ac- quired to gather their own information. Despite the fact companying cardiac beat, as well as the baffling strength of that the recordings themselves are copious, access to this artefacts and noise, combine to make signal analysis difficult. data is not ubiquitous, and comprehensive characteriza- As a result, automation of the recording of the Electro- tion of the recorded waveforms is a time-consuming and cardiogram (ECG) signal is clear, and many publicly ac- costly procedure. Aside from that, there is a great deal of cessible databases exist that store the recorded ECG signal variation in ECG rhythms and features of waveform for future medical use. (e MIT-BIH arrhythmia database is morphology, both across subjects and within persons over primarily utilized for medical and scientific purposes in- time, and therefore a meaningful representative collection of long-term ECGs for study must comprise a large cluding the identification and analysis of various cardiac arrhythmias. It is the goal of this database to create a number of recordings. completely automated environment in which precise in- Development of automated arrhythmia analysis algo- formation may be obtained for the diagnosis of ventricular rithms was slowed throughout the 1960s and 1970s due to a arrhythmias. scarcity of data that could be accessed by all researchers. Electrocardiograms (ECGs) are very popular because When doing such work, each group gathered its own col- they are a low-cost and noninvasive method of examining lection of recordings and often utilized the same data that the physiologic function of the heart. Initially developed in had been used to construct the algorithms in order to self- 1961, Holter introduced techniques for continuous re- evaluate their algorithms. From the beginning, it was evident cording of the ECG in ambulatory subjects for extended that the performance of these algorithms was inevitably periods of time. (e long-term ECG (Holter recording), data-dependent, and that the use of different data for the assessment of each algorithm made it impossible to make which typically lasts 24 hours, has since become the standard technique for observing transient aspects of cardiac electrical objective comparisons between algorithms belonging to activity. various groups of algorithms. 4 Journal of Healthcare Engineering 2 2 Data Collection and Selection. As soon as we realised we 1 (u − w) v g(u, v) � exp􏼨− 􏼢 + 􏼣􏼩, (3) would require a suitable set of well-characterized long-term 2 2 σ σ u v ECGs for our own research, we began collecting, digitising, and annotating long-term ECG recordings obtained by the where σu � 1/2πσx and σv � 1/2πσy, while the standard Arrhythmia Laboratory of Boston’s Beth Israel Hospital deviation of the elliptical Gaussian is represented as (BIH; now known as the Beth Israel Deaconess Medical σx and σy in the x- and y-axes. For exact amplitude esteems, Center), which was established in 1975. However, we the DC values of a 2D Gabor filter were used to minimize the intended to make these recordings accessible to the wider higher order harmonics. (e formula used to calculate the research community from the beginning, in order to spur filter parameter is more study in this area and to promote rigorously repeatable − 1/s− 1 and objectively comparable assessments of various methods u a � 􏼠 􏼡 , [3]. We anticipated that the availability of a shared database (4) would be a positive development. Our proposed system focuses on the blockage area to detect U , O � U /a(s−m) the R-R interval from the ECG signal. (is DCT-DOST seg- mentation with adaptive threshold is used in this paper to and σ is computed by using the equation determine the QRS complex and R peak from the recorded √�� � σ . (5) u�(a−1) U /(a+1) 2ln2 signals of the MIH-BIH database. (e distortion in the ECG is 0 filtered by a Gabor filter and therefore the QRS complex in- σ evaluated by using formation was preserved. After denoising, the signal gets seg- 1/2 mented into 256 constituent parts and the magnitude is found π σ (2ln2) σ u u (6) σ � tan􏼒 􏼓􏼢U 2 ln􏼢 􏼣􏼣􏼢2ln2 − 􏼣 . v n− to compare with trained data. It is performed to identify the 2 2k U U cardiac abnormality. (e difference in the amplitude and time period of the ECG sample helps to analyse the abnormality. Figure 4 represents the Gabor filter in the proposed Nearly 50,000 samples of ECG signals were considered for work. analysis. (e sampling frequency is split into 5 intervals to detect RR interval. (e mean, variance, and entropy are evaluated to extract the features. Genetic algorithm is used to select sig- 2.2. DCT-DOST-Based Segmentation. (is method uses the nificant features and is labelled with specific class. (e R peak, DCT-DOST scheme to examine the time domain repre- segment length, and mean value have been identified for the sentation of the ECG signal and to naturally distinguish the underlying ECG signal and finally using the RBFNN classifier, R peak. DFT rarely mentions the source signal in DOST. the test data is analogized with the trained ECG signal. (rough coefficient truncation, the signal in the case of Figure 2 represents the block diagram of proposed work. DOST will lose its structure. With DCT, however, it is more resistant to the loss of coefficients. DCT is highly regarded as it includes all frequencies to reduce unpredictability. (e 2.1. Preprocessing. Gabor filter is a type of linear filters advantages of DCT-DOST are that it blends vitality and whose response for impulse signal is characterized as a shows essential coefficients at lower frequencies. Gaussian function paired with a coherence function [22, 23]. (e linear S transform fills the gap among Fourier and (e requirement of minimal space bandwidth product wavelet transforms. (e S transfer of a signal h(t) is makes this filter highly suitable for our proposed work. −α |f| 2 (t− t) f /2 − i2πft Figure 3 represents the frequency domain. To define the s(τ, f) � 􏽚 h(t)e− e dt. (7) 2π result of signal propagation in frequency domain, the unpre- dictable theory should surpass or equals the constant value. Window’s width is expressed as ΔtΔf � c, (1) 1 σ(f) � T � . (8) |f| where c is a constant, Δt,ΔfΔ is the time and frequency space measurement. δ(τ, f ) is a 1D time function that demonstrates how the In 2D type, the time variable t is supplanted by spatial magnitude change with time for a fixed frequency. (e coordinates (x, y), and the frequency f is superseded by space DOST of h(KT) is variables (u, v). In most cases, the 2D Gabor function is N−1 n m + n evaluated as follows: i2πmj/N (9) H􏼔jT, 􏼕 � 􏽘 H􏼔 􏼕 G(m, n)e , NT NT m�0 2 2 1 −x + y ⎡ ⎢ ⎤ ⎥ ⎢ ⎥ ⎣ ⎦ g(x, y) � exp exp(j2πf(x cos θ + y sin θ)). where 2 2 2πσ 2σ g g 2 2 2 2 2π m n /n (10) G(m, n) � e− , (2) In the frequency domain, where n extends from 1, 2, . . ., N − 1. Journal of Healthcare Engineering 5 PREPROCESSING SEGMENTATION INPUT DATA (GABOR FILTER) (DCT-DOST) CLASSIFIER FEATURE EXTRACTION (RBFNN) (GENETIC ALGORITHM) PERFORMANCE PARAMETER Figure 2: Block diagram of proposed work. identification is performed. Initially the sample frequency is of 100 hertz. It is split up into five intervals to accurately locate the R-R interval. DCT is chosen because of the fol- lowing reason. It is a true value transformation, and it is strategically placed in space to reduce the amount of time required. It does not include any negative frequency. Only positive frequencies are used, and there is no symmetry FREQUENCY ∆T ∆f ≥ c coefficient as a result, higher frequencies are needed to DOMAIN convert to frequency space during segmentation. Since the DCT-DOST contains no negative frequencies, the frequency width for any signal of length 2N is as follows: N � 1, (11) i−2 N � 2 , for 2≤ i≤ N − 1. (e DCT-DOST method is as follows. Initially, the info ECG signal propagates via N point TIME DCT. (is level produces the coefficients A A , . . . , A . (e 1, 2 n DOMAIN 0 1 2 acquired coefficients are split into subbands [2 , 2 , 2 , n−1 Figure 3: Time and frequency domain. . . .. . .2 ]. For each subband, β point inverse DCT oper- ation is performed to ensure the β bandwidth that is gen- erated in the frequency and decomposition is perpendicular. (e proposed work’s main goal is to automatically find Figure 6 shows input ECG signal example. the peak value of R. To detect the R peak in this proposed work, every heartbeat segment is assumed to consist of 105 patterns prior to the R top identification and 151 patterns 2.3. Feature Extraction. In ECG signal, feature extraction were generated after the retrieval of R peak. A sum of 256 helps to figure out the amplitude and interval values of patterns were taken to find the extension of cardiac pulse. (e P-QRS-T segment present in the ECG signal. (e primary advantage of determining the length of every cardiac pulse is goal of this suggested work is to identify the R-R interval and to accurately detect the R top corresponding to P and T signs extract the transitory and morphological highlights from the due to its minimum magnitude and are noise sensitiveness. data. By utilising highlight extraction, 19 transient highlights Figure 5 represents the R peak detection using DCT. including PQ, RR, and PT interim and 3 morphological After the retrieval of noiseless image, DCT-DOST ap- highlights were extricated from the ECG signal. Figure 7 proach is applied for performing operation and the peak represents the feature extraction of the proposed work. 6 Journal of Healthcare Engineering MAGNITUDE INPUT 2D GABOR SUMMING EXTRACTION IMAGE FILTER OPERATOR FILTERED IMAGE PHASE OPERATOR Figure 4: Gabor filter functionality. PRE-PROCESSING NOISE LESS DCT-DOST ECG SIGNAL IMAGE TRANSFORM IDENTIFICATION OF REAL PEAK FINDING R-PEAKS PROCESS ECG SIGNAL WITH R-PEAKS/ R-R INTERVALS Figure 5: R peak detection using DCT-DOST. (i) Y � dct(y); (ii) z � 0 (iii) For cyin[1, 2, 3, . . .]; (iv) Y[z, z + (z − 1)]; idct(y[z; z + cz − 1]); (v) end (vi) return y ALGORITHM 1: DCT-DOST method. (e maximal and minimal points of each beat of the (e least value and most value point were figured out in ECG signal were captured using morphological highlights. the first R peak and the next R peak and then it normalized (e equation is by taking the esteems between 0 and 1. Feature describing the position of P, Q, R, S, T peak and min(t) QRS duration has been computed by using the initial po- f(t) � f(t) − − min(t). (12) max(t) sition of the Q-wave to the end of the S-wave. (e QRS Journal of Healthcare Engineering 7 2.6 2.4 2.2 1.8 1.6 1.4 1.2 0.8 0.6 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Figure 6: Input ECG signal. COMMON TRAINING DATA FIRST 5 MIN OF RECORD MORPHOLOGICAL DETECTION OF PREPROCESSING FEATURE BEAT EXTRACTION RBFNN FEATURE VECTOR TEMPORAL FEATURE EXTRACTION CLASSIFICATION RESULT Figure 7: Working of feature extraction. complex is computed, which has a significant role in the (viii) Step 8: Identify Tpoint by finding the highest value detection of abnormality. ranging from Rloc + 25 to Rloc + 100. (ix) Step 9: Compute the duration of QRS complex by using the equation 2.4. Algorithm Used to Compute Duration of QRS Complex QRS(i, j) � ceil((SOFF(i, j) − QON(i, j))). (13) (i) Step 1: Read the signal. (ii) Step 2: Identify the duration of QRS complex (x) Step 10: Find X � QRS. waveform. (iii) Step 3: Execute the wavelet analysis. False negative detection of QRS complex is carried out by using the following. (iv) Step 4: Calculate the coefficients by using wavelet decomposition. (a) Premature ventricular complexes (v) Step 5: Identify R peak location in the signal by (b) Low amplitude taking 60% of its value as threshold. False positive detection is carried out by using the (vi) Step 6: Identify Q point by finding the smallest following. value ranging from Rloc − 50 to Rloc − 10. (a) Negative QRS complexes (vii) Step 7: Identify S point by finding the smallest value ranging from Rloc + 5 to Rloc + 50. (b) Low SNR 8 Journal of Healthcare Engineering (is QRS algorithm is helpful to extract the R-R interval. It was performed by using heart rate variability (HRV). It is defined as the interval among two successive R peaks. (e R-R interval was computed by using the equation INITIALIZATION rr(i) � rr(i + 1) − r(i); 1, 2, . . . , m − 1. (14) FITNESS ASSIGNMENT where r(i) is the peak time of ith wave. Figure 8 represents the structure of genetic algorithm. (e next step was to reduce the number of features. It is done SELECTION with the aid of a genetic algorithm. Recently, there has been a surge in the use of genetic algorithms to reduce enhance- ment issues. (is algorithm is used in high-complexity CROSSOVER executions and large sets of arrangements. It was utilized to improve the features for identifying ECG signals. It assists in extracting the most desirable characteristics and is incor- MUTATION porated into the following generation. (e next generation would choose the best conditions, while the others would be ignored. It begins to repeat and build a population by STOPPING producing a new population at each stage through selection, CRITERIA = FALSE crossover, and mutation and then continues in this manner. And finally, it applies a fitness function, which is STOPPING computed by CRITERIA = TRUE f.f � 􏽘 (t − out). (15) n Figure 8: Structure of genetic algorithm. i�1 n stands for the number of outputs, t stands for the goal output, and out stands for the actual output. Positive and activation is combined in a linear way by the output layer. negative values may be present in the fitness function. As a (e input layer is represented as an x ∈ R vector of real result, we cannot use fitness benefit directly. (e selection numbers. (e network’s result is R ⟶ R, which is given by operator is used to identify the best features associated with 􏼌􏼌 􏼌􏼌 the highest fitness value and passes them over to the next 􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌 φ(x) � 􏽘 a x􏼐 x − c 􏼑, (18) 􏼌􏼌 􏼌􏼌 i i generation. (e crossover operator swaps the selected in- i�1 dividuals chromosomes to produce offspring chromosomes. where the neurons present in the hidden layer are repre- f(xi) Chromosome i reproduce � . sented as N, C is the centre vector, and a is the neuron’s (16) i i 􏽐 f(xk) k�1 weight. (e parameters a , c , and β aid in optimizing the i i i fitness between φ and the signal. Figure 9 represents the (e final operator is then used to notify the bits in the RBNN network. chromosome. (e probability that the chromosome in the A typical RBF of the scalar input vector that is a first layer nth position will be estimated is calculated using is N − N + 1 P � . (17) n (x − c) 􏽐 i h(x) � exp􏼠− (19) i�1 (e GA algorithm aids in the optimization of neural Normalized and denormalized forms of the generated network results, and it works well to achieve high precision, input are also possible. But it is discovered to be in a sensitivity, and specificity, as well as providing output with nonnormalized state. (e equation is better classification. (e classification is performed by RBFNN. � � � � � � 􏽐 a ρ x − c 􏼐� �􏼑 i�1 i i φ(x) � � � , (20) � � � � 􏽐 ρ􏼐�x − c �􏼑 i�1 i 2.5. Radial Basis Function Neural Network. RBFNN is a function that is used in time series prediction, classification, where and approximation of function. It can be used for any type of � � � � � � � � ρ􏼐�x − c �􏼑 model, including linear and nonlinear, as well as any net- � � � � u􏼐 x − c 􏼑 � � � . (21) � � � � work. (e three layers are input layer, hidden layer, and � � 􏽐 ρ􏼐 x − c 􏼑 � � j�1 i output layer. (e input to the hidden layer is converted nonlinearly by the hidden layer. (e hidden layer’s (is input layer expression can also be expressed as Journal of Healthcare Engineering 9 RBF NEURONS INPUT VECTOR WEIGHTED SUMS CAT. 1 WEIGHTS CATEGORY 1 SCORE CATEGORY C SCORE CAT. C WEIGHTS μ IS THE PROTOTYPE TO COMPARE AGAINST Figure 9: RBFNN network. 2N n normal and patient datasets [24]. Nearly 80% of data are φ(x) � 􏽘 􏽘 e v x − c􏼁 , (22) chosen for training and 20% was considered for testing. (e ij ij i i�1 j�1 training dataset is represented as n pairs using the below equation: where a , if i ∈ [1, N], ⎧ ⎨ T �􏼈 x , y􏼁􏼉 . (28) i i i i�1 e � ij b , if i ∈ [N + 1, 2N], ij (e output of the training dataset is Y , and time pre- diction is done by predicting the successive value and fea- ⎧ ⎪ δ , if i ∈ [1, N], ⎨ ij ρ x−c (‖ ‖) tures of a sequence: v x − c􏼁 � � � ij i ⎪ � � ⎩ � � 􏼐x − c 􏼑ρ􏼐�x − c �􏼑, if ∈ [N + 1, 2N]. ij ij i . . . , y − 3, y − 2, y − 1, . . . . (29) t t t (23) In the denormalized form, � � 3. Results and Discussion � � � � ⎧ ⎪ δ u􏼐�x − c �􏼑, if i ∈ [1, N], ij i v x − c􏼁 � � � ij i (e proposed ECG classification method discussed in this ⎪ � � ⎩ � � 􏼐x − c 􏼑u􏼐�x − c �􏼑, if i ∈ [N + 1, 2N]. ij ij i paper is implemented in MATLAB to analyze ECG signals. (e proposed methodology is implemented in MATLAB (24) and the MIT-BIH dataset is used to validate [24]. (e In the normalized form, RBFNN classifier is trained with the data from the previous section, and its performance is evaluated using the sample 1, if i � j, δ � 􏼨 (25) ij ECG signal as an example. (e expected performance for the 0, if i≠ j. ECG signals at each subsequent stage of the proposed method is exhibited for detailed analysis. (e ECG specimen (e probability density function among the input and image taken for analyzing has been elaborated for 50,000 the output layer is estimated as samples. One of the sample ECG signals is shown in � � � � Figure 10. � � p(x) � 􏽚 p(xΔy)dy � 􏽘 ρ􏼐�x − c �􏼑. (26) (e process of the proposed methodology starts with i�1 filtering of noises using Gabor filter. (e two types of noises (e output y given an input x is in the ECG signal are high-frequency noises such as elec- tromyogram noise and Gaussian noise and low-frequency φ(x) � E(y | x) � 􏽚 yP(y | x)dy, (27) noises like baseline wandering, and power line interference causes misinterpretation [25]. To eliminate all these noises, where the conditional probability of y given x is denoted as orientation-specific encoding schemes like Gabor filter is P(y|x). used for analyzing the texture features of ECG signal. For performing classification, training and test datasets Analogous to input signal, the output of Gabor is more are obtained from MIT-BIH database, which has both precise and accurate [26]. 10 Journal of Healthcare Engineering GABOR FILTER OUTPUT RESULT INPUT ECG SIGNAL 1.5 0.8 0.6 0.4 0.5 0.2 -0.2 -0.4 -0.5 -0.6 -1 -0.8 0123456789 10 0 1 23 4 567 8 910 TIME IN SEC TIME IN SEC Figure 11: Gabor filter output. Figure 10: Input ECG signal. QRS-complex in an ECG signal Figure 11 represents the Gabor filter output. For further 1.3 processing with minimum data redundancy and to con- 0.989 straint the dataset integration, the filtered output is normalized. (e distance between the R-peak values is estimated by finding the absolute values, as shown in Figure 12. When the heart’s electrical function is assumed as a vector, it is easy to analyze the trajectory of the vectors peak. (e signal ECG is considered as projection of the heart’s -0.2542 electrical vector on the corresponding lead vector as a time function (amplified by the absolute magnitude of the lead vectors). It is depicted in Figure 13 below. -0.8147 Generally, the coefficients are dispersed based on the bandwidth. (e energies in the ECG signal are gathered together using DCT-DOST so as to represent the most 160 185 211 260 important coefficient at the low frequency. (e features that Samples are extracted using the DCT-DOST approach indicate the QRS-Complex time-recurrence attributes of the ECG signal and are un- Figure 12: R-peak values estimation. symmetrical in nature. Also the peak values in QRS polarity and the unexpected variations in QRS amplitude are de- tected. Figure 14 represents the energy results. Figure 16, the DCT-DOST segmentation method produced (e traditional filtering minimizes signal noise by the following results. delaying the QRS components. As QRS complex represents (e moving average filter is dedicated to removing high- the ventricular activity of heart, it is necessary to preserve frequency noises from the ECG signal by computing the them. (e zero-phase filtering minimizes phase distortion running mean on the predetermined window length. (is is and provides a compromise among filtering and data re- a moderately straightforward estimation that will smoothen tention. (e output of the zero-phase filter is depicted in both the signal and its anomalies. (e R top in the ECG sign Figure 15. is smoothed to around 33% of its unique height. (e low- It is composed of 112 patterns before the R top occurs frequency contents of the ECG signal are represented in and 144 patterns after the R top occurs; an aggregate of 256 Figure 17. patterns are chosen to find the length of every occasion (e QRS wave of the ECG is detected using zero crossing relating to window size. (e ECG portion is composed of point detection approach. (e dominant and low-frequency 112 patterns before the R top occurs and 144 patterns after contents in the ECG are roughly estimated. Ideally the the R top occurs. (e duration of each event is determined in number of zero crossing points should be low for QRS, while order to condense the great majority of the data collected in it can be high at other times. (e number of zero crossing relation to each cardiac event as much as possible. (e points is used to determine the QRS with low computational benefit of establishing the duration of each heart event is that cost. Figure 18 represents the zero crossing output. it allows you to discover the R top with more precision when (e R top in the QRS interim is the most significant compared to the P and T waves, which have a low magnitude component for examining the ECG signal. R top discovery in and are vulnerable to turbulence. (ose uneven time-re- ECG is a strategy that is generally used to analyze heart currence coefficients must be processed for the ECG signal in anomalies and gauge pulse fluctuation. It is natural that the order to describe their morphological characteristics, which magnitudes of genuine R tops are more than those for bogus are then employed for further investigation. As illustrated in pinnacles. (e primary request separation of the sign is VOLTS Voltage (mV) VOLTAGE IN VOLTS Journal of Healthcare Engineering 11 ECG ABSOLUTE VALUE RESULTS DCT-DOST TRANSFORM BASED ECG AMPLITIUDE DETECTION 1 0.3 0.9 0.2 0.8 0.7 0.1 0.6 0.5 0.4 -0.1 0.3 -0.2 0.2 0.1 -0.3 012345 678 9 10 -0.4 TIME IN SEC 012345 678 9 10 TIME IN SEC Figure 13: Estimation of absolute value in ECG signal. Figure 16: Segmented output of sample ECG signal. ENERGY RESULTS USING DCT-DOST MOVING AVERAGE FILTERING 0.9 0.1 0.8 0.7 0.05 0.6 0.5 0.4 0.3 -0.05 0.2 0.1 -0.1 012345 678 9 10 TIME IN SEC -0.15 012345 6789 10 Figure 14: Output of DCT-DOST approach. TIME IN SEC Figure 17: Output of moving average filter. ZERO PHASE FILTERING RESULTS 0.35 0.3 ZERO CROSSING POINT DETECTION USING DCT-DOST SCHEME 0.25 0.9 0.2 0.8 0.15 0.7 0.1 0.6 0.05 0.5 0.4 -0.05 0.3 012345 678 9 10 TIME IN SEC 0.2 Figure 15: Output of zero-phase filter. 0.1 012345 6789 10 utilized to store the incline data of the genuine pinnacles yet TIME IN SEC diminishes the slant data of the bogus pinnacles. (e pro- Figure 18: Zero crossing detector output. posed strategy can proficiently recognize R tops under different conditions like pattern float, uproarious sign, tall T waves, or a quite delayed waves. Figure 19 represents the achieved with the slope index when compared with the high peak detection. recurrence index. (is is depicted in Figure 20. To detect ischemia, the slope index is preferred, which (e QRS detection ensures the efficient extraction of beat outperforms the higher recurrence index model of the interval and the abnormalities in the heart function. (e bandpass filtered QRS signal as the average relative factor of improvement in the QRS sections is executed by the pro- variation is much higher. (e superior performances can be posed technique to eliminate the pattern meandering. In this VOLTAGE IN VOLTS VOLTAGE IN VOLTS VOLTAGE IN VOLTS VOLTAGE IN VOLTS VOLTAGE IN VOLTS VOLTAGE IN VOLTS 12 Journal of Healthcare Engineering R-PEAK DETECTION USING PROPOSED SYSTEM QRS DETECTION USING PROPOSED SYSTEM 0.5 -100 -0.5 -200 0 500 1000 1500 2000 2500 3000 3500 TIME IN ms -1 012345 6789 10 Figure 21: Detection of QRS using the proposed method. TIME IN SEC Figure 19: R-peak detection. R-R INTERVAL IDENTIFICATION ECG SIGNAL SLOPE IDENTIFICATION 0.08 0.06 0.04 0.02 200 -0.02 -0.04 -200 -0.06 -400 -0.08 0 50 100 150 200 250 300 350 400 -0.1 TIME in ms -0.12 0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 Figure 22: Identification of RR interval. TIME IN SEC Figure 20: Slope identification of ECG. samples, our system outperforms the competition [28]. (e proposed method’s reliability is guaranteed since its effi- ciency is consistently high and without compromise. Fig- paper, the QRS fiducial focuses are detected to perceive the R ure 23 represents the accuracy comparison. point using QRS complex so that heart function classifica- tion can be accomplished simultaneously. Figure 21 repre- (e sensitivity shows the true positive value of the classification. It is calculated as the percentage of positives sents the QRS detection. (e RR-interim is resolved to obtain the dynamic that are correctly categorised [29]. With a maximum sen- sitivity of 98.3%, it outperforms the current system, while qualities of the ECG signal. (e 4 RR attributes that are discussed in this paper are pre-RR, post-RR, neighbor- CNN and SVM have maximum sensitivity of 92 percent and 86 percent, respectively. Figure 24 illustrates the sensitivity hood RR, and mean RR interim. (e interim between a past R top and the present R top is processed to find the relation. Figure 24 represents the sensitivity comparison. (e proposed method’s specificity values change in a zig- pre-RR attribute, while the interim between a specific R top and the successive R top is estimated to find the post- zag pattern as the number of samples increased [30], with a RR highlight. (e combined features of the pre- and post- maximum specificity of 99% for the proposed method and 93 percent and 95.6 percent for CNN and SVM classifiers, RR interim represent the momentary cadence charac- teristics. (e mean RR interim features are determined by respectively [31]. Figure 25 represents the comparison of specificity. (e averaging the RR interims of the previous 3-minimum RR interval of a specific occasion. Figure 22 represents the RR measure of various contents in the ECG signal [32] such as class, sinus rhythm, artifact, ventricular tachycardia, atrial interval. Similarly, the neighborhood RR features are inferred by fibrillation, bigeminy, and PVC is computed in terms of R, P, S, and F1. From the comparison table, it is clear that the averaging all the RR-interims of the previous episodes of a specific occasion [27]. (e neighborhood and mean high- estimation [33] by the proposed RBFNN is more than the conventional methods. Table 1 represents the accuracy lights indicate the mean qualities. (ese 4 highlights are comparison [34]. connected to the morphological list of the ECG signal. (e proposed method’s performance is compared with (e training, validation, and testing efficiencies of the proposed method are compared with conventional methods. the traditional methods such as CNN and SVM. With a maximum accuracy of 98.5% for different numbers of test (e training efficiency of our method is much higher than VOLTAGE IN VOLTS VOLTAGE IN VOLTS VOLTAGE IN mV. VOLTAGE in mV Journal of Healthcare Engineering 13 ACCURACY COMPARISION 020 40 60 80 100 NO. OF SAMPLES RBFNN CLASSIFICATION CNN CLASSIFICATION SVM CLASSIFICATION Figure 23: Accuracy comparison of different methods. SENSITIVITY COMPARISION 020 40 60 80 100 NO. OF SAMPLES RBFNN CLASSIFICATION CNN CLASSIFICATION SVM CLASSIFICATION Figure 24: Sensitivity comparison. SPECIFICITY COMPARISION 0 204060 80 NO. OF SAMPLES RBFNN CLASSIFICATION CNN CLASSIFICATION SVM CLASSIFICATION Figure 25: Comparison of specificity. ACCURACY IN (%) SENSITIVITY IN (%) SENSITIVITY IN (%) 14 Journal of Healthcare Engineering Table 1: Comparison of aggregate accuracy. Aggregate accuracy comparison Model Training (%) Validation (%) Test (%) Baseline—LSTM 66.8 66.3 65.6 Baseline—CNN 68.6 72.2 68.8 Stacked unidirectional—LSTM 80.5 78.1 79.2 Stacked bidirectional—LSTM 82.2 79.5 80.2 Stacked unidirectional—LSTM 80.4 79.4 79.3 Deep residual—CNN 84.7 75.3 74.7 Combined unidirectional LSTM—CNN 83.4 77.7 79.6 Combined bidirectional LSTM—CNN 93.2 74.8 76.8 Proposed RBFNN 99 84.4 98.5 Table 2: Comparison of classification metrics. BDLSTM Residual LSTM-CNN Proposed RBFNN Rhythm Class R P S F1 R P S F1 R P S F1 R P S F1 Sinus rhythm 0.82 0.83 0.94 0.84 0.64 0.88 0.86 0.76 0.79 0.80 0.95 0.79 0.85 0.87 0.96 0.89 Artifact/noise 0.88 0.82 0.94 0.83 0.89 0.97 0.94 0.82 0.81 0.83 0.94 0.81 0.89 0.85 0.92 0.84 Ventricular tachycardia 0.16 0.51 0.95 0.26 0.48 0.92 0.96 0.08 0.56 0.57 0.97 0.43 0.55 0.34 0.94 0.67 Atrial fibrillation 0.81 0.83 0.94 0.82 0.78 0.93 0.92 0.76 0.73 0.69 0.89 0.84 0.88 0.81 0.97 0.81 Bigeminy 0.72 0.65 0.82 0.67 0.89 0.98 0.98 0.16 0.67 0.67 0.96 0.55 0.84 0.83 0.91 0.80 PVC 0.78 0.76 0.88 0.76 0.78 0.93 0.93 0.83 0.79 0.77 0.92 0.72 0.81 0.82 0.95 0.89 Table 3: F1 score class comparison. F1 score class comparison Rhythm class BDLSTM Residual LSTM-CNN Proposed RBFNN Sinus rhythm 0.812 0.734 0.793 0.883 Artifact/noise 0.834 0.818 0.843 0.923 Ventricular tachycardia 0.265 0.169 0.417 0.721 Atrial fibrillation 0.837 0.763 0.764 0.852 Bigeminy 0.663 0.136 0.553 0.754 PVC 0.769 0.821 0.724 0.912 Overall 0.813 0.728 0.742 0.902 Table 4: F1 score class comparison. F1 score class comparison BDLSTM Residual LSTM-CNN Proposed RBFNN Rhythm class Multi Single Multi Single Multi Single Multi Single Sinus rhythm 0.812 0.612 0.734 0.692 0.793 0.702 0.883 0.813 Artifact/noise 0.834 0.734 0.818 0.746 0.843 0.774 0.923 0.874 Ventricular tachycardia 0.265 0.065 0.169 0.085 0.417 0.145 0.721 0.835 Atrial fibrillation 0.837 0.337 0.763 0.797 0.764 0.717 0.852 0.857 Bigeminy 0.663 0.263 0.136 0.073 0.553 0.523 0.754 0.873 PVC 0.769 0.669 0.821 0.709 0.724 0.709 0.912 0.879 the other methods [35]. Table 2 represents the classification 4. Conclusion metrics. Our proposed work enhances the diagnosis accuracy by From Table 3, the overall F1 score of the proposed eliminating the redundant and noise highlights. (e al- method is 90.2%, which is more the existing methods in gorithm presented here provides sensitivity and accuracy which the least performance is shown by the residual above 98.5%. (ese are computationally facile algorithm method [36]. that can be applied for practical application and aids in By concatenating the classification methods, the per- processing of a massive set of databases. By this work, the formance can be improved, which is shown in Table 4. Journal of Healthcare Engineering 15 objective gets achieved and the artifacts can be detected by References analogizing with the results from the algorithm for ad- [1] Y. Miao, Y. Tian, L. Peng, M. S. Hossain, and G. Muhammad, ditional analysis. It gives better acknowledgement ex- “Research and implementation of ECG-based biological actness when compared with various other existing recognition parallelization,” IEEE Access, vol. 18, no. 3, frameworks. In defining the path of development of ar- pp. 161–164, 2018. rhythmia detectors, the experience of the last 20 years, [2] L. Sun, Y. Lu, K. Yang, and S. Li, “ECG analysis using multiple since the release of the MIT-BIH Arrhythmia Database instance learning for myocardial infarction detection,” IEEE and the AHA Database soon afterwards, may be con- Transactions on Biomedical Engineering, vol. 59, no. 12, sidered a big experiment in shaping the direction of pp. 3348–3356, 2012. development. Performance statistics were of little or no [3] G. Da Poian, R. Bernardini, and R. Rinaldo, “Separation and value until databases became available, because it was analysis of fetal-ECG signals from compressed sensed ab- dominal ECG recordings,” IEEE Transactions on Biomedical widely accepted that each manufacturer designed its Engineering, vol. 63, no. 6, pp. 1269–1279, 2016. products using its own data and designed its statistics to [4] N. Bayasi, T. Tekeste, H. Saleh, B. Mohammad, A. Khandoker, present the products in a favourable light, so performance and M. Ismail, “low-power ECG-based processor for pre- statistics were of little or no value until databases became dicting ventricular arrhythmia,” IEEE Transactions on Very available. (e temptation to match their rivals’ products Large Scale Integration Systems, vol. 24, no. 5, pp. 1962–1974, feature for feature, rather of investing time and money on enhancements that could not be measured and hence did [5] X. Zhai and C. Tin, “Automated ECG classification using dual not add perceived value to the product, challenged con- heartbeat coupling based on convolutional neural network,” scientious developers who attempted to increase the ac- IEEE Access, vol. 6, no. 5, pp. 27466–27472, 2018. curacy of their algorithms. [6] M. J. Mollakazemi, F. Asadi, H. E. Orimi, S. A. Atyabi, Was the experiment a success or a failure, and what I. Uzelac, and A. Ghaffari, “Estimation of extent damage tissue were the outcomes? (e introduction of databases in the by multi resolution analysis of the electrocardiogram and early 1980s signaled a sea shift in the development arterial blood pressure,” in Proceedings of the 2015 Computing in Cardiology Conference (CinC), vol. 42, no. 9, pp. 1113–1116, community’s efforts. End users and regulatory bodies Nice, France, September 2015. started to inquire of manufacturers about the effective- [7] R. He, Y. Liu, K. Wang et al., “Automatic cardiac arrhythmia ness of their gadgets when subjected to conventional classification using combination of deep residual network and testing. Manufacturing companies could not avoid doing bidirectional LSTM,” IEEE Access, vol. 7, no. 7, the tests and reporting the results, and those whose al- pp. 102119–102135, 2019. gorithms did not perform as well as their rivals spent their [8] S. K. Yadav, R. Sinha, and P. K. Bora, “Electrocardiogram development expenditures on focused attempts to en- signal denoising using non-local wavelet transform domain hance performance. filtering,” IET Signal Processing, vol. 9, no. 2, pp. 88–96, 2015. As a result of the availability of databases, the overall [9] H. D. Hesar and M. Mohebbi, “A multi rate marginalized level of performance of commercial arrhythmia detectors particle extended kalman filter for P and T wave segmentation has increased significantly in recent years. in ECG signals,” IEEE Journal of Biomedical and Health In- Although it would be incorrect to suggest that manu- formatics, vol. 23, no. 1, pp. 112–122, 2019. [10] S. Xu, J. Li, K. Liu, and L. Wu, “A parallel GRU recurrent facturers were capable of producing significantly better network model and its application to multi-channel time- products in the late 1970s and instead chose to add bells and varying signal classification,” IEEE Access, vol. 7, no. 8, whistles in response to their customers’ apparent lack of pp. 118739–118748, 2019. interest in performance, it is possible to infer this from the [11] S. K. Bashar, E. Ding, A. J. Walkey, D. D. McManus, and fact that manufacturers were capable of producing signifi- K. H. Chon, “Noise detection in electrocardiogram signals for cantly better products in the late 1970s. intensive care unit patients,” IEEE Access, vol. 7, no. 7, pp. 88357–88368, 2019. [12] H. Liu, D. Chen, and G. Sun, “Detection of fetal ECG R wave Data Availability from single-lead abdominal ECG using a combination of RR (e data that support the findings of this study are available time-series smoothing and template-matching approach,” IEEE Access, vol. 7, no. 5, pp. 66633–66643, 2019. upon request from the corresponding author. [13] S. Jain, V. Bajaj, and A. Kumar, “Riemann liouvelle fractional integral based empirical mode decomposition for ECG denoising,” IEEE Journal of Biomedical and Health Infor- Conflicts of Interest matics, vol. 22, no. 4, pp. 1133–1139, 2018. (e authors declare that they do not have any conflicts of [14] J. Wang, M. She, S. Nahavandi, and A. Kouzani, “Human interest regarding the publication of this paper. identification from ECG signals via sparse representation of local segments,” IEEE Signal Processing Letters, vol. 20, no. 10, pp. 937–940, 2013. Acknowledgments [15] F. Khosrow-Khavar, K. Tavakolian, A. Blaber, and C. Menon, “Automatic and robust delineation of the fiducial points of the (e authors extend their appreciation to the researchers seismocardiogram signal for noninvasive estimation of car- supporting project number (RSP-2021/384), King Saud diac time intervals,” IEEE Transactions on Biomedical Engi- University, Riyadh, Saudi Arabia. neering, vol. 64, no. 8, pp. 1701–1710, 2017. 16 Journal of Healthcare Engineering [16] F. Alonso-Atienza, E. Morgado, L. Fernandez-Martinez, and Engineering in Health and Medicine, vol. 59, no. 12, pp. 3348–3356, 2015. A. Garc´ıa-Alberola, J. L. Rojo-Alvarez, Detection of life- threatening arrhythmias using feature selection and support [31] C. Venkatesan, P. Karthigaikumar, A. Paul, S. Satheeskumaran, and R. Kumar, “ECG signal preprocessing vector machines,” IEEE Transactions on Biomedical Engi- neering, vol. 61, no. 3, pp. 832–840, 2014. and SVM classifier-based abnormality detection in remote healthcare applications,” IEEE Access, vol. 6, no. 1, [17] S. Raj and K. C. Ray, “ECG signal analysis using DCT-based pp. 9767–9773, 2018. DOST and PSO optimized SVM,” IEEE Transactions on In- [32] Y. Park, D. YunII, and S.-H. Kang, “Preprocessing method for strumentation and Measurement, vol. 66, no. 3, pp. 470–478, performance enhancement in CNN-based STEMI detection from 12-lead ECG,” IEEE Access, vol. 7, no. 7, [18] S. R. Moosavi, E. Nigussie, M. Levorato, S. Virtanen, and pp. 99964–99977, 2019. J. Isoaho, “Low-latency approach for secure ECG feature [33] W. Zhu, X. Chen, Y. Wang, and L. Wang, “Arrhythmia based cryptographic key generation,” IEEE Access, vol. 6, recognition and classification using ECG morphology and no. 2, pp. 428–442, 2018. segment feature analysis,” IEEE/ACM Transactions on [19] A. O. Bicen, N. Z. Gurel, A. Dorier, and O. T. Inan, “Improved Computational Biology and Bioinformatics, vol. 16, no. 2, pre-ejection period estimation from ballistocardiogram and pp. 131–138, 2019. electrocardiogram signals by fusing multiple timing interval [34] S. Kiranyaz, T. Ince, and M. Gabbouj, “Real-time patient- features,” IEEE Sensors Journal, vol. 17, no. 7, pp. 4172–4180, specific ECG classification by 1-d convolutional neural net- works,” IEEE Transactions on Biomedical Engineering, vol. 63, [20] U. Satija, B. Ramkumar, and M. S. Manikandan, “A new no. 3, pp. 664–675, 2016. automated signal quality-aware ECG beat classification [35] L. Han, X. Pu, and X. Chen, “Method of fetal electrocar- method for unsupervised ECG diagnosis environments,” diogram extraction based on ]-support vector regression,” IEEE Sensors Journal, vol. 19, no. 1, pp. 277–286, 2019. IET Signal Processing, vol. 9, no. 7, pp. 430–439, 2015. [21] Q. A. Rahman, L. G. Tereshchenko, M. Kongkatong, [36] A. Smruthy and M. Suchetha, “Real-time classification of T. Abraham, M. R. Abraham, and H. Shatkay, “Utilizing ECG- healthy and apnea subjects using ECG signals with variational based heartbeat classification for hypertrophic cardiomyop- mode decomposition,” IEEE Sensors Journal, vol. 17, no. 5, athy identification,” IEEE Transactions on Nanobioscience, pp. 3092–3099, 2017. vol. 14, no. 7, pp. 505–512, 2015. [22] R. G. Sutar and A. G. Kothari, “Intelligent electrocardiogram pattern classification and recognition using low-cost cardio- care system,” IET Science, Measurement & Technology, vol. 9, no. 5, pp. 134–143, 2015. [23] H. E. Saadawy, M. Tantawi, H. A. Shedeed, and M. F. Tolba, “Hybrid hierarchical method for electrocardiogram heartbeat classification,” IET Signal Processing, vol. 12, no. 5, pp. 506–513, 2018. [24] P. K. Shukla, M. Zakariah, W. A. Hatamleh, H. Tarazi, and B. Tiwari, “AI-DRIVEN novel approach for liver cancer screening and prediction using cascaded fully convolutional neural network,” Journal of Healthcare Engineering, vol. 2022, Article ID 4277436, 14 pages, 2022. [25] P. K. Shukla, V. Roy, P. K. Shukla et al., “An Advanced EEG Motion Artifacts Eradication Algorithm,” Ke Computer Journal, p. bxab170, 2021. [26] N. K. Rathore, N. K. Jain, P. K. Shukla, U. Rawat, and R. Dubey, “Image forgery detection using singular value decomposition with some attacks,” National Academy Science Letters, vol. 44, no. 4, pp. 331–338, 2021. [27] A. Kumar, M. Saini, N. Gupta et al., “Efficient stochastic model for operational availability optimization of cooling tower using metaheuristic algorithms,” IEEE Access, vol. 10, p. 1, 2022, https://ieeexplore.ieee.org/ stamp/stamp.jsp? tp�&arnumber�9682734&isnumber�6514899. [28] D. Singh, V. Kumar, M. Kaur, M. Y. Jabarulla, and H.-N. Lee, “Screening of COVID-19 suspected subjects using multi- crossover genetic algorithm based dense convolutional neural network,” IEEE Access, vol. 9, pp. 142566–142580, 2021. [29] H. Kaushik, D. Singh, M. Kaur, H. Alshazly, A. Zaguia, and H. Hamam, “Diabetic retinopathy diagnosis from fundus images using stacked generalization of deep models,” IEEE Access, vol. 9, pp. 108276–108292, 2021. [30] D. B. Saadi, G. Tanev, M. Flintrup et al., “Automatic real-time embedded QRS complex detection for a novel patch-type electrocardiogram recorder,” IEEE Journal of Translational http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Healthcare Engineering Hindawi Publishing Corporation

An Innovative Machine Learning Approach for Classifying ECG Signals in Healthcare Devices

Loading next page...
 
/lp/hindawi-publishing-corporation/an-innovative-machine-learning-approach-for-classifying-ecg-signals-in-KkG7yx7Ewn
Publisher
Hindawi Publishing Corporation
Copyright
Copyright © 2022 Kishore B et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
ISSN
2040-2295
eISSN
2040-2309
DOI
10.1155/2022/7194419
Publisher site
See Article on Publisher Site

Abstract

Hindawi Journal of Healthcare Engineering Volume 2022, Article ID 7194419, 16 pages https://doi.org/10.1155/2022/7194419 Research Article An Innovative Machine Learning Approach for Classifying ECG Signals in Healthcare Devices 1 2 3 4 Kishore B, A. Nanda Gopal Reddy , Anila Kumar Chillara, Wesam Atef Hatamleh, 4 5 6 7 Kamel Dine Haouam, Rohit Verma, B. Lakshmi Dhevi, and Henry Kwame Atiglah Department of Computer Science and Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka 576104, India Mahaveer Institute of Science and Technology, Hyderabad, India Birla Institute of Technology & Science, Hyderabad, India Department of Computer Science, College of Computer and Information Sciences, King Saud University, P. O. Box 51178, Riyadh 11543, Saudi Arabia School of Electronics, Dublin City University, Dublin, Ireland Institute of Artificial Intelligence and Machine Learning, Saveetha School of Engineering, Saveetha Institute of Medical and Technical Sciences, Chennai, India Department of Electrical & Electronics Engineering, Tamale Technical University, Tamale, Ghana Correspondence should be addressed to Henry Kwame Atiglah; hkatiglah@tatu.edu.gh Received 22 January 2022; Revised 20 February 2022; Accepted 23 February 2022; Published 13 April 2022 Academic Editor: Bhagyaveni M. A Copyright © 2022 Kishore B et al. (is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. An ECG is a diagnostic technique that examines and records the heart’s electrical impulses. It is easy to categorise and prevent computational abstractions in the ECG signal using the conventional method for obtaining ECG features. It is a significant issue, but it is also a difficult and time-consuming chore for cardiologists and medical professionals. (e proposed classifier eliminates all of the following limitations. Machine learning in healthcare equipment reduces moral transgressions. (is study’s primary purpose is to calculate the R-R interval and analyze the blockage utilising simple algorithms and approaches that give high accuracy. (e MIT-BIH dataset may be used to rebuild the data. (e acquired data may include both normal and abnormal ECGs. A Gabor filter is employed to generate a noiseless signal, and DCT-DOST is used to calculate the signal’s amplitude. (e amplitude is computed to detect any cardiac anomalies. A genetic algorithm derives the main highlights from the R peak and cycle segment length underlying the ECG signal. So, combining data with specific qualities maximises identification. (e genetic algorithm aids in hereditary computations, which aids in multitarget improvement. Finally, Radial Basis Function Neural Network (RBFNN) is presented as an example. An efficient feedforward neural network lowers the number of local minima in the signal. It shows progress in identifying both normal and abnormal ECG signals. abnormalities at regular interval. It is most helpful in di- 1. Introduction agnosing cardiac disorders such as myocardial infarction. Automatic electrocardiogram analysis is the best practice (e extended ischemia will continue till the cells start to die, utilized by clinicians for scrutinizing and recording the which is called myocardial infarction. In India, compiling functions of the heart by positioning the electrodes at the accurate data on sudden cardiac death, 5.5% of all-out external area of the skin membrane can be observed by mortality happened, and around 1-fifth of all the cardio- electrocardiogram device and a greater number of researches vascular passing and 6 lakh heart passing in the nation were are focused by scientists in recent years [1, 2]. Advancement suddenly occurred [3, 4]. A determination of myocardial in the technology provides enhancement in visualizing heart dead tissue is produced by incorporating the historical 2 Journal of Healthcare Engineering the first step in the process. Because of the limits of single- backdrop of the displaying ailment and physical investiga- tion with electrocardiogram discoveries. Free wall rupture is lead ECG analysis, which is lead dependent, a new multilead TWA detection is suggested to address these issues. For the a complicated one. In [5], it occurs in 1% of patients of acute myocardial infarction and it accounts for up to 7% of all purpose of translating the alternans-related information infarct-related deaths. Automatic ECG analysis works well in from the ST segment into a new signal known as the derived identification of cardiac-related problems in an enhanced lead, the Principal Component Analysis (PCA) approach is manner and for the better treatment. We mean to decide the utilized in this procedure. With the use of calibrated measure of heart tissue harm by multigoals examination of alternans records, the algorithm has been confirmed. Fig- electrocardiogram signals [6]. (e most critical component ure 1 represents the basic signal. In [11], the authors focus on determining the blockage of the ECG signal is the QRS complex, the pinnacle of which is indicated as R-peaks [7]. (e R-R intermission means the and R-R interval to achieve with good accuracy. Systematic finding of QRS complex is essential to extract the R-R in- time space between the two successive R tops. It is utilized to find the abnormalities in the heart normal operation called terval from the electrocardiogram recordings. To accurately analyze the cardiac rate variation, RR series plays a signif- arrhythmia. (e diagnosis involves an estimation of the size of infarct and to identify the acute complications [8]. In icant role and it is helpful to provide a quantitative evalu- ECG, Q and T waves play a major role in electrocardiogram ation of heart autonomic capacity in wellbeing and in signal. In P wave, if any problem occurs, it does not lead to sickness states [12]. In the past decades, wide collections of any complication. So, QRS detection is necessary to achieve algorithm and techniques were used in understanding au- our target. T-wave change occurs in larger area, and it tomatic regulation of heart beat. But the ECG recording may denotes ischemia and ST segment change occurs in lesser contain fictitious occurrences because of multiple disrup- tions like commotion interference in the signal, unexpected number of leads and it indicates the myocardial injury and Q-wave overlie and it denotes the main area of myocardial change in amplitude of QRS, and so on [13, 14]. Since there are so many methods for detecting a QRS signal as well as necrosis [9, 10]. Many researchers have worked in the area of medical field in performing analysis of cancer detection, preventing its propagation, it is important to pick a method that works in real time and can handle big datasets while electrocardiogram analysis, and so on. Sudden Cardiac Arrest (SCA) or Sudden Cardiac Death requiring little computational effort [15, 16]. In this study, (SCD) is one of the most common causes of cardiac mor- the noise interference that is present in the electrocardio- tality in the world, accounting for about one-third of all gram will be removed by handling preprocessing and then it cardiac deaths. If the danger of SCD can be detected at an is split up into samples by using the algorithm DCT-based early stage, it may be feasible to preserve the lives of patients DOST [17] and amplitude is computed in each interval. If by administering suitable treatment at the right time. (e there is any complication found in computing amplitude, it detects a block in such area. Initially it is of 100 hertz. It is risk of SCD may be detected by analyzing the conventional 12-lead Electrocardiogram (ECG) data, which are available split up into 5 intervals PQRST and amplitude is of 1 millivolt. Frequency is computed by f � 1/Tand the next step in most hospitals. Various studies have shown that differ- ences in the shape of the ECG, particularly in the ST-T wave involved in our work is feature extraction [18]. It helps to and QT segments, are directly associated with the risk of compute the mean and average of each interval and finally, SCD and Ventricular Arrhythmias (VA). Some of these Radial Basis Function Neural Network (RBFNN) is used to changes are so minute that they are not detectable just by analogize the trained and test data. (e data is collected from looking at the ECG for a short period of time. As a result, the MIT-BIH dataset. (e collected information has normal advanced-level computerised ECG algorithms are needed dataset and abnormal dataset [19]. (e trained and test for this new field of investigation. (e single and multilead dataset is analogized with the ratio 1 : 6 and the expected approaches used in this thesis are proved to be effective in accuracy is met. (e last objective of this work is to decide the perfect calculation for analogizing various classes of ECG the analysis of ST and QT segments. By using the multilead idea, the goal is to enhance the quantitative and qualitative oddities by quantitatively looking at the different QRS identification method to detect the blockage and R-R in- performances of the currently existing methodologies. T- Wave Alternans (TWA) and the QT interval, two nonin- terval and delineating their failure instance [20, 21]. By this vasive SCD indicators, are investigated in depth in this study. study, it achieves 98.5% accuracy. For the categorization of MI and healthy people, a novel Stationary Wavelet (SWT) method is proposed. Multilead 2. Proposed Works QT interval analysis is also carried out using three frank leads, designated as X, Y, and Z. Multilead and single-lead (ere are many databases accessible for public use, including approaches are used to analyze patients with a variety of the MIT-BIH arrhythmia database, which contains standard cardiac problems as well as healthy subjects. In addition, as a investigative material for the identification of cardiac ar- consequence of participation in the worldwide-level chal- rhythmias. It has been in use since 1980 for the purposes of lenge, the methods for measuring the foetal QRS and QT basic research and medical device development in the field of intervals were explained in the dissertation. PhysioNet/CinC cardiac rhythm and associated illnesses. 2013 is a collaborative effort. Brief description of the work (ere are many databases accessible for public use, in- given in the thesis is provided in the paragraphs following cluding the MIT-BIH arrhythmia database, which contains the thesis. (e creation of a multilead TWA detection idea is standard investigative material for the identification of Journal of Healthcare Engineering 3 T-OFFSET T-ONSET P-R SEGMENT S-T SEGMENT P-ONSET P-OFFSET P-R INTERVAL S-T INTERVAL QRS INTERVAL Figure 1: Structure of ECG signal. cardiac arrhythmias. It has been in use since 1980 for the Since the mid-1970s, our research group has inves- purposes of scientific research and medical device devel- tigated irregularities in heart rhythm (arrhythmias) as opment in the field of cardiac rhythm and associated ill- seen in long-term electrocardiograms (ECGs), as well as nesses. (e goal of creating the database is to develop automated approaches for detecting arrhythmias in real time. Other research groups in academia and business automated arrhythmia detectors that read the variety of the signal and, on the basis of that, can perform automated have pursued topics that are comparable to this one. Until 1980, anyone seeking to pursue such a career were re- cardiac diagnostics. (e many intricacies of the ECG, such as the variation of the waveform of the pulse and the ac- quired to gather their own information. Despite the fact companying cardiac beat, as well as the baffling strength of that the recordings themselves are copious, access to this artefacts and noise, combine to make signal analysis difficult. data is not ubiquitous, and comprehensive characteriza- As a result, automation of the recording of the Electro- tion of the recorded waveforms is a time-consuming and cardiogram (ECG) signal is clear, and many publicly ac- costly procedure. Aside from that, there is a great deal of cessible databases exist that store the recorded ECG signal variation in ECG rhythms and features of waveform for future medical use. (e MIT-BIH arrhythmia database is morphology, both across subjects and within persons over primarily utilized for medical and scientific purposes in- time, and therefore a meaningful representative collection of long-term ECGs for study must comprise a large cluding the identification and analysis of various cardiac arrhythmias. It is the goal of this database to create a number of recordings. completely automated environment in which precise in- Development of automated arrhythmia analysis algo- formation may be obtained for the diagnosis of ventricular rithms was slowed throughout the 1960s and 1970s due to a arrhythmias. scarcity of data that could be accessed by all researchers. Electrocardiograms (ECGs) are very popular because When doing such work, each group gathered its own col- they are a low-cost and noninvasive method of examining lection of recordings and often utilized the same data that the physiologic function of the heart. Initially developed in had been used to construct the algorithms in order to self- 1961, Holter introduced techniques for continuous re- evaluate their algorithms. From the beginning, it was evident cording of the ECG in ambulatory subjects for extended that the performance of these algorithms was inevitably periods of time. (e long-term ECG (Holter recording), data-dependent, and that the use of different data for the assessment of each algorithm made it impossible to make which typically lasts 24 hours, has since become the standard technique for observing transient aspects of cardiac electrical objective comparisons between algorithms belonging to activity. various groups of algorithms. 4 Journal of Healthcare Engineering 2 2 Data Collection and Selection. As soon as we realised we 1 (u − w) v g(u, v) � exp􏼨− 􏼢 + 􏼣􏼩, (3) would require a suitable set of well-characterized long-term 2 2 σ σ u v ECGs for our own research, we began collecting, digitising, and annotating long-term ECG recordings obtained by the where σu � 1/2πσx and σv � 1/2πσy, while the standard Arrhythmia Laboratory of Boston’s Beth Israel Hospital deviation of the elliptical Gaussian is represented as (BIH; now known as the Beth Israel Deaconess Medical σx and σy in the x- and y-axes. For exact amplitude esteems, Center), which was established in 1975. However, we the DC values of a 2D Gabor filter were used to minimize the intended to make these recordings accessible to the wider higher order harmonics. (e formula used to calculate the research community from the beginning, in order to spur filter parameter is more study in this area and to promote rigorously repeatable − 1/s− 1 and objectively comparable assessments of various methods u a � 􏼠 􏼡 , [3]. We anticipated that the availability of a shared database (4) would be a positive development. Our proposed system focuses on the blockage area to detect U , O � U /a(s−m) the R-R interval from the ECG signal. (is DCT-DOST seg- mentation with adaptive threshold is used in this paper to and σ is computed by using the equation determine the QRS complex and R peak from the recorded √�� � σ . (5) u�(a−1) U /(a+1) 2ln2 signals of the MIH-BIH database. (e distortion in the ECG is 0 filtered by a Gabor filter and therefore the QRS complex in- σ evaluated by using formation was preserved. After denoising, the signal gets seg- 1/2 mented into 256 constituent parts and the magnitude is found π σ (2ln2) σ u u (6) σ � tan􏼒 􏼓􏼢U 2 ln􏼢 􏼣􏼣􏼢2ln2 − 􏼣 . v n− to compare with trained data. It is performed to identify the 2 2k U U cardiac abnormality. (e difference in the amplitude and time period of the ECG sample helps to analyse the abnormality. Figure 4 represents the Gabor filter in the proposed Nearly 50,000 samples of ECG signals were considered for work. analysis. (e sampling frequency is split into 5 intervals to detect RR interval. (e mean, variance, and entropy are evaluated to extract the features. Genetic algorithm is used to select sig- 2.2. DCT-DOST-Based Segmentation. (is method uses the nificant features and is labelled with specific class. (e R peak, DCT-DOST scheme to examine the time domain repre- segment length, and mean value have been identified for the sentation of the ECG signal and to naturally distinguish the underlying ECG signal and finally using the RBFNN classifier, R peak. DFT rarely mentions the source signal in DOST. the test data is analogized with the trained ECG signal. (rough coefficient truncation, the signal in the case of Figure 2 represents the block diagram of proposed work. DOST will lose its structure. With DCT, however, it is more resistant to the loss of coefficients. DCT is highly regarded as it includes all frequencies to reduce unpredictability. (e 2.1. Preprocessing. Gabor filter is a type of linear filters advantages of DCT-DOST are that it blends vitality and whose response for impulse signal is characterized as a shows essential coefficients at lower frequencies. Gaussian function paired with a coherence function [22, 23]. (e linear S transform fills the gap among Fourier and (e requirement of minimal space bandwidth product wavelet transforms. (e S transfer of a signal h(t) is makes this filter highly suitable for our proposed work. −α |f| 2 (t− t) f /2 − i2πft Figure 3 represents the frequency domain. To define the s(τ, f) � 􏽚 h(t)e− e dt. (7) 2π result of signal propagation in frequency domain, the unpre- dictable theory should surpass or equals the constant value. Window’s width is expressed as ΔtΔf � c, (1) 1 σ(f) � T � . (8) |f| where c is a constant, Δt,ΔfΔ is the time and frequency space measurement. δ(τ, f ) is a 1D time function that demonstrates how the In 2D type, the time variable t is supplanted by spatial magnitude change with time for a fixed frequency. (e coordinates (x, y), and the frequency f is superseded by space DOST of h(KT) is variables (u, v). In most cases, the 2D Gabor function is N−1 n m + n evaluated as follows: i2πmj/N (9) H􏼔jT, 􏼕 � 􏽘 H􏼔 􏼕 G(m, n)e , NT NT m�0 2 2 1 −x + y ⎡ ⎢ ⎤ ⎥ ⎢ ⎥ ⎣ ⎦ g(x, y) � exp exp(j2πf(x cos θ + y sin θ)). where 2 2 2πσ 2σ g g 2 2 2 2 2π m n /n (10) G(m, n) � e− , (2) In the frequency domain, where n extends from 1, 2, . . ., N − 1. Journal of Healthcare Engineering 5 PREPROCESSING SEGMENTATION INPUT DATA (GABOR FILTER) (DCT-DOST) CLASSIFIER FEATURE EXTRACTION (RBFNN) (GENETIC ALGORITHM) PERFORMANCE PARAMETER Figure 2: Block diagram of proposed work. identification is performed. Initially the sample frequency is of 100 hertz. It is split up into five intervals to accurately locate the R-R interval. DCT is chosen because of the fol- lowing reason. It is a true value transformation, and it is strategically placed in space to reduce the amount of time required. It does not include any negative frequency. Only positive frequencies are used, and there is no symmetry FREQUENCY ∆T ∆f ≥ c coefficient as a result, higher frequencies are needed to DOMAIN convert to frequency space during segmentation. Since the DCT-DOST contains no negative frequencies, the frequency width for any signal of length 2N is as follows: N � 1, (11) i−2 N � 2 , for 2≤ i≤ N − 1. (e DCT-DOST method is as follows. Initially, the info ECG signal propagates via N point TIME DCT. (is level produces the coefficients A A , . . . , A . (e 1, 2 n DOMAIN 0 1 2 acquired coefficients are split into subbands [2 , 2 , 2 , n−1 Figure 3: Time and frequency domain. . . .. . .2 ]. For each subband, β point inverse DCT oper- ation is performed to ensure the β bandwidth that is gen- erated in the frequency and decomposition is perpendicular. (e proposed work’s main goal is to automatically find Figure 6 shows input ECG signal example. the peak value of R. To detect the R peak in this proposed work, every heartbeat segment is assumed to consist of 105 patterns prior to the R top identification and 151 patterns 2.3. Feature Extraction. In ECG signal, feature extraction were generated after the retrieval of R peak. A sum of 256 helps to figure out the amplitude and interval values of patterns were taken to find the extension of cardiac pulse. (e P-QRS-T segment present in the ECG signal. (e primary advantage of determining the length of every cardiac pulse is goal of this suggested work is to identify the R-R interval and to accurately detect the R top corresponding to P and T signs extract the transitory and morphological highlights from the due to its minimum magnitude and are noise sensitiveness. data. By utilising highlight extraction, 19 transient highlights Figure 5 represents the R peak detection using DCT. including PQ, RR, and PT interim and 3 morphological After the retrieval of noiseless image, DCT-DOST ap- highlights were extricated from the ECG signal. Figure 7 proach is applied for performing operation and the peak represents the feature extraction of the proposed work. 6 Journal of Healthcare Engineering MAGNITUDE INPUT 2D GABOR SUMMING EXTRACTION IMAGE FILTER OPERATOR FILTERED IMAGE PHASE OPERATOR Figure 4: Gabor filter functionality. PRE-PROCESSING NOISE LESS DCT-DOST ECG SIGNAL IMAGE TRANSFORM IDENTIFICATION OF REAL PEAK FINDING R-PEAKS PROCESS ECG SIGNAL WITH R-PEAKS/ R-R INTERVALS Figure 5: R peak detection using DCT-DOST. (i) Y � dct(y); (ii) z � 0 (iii) For cyin[1, 2, 3, . . .]; (iv) Y[z, z + (z − 1)]; idct(y[z; z + cz − 1]); (v) end (vi) return y ALGORITHM 1: DCT-DOST method. (e maximal and minimal points of each beat of the (e least value and most value point were figured out in ECG signal were captured using morphological highlights. the first R peak and the next R peak and then it normalized (e equation is by taking the esteems between 0 and 1. Feature describing the position of P, Q, R, S, T peak and min(t) QRS duration has been computed by using the initial po- f(t) � f(t) − − min(t). (12) max(t) sition of the Q-wave to the end of the S-wave. (e QRS Journal of Healthcare Engineering 7 2.6 2.4 2.2 1.8 1.6 1.4 1.2 0.8 0.6 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Figure 6: Input ECG signal. COMMON TRAINING DATA FIRST 5 MIN OF RECORD MORPHOLOGICAL DETECTION OF PREPROCESSING FEATURE BEAT EXTRACTION RBFNN FEATURE VECTOR TEMPORAL FEATURE EXTRACTION CLASSIFICATION RESULT Figure 7: Working of feature extraction. complex is computed, which has a significant role in the (viii) Step 8: Identify Tpoint by finding the highest value detection of abnormality. ranging from Rloc + 25 to Rloc + 100. (ix) Step 9: Compute the duration of QRS complex by using the equation 2.4. Algorithm Used to Compute Duration of QRS Complex QRS(i, j) � ceil((SOFF(i, j) − QON(i, j))). (13) (i) Step 1: Read the signal. (ii) Step 2: Identify the duration of QRS complex (x) Step 10: Find X � QRS. waveform. (iii) Step 3: Execute the wavelet analysis. False negative detection of QRS complex is carried out by using the following. (iv) Step 4: Calculate the coefficients by using wavelet decomposition. (a) Premature ventricular complexes (v) Step 5: Identify R peak location in the signal by (b) Low amplitude taking 60% of its value as threshold. False positive detection is carried out by using the (vi) Step 6: Identify Q point by finding the smallest following. value ranging from Rloc − 50 to Rloc − 10. (a) Negative QRS complexes (vii) Step 7: Identify S point by finding the smallest value ranging from Rloc + 5 to Rloc + 50. (b) Low SNR 8 Journal of Healthcare Engineering (is QRS algorithm is helpful to extract the R-R interval. It was performed by using heart rate variability (HRV). It is defined as the interval among two successive R peaks. (e R-R interval was computed by using the equation INITIALIZATION rr(i) � rr(i + 1) − r(i); 1, 2, . . . , m − 1. (14) FITNESS ASSIGNMENT where r(i) is the peak time of ith wave. Figure 8 represents the structure of genetic algorithm. (e next step was to reduce the number of features. It is done SELECTION with the aid of a genetic algorithm. Recently, there has been a surge in the use of genetic algorithms to reduce enhance- ment issues. (is algorithm is used in high-complexity CROSSOVER executions and large sets of arrangements. It was utilized to improve the features for identifying ECG signals. It assists in extracting the most desirable characteristics and is incor- MUTATION porated into the following generation. (e next generation would choose the best conditions, while the others would be ignored. It begins to repeat and build a population by STOPPING producing a new population at each stage through selection, CRITERIA = FALSE crossover, and mutation and then continues in this manner. And finally, it applies a fitness function, which is STOPPING computed by CRITERIA = TRUE f.f � 􏽘 (t − out). (15) n Figure 8: Structure of genetic algorithm. i�1 n stands for the number of outputs, t stands for the goal output, and out stands for the actual output. Positive and activation is combined in a linear way by the output layer. negative values may be present in the fitness function. As a (e input layer is represented as an x ∈ R vector of real result, we cannot use fitness benefit directly. (e selection numbers. (e network’s result is R ⟶ R, which is given by operator is used to identify the best features associated with 􏼌􏼌 􏼌􏼌 the highest fitness value and passes them over to the next 􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌 φ(x) � 􏽘 a x􏼐 x − c 􏼑, (18) 􏼌􏼌 􏼌􏼌 i i generation. (e crossover operator swaps the selected in- i�1 dividuals chromosomes to produce offspring chromosomes. where the neurons present in the hidden layer are repre- f(xi) Chromosome i reproduce � . sented as N, C is the centre vector, and a is the neuron’s (16) i i 􏽐 f(xk) k�1 weight. (e parameters a , c , and β aid in optimizing the i i i fitness between φ and the signal. Figure 9 represents the (e final operator is then used to notify the bits in the RBNN network. chromosome. (e probability that the chromosome in the A typical RBF of the scalar input vector that is a first layer nth position will be estimated is calculated using is N − N + 1 P � . (17) n (x − c) 􏽐 i h(x) � exp􏼠− (19) i�1 (e GA algorithm aids in the optimization of neural Normalized and denormalized forms of the generated network results, and it works well to achieve high precision, input are also possible. But it is discovered to be in a sensitivity, and specificity, as well as providing output with nonnormalized state. (e equation is better classification. (e classification is performed by RBFNN. � � � � � � 􏽐 a ρ x − c 􏼐� �􏼑 i�1 i i φ(x) � � � , (20) � � � � 􏽐 ρ􏼐�x − c �􏼑 i�1 i 2.5. Radial Basis Function Neural Network. RBFNN is a function that is used in time series prediction, classification, where and approximation of function. It can be used for any type of � � � � � � � � ρ􏼐�x − c �􏼑 model, including linear and nonlinear, as well as any net- � � � � u􏼐 x − c 􏼑 � � � . (21) � � � � work. (e three layers are input layer, hidden layer, and � � 􏽐 ρ􏼐 x − c 􏼑 � � j�1 i output layer. (e input to the hidden layer is converted nonlinearly by the hidden layer. (e hidden layer’s (is input layer expression can also be expressed as Journal of Healthcare Engineering 9 RBF NEURONS INPUT VECTOR WEIGHTED SUMS CAT. 1 WEIGHTS CATEGORY 1 SCORE CATEGORY C SCORE CAT. C WEIGHTS μ IS THE PROTOTYPE TO COMPARE AGAINST Figure 9: RBFNN network. 2N n normal and patient datasets [24]. Nearly 80% of data are φ(x) � 􏽘 􏽘 e v x − c􏼁 , (22) chosen for training and 20% was considered for testing. (e ij ij i i�1 j�1 training dataset is represented as n pairs using the below equation: where a , if i ∈ [1, N], ⎧ ⎨ T �􏼈 x , y􏼁􏼉 . (28) i i i i�1 e � ij b , if i ∈ [N + 1, 2N], ij (e output of the training dataset is Y , and time pre- diction is done by predicting the successive value and fea- ⎧ ⎪ δ , if i ∈ [1, N], ⎨ ij ρ x−c (‖ ‖) tures of a sequence: v x − c􏼁 � � � ij i ⎪ � � ⎩ � � 􏼐x − c 􏼑ρ􏼐�x − c �􏼑, if ∈ [N + 1, 2N]. ij ij i . . . , y − 3, y − 2, y − 1, . . . . (29) t t t (23) In the denormalized form, � � 3. Results and Discussion � � � � ⎧ ⎪ δ u􏼐�x − c �􏼑, if i ∈ [1, N], ij i v x − c􏼁 � � � ij i (e proposed ECG classification method discussed in this ⎪ � � ⎩ � � 􏼐x − c 􏼑u􏼐�x − c �􏼑, if i ∈ [N + 1, 2N]. ij ij i paper is implemented in MATLAB to analyze ECG signals. (e proposed methodology is implemented in MATLAB (24) and the MIT-BIH dataset is used to validate [24]. (e In the normalized form, RBFNN classifier is trained with the data from the previous section, and its performance is evaluated using the sample 1, if i � j, δ � 􏼨 (25) ij ECG signal as an example. (e expected performance for the 0, if i≠ j. ECG signals at each subsequent stage of the proposed method is exhibited for detailed analysis. (e ECG specimen (e probability density function among the input and image taken for analyzing has been elaborated for 50,000 the output layer is estimated as samples. One of the sample ECG signals is shown in � � � � Figure 10. � � p(x) � 􏽚 p(xΔy)dy � 􏽘 ρ􏼐�x − c �􏼑. (26) (e process of the proposed methodology starts with i�1 filtering of noises using Gabor filter. (e two types of noises (e output y given an input x is in the ECG signal are high-frequency noises such as elec- tromyogram noise and Gaussian noise and low-frequency φ(x) � E(y | x) � 􏽚 yP(y | x)dy, (27) noises like baseline wandering, and power line interference causes misinterpretation [25]. To eliminate all these noises, where the conditional probability of y given x is denoted as orientation-specific encoding schemes like Gabor filter is P(y|x). used for analyzing the texture features of ECG signal. For performing classification, training and test datasets Analogous to input signal, the output of Gabor is more are obtained from MIT-BIH database, which has both precise and accurate [26]. 10 Journal of Healthcare Engineering GABOR FILTER OUTPUT RESULT INPUT ECG SIGNAL 1.5 0.8 0.6 0.4 0.5 0.2 -0.2 -0.4 -0.5 -0.6 -1 -0.8 0123456789 10 0 1 23 4 567 8 910 TIME IN SEC TIME IN SEC Figure 11: Gabor filter output. Figure 10: Input ECG signal. QRS-complex in an ECG signal Figure 11 represents the Gabor filter output. For further 1.3 processing with minimum data redundancy and to con- 0.989 straint the dataset integration, the filtered output is normalized. (e distance between the R-peak values is estimated by finding the absolute values, as shown in Figure 12. When the heart’s electrical function is assumed as a vector, it is easy to analyze the trajectory of the vectors peak. (e signal ECG is considered as projection of the heart’s -0.2542 electrical vector on the corresponding lead vector as a time function (amplified by the absolute magnitude of the lead vectors). It is depicted in Figure 13 below. -0.8147 Generally, the coefficients are dispersed based on the bandwidth. (e energies in the ECG signal are gathered together using DCT-DOST so as to represent the most 160 185 211 260 important coefficient at the low frequency. (e features that Samples are extracted using the DCT-DOST approach indicate the QRS-Complex time-recurrence attributes of the ECG signal and are un- Figure 12: R-peak values estimation. symmetrical in nature. Also the peak values in QRS polarity and the unexpected variations in QRS amplitude are de- tected. Figure 14 represents the energy results. Figure 16, the DCT-DOST segmentation method produced (e traditional filtering minimizes signal noise by the following results. delaying the QRS components. As QRS complex represents (e moving average filter is dedicated to removing high- the ventricular activity of heart, it is necessary to preserve frequency noises from the ECG signal by computing the them. (e zero-phase filtering minimizes phase distortion running mean on the predetermined window length. (is is and provides a compromise among filtering and data re- a moderately straightforward estimation that will smoothen tention. (e output of the zero-phase filter is depicted in both the signal and its anomalies. (e R top in the ECG sign Figure 15. is smoothed to around 33% of its unique height. (e low- It is composed of 112 patterns before the R top occurs frequency contents of the ECG signal are represented in and 144 patterns after the R top occurs; an aggregate of 256 Figure 17. patterns are chosen to find the length of every occasion (e QRS wave of the ECG is detected using zero crossing relating to window size. (e ECG portion is composed of point detection approach. (e dominant and low-frequency 112 patterns before the R top occurs and 144 patterns after contents in the ECG are roughly estimated. Ideally the the R top occurs. (e duration of each event is determined in number of zero crossing points should be low for QRS, while order to condense the great majority of the data collected in it can be high at other times. (e number of zero crossing relation to each cardiac event as much as possible. (e points is used to determine the QRS with low computational benefit of establishing the duration of each heart event is that cost. Figure 18 represents the zero crossing output. it allows you to discover the R top with more precision when (e R top in the QRS interim is the most significant compared to the P and T waves, which have a low magnitude component for examining the ECG signal. R top discovery in and are vulnerable to turbulence. (ose uneven time-re- ECG is a strategy that is generally used to analyze heart currence coefficients must be processed for the ECG signal in anomalies and gauge pulse fluctuation. It is natural that the order to describe their morphological characteristics, which magnitudes of genuine R tops are more than those for bogus are then employed for further investigation. As illustrated in pinnacles. (e primary request separation of the sign is VOLTS Voltage (mV) VOLTAGE IN VOLTS Journal of Healthcare Engineering 11 ECG ABSOLUTE VALUE RESULTS DCT-DOST TRANSFORM BASED ECG AMPLITIUDE DETECTION 1 0.3 0.9 0.2 0.8 0.7 0.1 0.6 0.5 0.4 -0.1 0.3 -0.2 0.2 0.1 -0.3 012345 678 9 10 -0.4 TIME IN SEC 012345 678 9 10 TIME IN SEC Figure 13: Estimation of absolute value in ECG signal. Figure 16: Segmented output of sample ECG signal. ENERGY RESULTS USING DCT-DOST MOVING AVERAGE FILTERING 0.9 0.1 0.8 0.7 0.05 0.6 0.5 0.4 0.3 -0.05 0.2 0.1 -0.1 012345 678 9 10 TIME IN SEC -0.15 012345 6789 10 Figure 14: Output of DCT-DOST approach. TIME IN SEC Figure 17: Output of moving average filter. ZERO PHASE FILTERING RESULTS 0.35 0.3 ZERO CROSSING POINT DETECTION USING DCT-DOST SCHEME 0.25 0.9 0.2 0.8 0.15 0.7 0.1 0.6 0.05 0.5 0.4 -0.05 0.3 012345 678 9 10 TIME IN SEC 0.2 Figure 15: Output of zero-phase filter. 0.1 012345 6789 10 utilized to store the incline data of the genuine pinnacles yet TIME IN SEC diminishes the slant data of the bogus pinnacles. (e pro- Figure 18: Zero crossing detector output. posed strategy can proficiently recognize R tops under different conditions like pattern float, uproarious sign, tall T waves, or a quite delayed waves. Figure 19 represents the achieved with the slope index when compared with the high peak detection. recurrence index. (is is depicted in Figure 20. To detect ischemia, the slope index is preferred, which (e QRS detection ensures the efficient extraction of beat outperforms the higher recurrence index model of the interval and the abnormalities in the heart function. (e bandpass filtered QRS signal as the average relative factor of improvement in the QRS sections is executed by the pro- variation is much higher. (e superior performances can be posed technique to eliminate the pattern meandering. In this VOLTAGE IN VOLTS VOLTAGE IN VOLTS VOLTAGE IN VOLTS VOLTAGE IN VOLTS VOLTAGE IN VOLTS VOLTAGE IN VOLTS 12 Journal of Healthcare Engineering R-PEAK DETECTION USING PROPOSED SYSTEM QRS DETECTION USING PROPOSED SYSTEM 0.5 -100 -0.5 -200 0 500 1000 1500 2000 2500 3000 3500 TIME IN ms -1 012345 6789 10 Figure 21: Detection of QRS using the proposed method. TIME IN SEC Figure 19: R-peak detection. R-R INTERVAL IDENTIFICATION ECG SIGNAL SLOPE IDENTIFICATION 0.08 0.06 0.04 0.02 200 -0.02 -0.04 -200 -0.06 -400 -0.08 0 50 100 150 200 250 300 350 400 -0.1 TIME in ms -0.12 0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 Figure 22: Identification of RR interval. TIME IN SEC Figure 20: Slope identification of ECG. samples, our system outperforms the competition [28]. (e proposed method’s reliability is guaranteed since its effi- ciency is consistently high and without compromise. Fig- paper, the QRS fiducial focuses are detected to perceive the R ure 23 represents the accuracy comparison. point using QRS complex so that heart function classifica- tion can be accomplished simultaneously. Figure 21 repre- (e sensitivity shows the true positive value of the classification. It is calculated as the percentage of positives sents the QRS detection. (e RR-interim is resolved to obtain the dynamic that are correctly categorised [29]. With a maximum sen- sitivity of 98.3%, it outperforms the current system, while qualities of the ECG signal. (e 4 RR attributes that are discussed in this paper are pre-RR, post-RR, neighbor- CNN and SVM have maximum sensitivity of 92 percent and 86 percent, respectively. Figure 24 illustrates the sensitivity hood RR, and mean RR interim. (e interim between a past R top and the present R top is processed to find the relation. Figure 24 represents the sensitivity comparison. (e proposed method’s specificity values change in a zig- pre-RR attribute, while the interim between a specific R top and the successive R top is estimated to find the post- zag pattern as the number of samples increased [30], with a RR highlight. (e combined features of the pre- and post- maximum specificity of 99% for the proposed method and 93 percent and 95.6 percent for CNN and SVM classifiers, RR interim represent the momentary cadence charac- teristics. (e mean RR interim features are determined by respectively [31]. Figure 25 represents the comparison of specificity. (e averaging the RR interims of the previous 3-minimum RR interval of a specific occasion. Figure 22 represents the RR measure of various contents in the ECG signal [32] such as class, sinus rhythm, artifact, ventricular tachycardia, atrial interval. Similarly, the neighborhood RR features are inferred by fibrillation, bigeminy, and PVC is computed in terms of R, P, S, and F1. From the comparison table, it is clear that the averaging all the RR-interims of the previous episodes of a specific occasion [27]. (e neighborhood and mean high- estimation [33] by the proposed RBFNN is more than the conventional methods. Table 1 represents the accuracy lights indicate the mean qualities. (ese 4 highlights are comparison [34]. connected to the morphological list of the ECG signal. (e proposed method’s performance is compared with (e training, validation, and testing efficiencies of the proposed method are compared with conventional methods. the traditional methods such as CNN and SVM. With a maximum accuracy of 98.5% for different numbers of test (e training efficiency of our method is much higher than VOLTAGE IN VOLTS VOLTAGE IN VOLTS VOLTAGE IN mV. VOLTAGE in mV Journal of Healthcare Engineering 13 ACCURACY COMPARISION 020 40 60 80 100 NO. OF SAMPLES RBFNN CLASSIFICATION CNN CLASSIFICATION SVM CLASSIFICATION Figure 23: Accuracy comparison of different methods. SENSITIVITY COMPARISION 020 40 60 80 100 NO. OF SAMPLES RBFNN CLASSIFICATION CNN CLASSIFICATION SVM CLASSIFICATION Figure 24: Sensitivity comparison. SPECIFICITY COMPARISION 0 204060 80 NO. OF SAMPLES RBFNN CLASSIFICATION CNN CLASSIFICATION SVM CLASSIFICATION Figure 25: Comparison of specificity. ACCURACY IN (%) SENSITIVITY IN (%) SENSITIVITY IN (%) 14 Journal of Healthcare Engineering Table 1: Comparison of aggregate accuracy. Aggregate accuracy comparison Model Training (%) Validation (%) Test (%) Baseline—LSTM 66.8 66.3 65.6 Baseline—CNN 68.6 72.2 68.8 Stacked unidirectional—LSTM 80.5 78.1 79.2 Stacked bidirectional—LSTM 82.2 79.5 80.2 Stacked unidirectional—LSTM 80.4 79.4 79.3 Deep residual—CNN 84.7 75.3 74.7 Combined unidirectional LSTM—CNN 83.4 77.7 79.6 Combined bidirectional LSTM—CNN 93.2 74.8 76.8 Proposed RBFNN 99 84.4 98.5 Table 2: Comparison of classification metrics. BDLSTM Residual LSTM-CNN Proposed RBFNN Rhythm Class R P S F1 R P S F1 R P S F1 R P S F1 Sinus rhythm 0.82 0.83 0.94 0.84 0.64 0.88 0.86 0.76 0.79 0.80 0.95 0.79 0.85 0.87 0.96 0.89 Artifact/noise 0.88 0.82 0.94 0.83 0.89 0.97 0.94 0.82 0.81 0.83 0.94 0.81 0.89 0.85 0.92 0.84 Ventricular tachycardia 0.16 0.51 0.95 0.26 0.48 0.92 0.96 0.08 0.56 0.57 0.97 0.43 0.55 0.34 0.94 0.67 Atrial fibrillation 0.81 0.83 0.94 0.82 0.78 0.93 0.92 0.76 0.73 0.69 0.89 0.84 0.88 0.81 0.97 0.81 Bigeminy 0.72 0.65 0.82 0.67 0.89 0.98 0.98 0.16 0.67 0.67 0.96 0.55 0.84 0.83 0.91 0.80 PVC 0.78 0.76 0.88 0.76 0.78 0.93 0.93 0.83 0.79 0.77 0.92 0.72 0.81 0.82 0.95 0.89 Table 3: F1 score class comparison. F1 score class comparison Rhythm class BDLSTM Residual LSTM-CNN Proposed RBFNN Sinus rhythm 0.812 0.734 0.793 0.883 Artifact/noise 0.834 0.818 0.843 0.923 Ventricular tachycardia 0.265 0.169 0.417 0.721 Atrial fibrillation 0.837 0.763 0.764 0.852 Bigeminy 0.663 0.136 0.553 0.754 PVC 0.769 0.821 0.724 0.912 Overall 0.813 0.728 0.742 0.902 Table 4: F1 score class comparison. F1 score class comparison BDLSTM Residual LSTM-CNN Proposed RBFNN Rhythm class Multi Single Multi Single Multi Single Multi Single Sinus rhythm 0.812 0.612 0.734 0.692 0.793 0.702 0.883 0.813 Artifact/noise 0.834 0.734 0.818 0.746 0.843 0.774 0.923 0.874 Ventricular tachycardia 0.265 0.065 0.169 0.085 0.417 0.145 0.721 0.835 Atrial fibrillation 0.837 0.337 0.763 0.797 0.764 0.717 0.852 0.857 Bigeminy 0.663 0.263 0.136 0.073 0.553 0.523 0.754 0.873 PVC 0.769 0.669 0.821 0.709 0.724 0.709 0.912 0.879 the other methods [35]. Table 2 represents the classification 4. Conclusion metrics. Our proposed work enhances the diagnosis accuracy by From Table 3, the overall F1 score of the proposed eliminating the redundant and noise highlights. (e al- method is 90.2%, which is more the existing methods in gorithm presented here provides sensitivity and accuracy which the least performance is shown by the residual above 98.5%. (ese are computationally facile algorithm method [36]. that can be applied for practical application and aids in By concatenating the classification methods, the per- processing of a massive set of databases. By this work, the formance can be improved, which is shown in Table 4. Journal of Healthcare Engineering 15 objective gets achieved and the artifacts can be detected by References analogizing with the results from the algorithm for ad- [1] Y. Miao, Y. Tian, L. Peng, M. S. Hossain, and G. Muhammad, ditional analysis. It gives better acknowledgement ex- “Research and implementation of ECG-based biological actness when compared with various other existing recognition parallelization,” IEEE Access, vol. 18, no. 3, frameworks. In defining the path of development of ar- pp. 161–164, 2018. rhythmia detectors, the experience of the last 20 years, [2] L. Sun, Y. Lu, K. Yang, and S. Li, “ECG analysis using multiple since the release of the MIT-BIH Arrhythmia Database instance learning for myocardial infarction detection,” IEEE and the AHA Database soon afterwards, may be con- Transactions on Biomedical Engineering, vol. 59, no. 12, sidered a big experiment in shaping the direction of pp. 3348–3356, 2012. development. Performance statistics were of little or no [3] G. Da Poian, R. Bernardini, and R. Rinaldo, “Separation and value until databases became available, because it was analysis of fetal-ECG signals from compressed sensed ab- dominal ECG recordings,” IEEE Transactions on Biomedical widely accepted that each manufacturer designed its Engineering, vol. 63, no. 6, pp. 1269–1279, 2016. products using its own data and designed its statistics to [4] N. Bayasi, T. Tekeste, H. Saleh, B. Mohammad, A. Khandoker, present the products in a favourable light, so performance and M. Ismail, “low-power ECG-based processor for pre- statistics were of little or no value until databases became dicting ventricular arrhythmia,” IEEE Transactions on Very available. (e temptation to match their rivals’ products Large Scale Integration Systems, vol. 24, no. 5, pp. 1962–1974, feature for feature, rather of investing time and money on enhancements that could not be measured and hence did [5] X. Zhai and C. Tin, “Automated ECG classification using dual not add perceived value to the product, challenged con- heartbeat coupling based on convolutional neural network,” scientious developers who attempted to increase the ac- IEEE Access, vol. 6, no. 5, pp. 27466–27472, 2018. curacy of their algorithms. [6] M. J. Mollakazemi, F. Asadi, H. E. Orimi, S. A. Atyabi, Was the experiment a success or a failure, and what I. Uzelac, and A. Ghaffari, “Estimation of extent damage tissue were the outcomes? (e introduction of databases in the by multi resolution analysis of the electrocardiogram and early 1980s signaled a sea shift in the development arterial blood pressure,” in Proceedings of the 2015 Computing in Cardiology Conference (CinC), vol. 42, no. 9, pp. 1113–1116, community’s efforts. End users and regulatory bodies Nice, France, September 2015. started to inquire of manufacturers about the effective- [7] R. He, Y. Liu, K. Wang et al., “Automatic cardiac arrhythmia ness of their gadgets when subjected to conventional classification using combination of deep residual network and testing. Manufacturing companies could not avoid doing bidirectional LSTM,” IEEE Access, vol. 7, no. 7, the tests and reporting the results, and those whose al- pp. 102119–102135, 2019. gorithms did not perform as well as their rivals spent their [8] S. K. Yadav, R. Sinha, and P. K. Bora, “Electrocardiogram development expenditures on focused attempts to en- signal denoising using non-local wavelet transform domain hance performance. filtering,” IET Signal Processing, vol. 9, no. 2, pp. 88–96, 2015. As a result of the availability of databases, the overall [9] H. D. Hesar and M. Mohebbi, “A multi rate marginalized level of performance of commercial arrhythmia detectors particle extended kalman filter for P and T wave segmentation has increased significantly in recent years. in ECG signals,” IEEE Journal of Biomedical and Health In- Although it would be incorrect to suggest that manu- formatics, vol. 23, no. 1, pp. 112–122, 2019. [10] S. Xu, J. Li, K. Liu, and L. Wu, “A parallel GRU recurrent facturers were capable of producing significantly better network model and its application to multi-channel time- products in the late 1970s and instead chose to add bells and varying signal classification,” IEEE Access, vol. 7, no. 8, whistles in response to their customers’ apparent lack of pp. 118739–118748, 2019. interest in performance, it is possible to infer this from the [11] S. K. Bashar, E. Ding, A. J. Walkey, D. D. McManus, and fact that manufacturers were capable of producing signifi- K. H. Chon, “Noise detection in electrocardiogram signals for cantly better products in the late 1970s. intensive care unit patients,” IEEE Access, vol. 7, no. 7, pp. 88357–88368, 2019. [12] H. Liu, D. Chen, and G. Sun, “Detection of fetal ECG R wave Data Availability from single-lead abdominal ECG using a combination of RR (e data that support the findings of this study are available time-series smoothing and template-matching approach,” IEEE Access, vol. 7, no. 5, pp. 66633–66643, 2019. upon request from the corresponding author. [13] S. Jain, V. Bajaj, and A. Kumar, “Riemann liouvelle fractional integral based empirical mode decomposition for ECG denoising,” IEEE Journal of Biomedical and Health Infor- Conflicts of Interest matics, vol. 22, no. 4, pp. 1133–1139, 2018. (e authors declare that they do not have any conflicts of [14] J. Wang, M. She, S. Nahavandi, and A. Kouzani, “Human interest regarding the publication of this paper. identification from ECG signals via sparse representation of local segments,” IEEE Signal Processing Letters, vol. 20, no. 10, pp. 937–940, 2013. Acknowledgments [15] F. Khosrow-Khavar, K. Tavakolian, A. Blaber, and C. Menon, “Automatic and robust delineation of the fiducial points of the (e authors extend their appreciation to the researchers seismocardiogram signal for noninvasive estimation of car- supporting project number (RSP-2021/384), King Saud diac time intervals,” IEEE Transactions on Biomedical Engi- University, Riyadh, Saudi Arabia. neering, vol. 64, no. 8, pp. 1701–1710, 2017. 16 Journal of Healthcare Engineering [16] F. Alonso-Atienza, E. Morgado, L. Fernandez-Martinez, and Engineering in Health and Medicine, vol. 59, no. 12, pp. 3348–3356, 2015. A. Garc´ıa-Alberola, J. L. Rojo-Alvarez, Detection of life- threatening arrhythmias using feature selection and support [31] C. Venkatesan, P. Karthigaikumar, A. Paul, S. Satheeskumaran, and R. Kumar, “ECG signal preprocessing vector machines,” IEEE Transactions on Biomedical Engi- neering, vol. 61, no. 3, pp. 832–840, 2014. and SVM classifier-based abnormality detection in remote healthcare applications,” IEEE Access, vol. 6, no. 1, [17] S. Raj and K. C. Ray, “ECG signal analysis using DCT-based pp. 9767–9773, 2018. DOST and PSO optimized SVM,” IEEE Transactions on In- [32] Y. Park, D. YunII, and S.-H. Kang, “Preprocessing method for strumentation and Measurement, vol. 66, no. 3, pp. 470–478, performance enhancement in CNN-based STEMI detection from 12-lead ECG,” IEEE Access, vol. 7, no. 7, [18] S. R. Moosavi, E. Nigussie, M. Levorato, S. Virtanen, and pp. 99964–99977, 2019. J. Isoaho, “Low-latency approach for secure ECG feature [33] W. Zhu, X. Chen, Y. Wang, and L. Wang, “Arrhythmia based cryptographic key generation,” IEEE Access, vol. 6, recognition and classification using ECG morphology and no. 2, pp. 428–442, 2018. segment feature analysis,” IEEE/ACM Transactions on [19] A. O. Bicen, N. Z. Gurel, A. Dorier, and O. T. Inan, “Improved Computational Biology and Bioinformatics, vol. 16, no. 2, pre-ejection period estimation from ballistocardiogram and pp. 131–138, 2019. electrocardiogram signals by fusing multiple timing interval [34] S. Kiranyaz, T. Ince, and M. Gabbouj, “Real-time patient- features,” IEEE Sensors Journal, vol. 17, no. 7, pp. 4172–4180, specific ECG classification by 1-d convolutional neural net- works,” IEEE Transactions on Biomedical Engineering, vol. 63, [20] U. Satija, B. Ramkumar, and M. S. Manikandan, “A new no. 3, pp. 664–675, 2016. automated signal quality-aware ECG beat classification [35] L. Han, X. Pu, and X. Chen, “Method of fetal electrocar- method for unsupervised ECG diagnosis environments,” diogram extraction based on ]-support vector regression,” IEEE Sensors Journal, vol. 19, no. 1, pp. 277–286, 2019. IET Signal Processing, vol. 9, no. 7, pp. 430–439, 2015. [21] Q. A. Rahman, L. G. Tereshchenko, M. Kongkatong, [36] A. Smruthy and M. Suchetha, “Real-time classification of T. Abraham, M. R. Abraham, and H. Shatkay, “Utilizing ECG- healthy and apnea subjects using ECG signals with variational based heartbeat classification for hypertrophic cardiomyop- mode decomposition,” IEEE Sensors Journal, vol. 17, no. 5, athy identification,” IEEE Transactions on Nanobioscience, pp. 3092–3099, 2017. vol. 14, no. 7, pp. 505–512, 2015. [22] R. G. Sutar and A. G. Kothari, “Intelligent electrocardiogram pattern classification and recognition using low-cost cardio- care system,” IET Science, Measurement & Technology, vol. 9, no. 5, pp. 134–143, 2015. [23] H. E. Saadawy, M. Tantawi, H. A. Shedeed, and M. F. Tolba, “Hybrid hierarchical method for electrocardiogram heartbeat classification,” IET Signal Processing, vol. 12, no. 5, pp. 506–513, 2018. [24] P. K. Shukla, M. Zakariah, W. A. Hatamleh, H. Tarazi, and B. Tiwari, “AI-DRIVEN novel approach for liver cancer screening and prediction using cascaded fully convolutional neural network,” Journal of Healthcare Engineering, vol. 2022, Article ID 4277436, 14 pages, 2022. [25] P. K. Shukla, V. Roy, P. K. Shukla et al., “An Advanced EEG Motion Artifacts Eradication Algorithm,” Ke Computer Journal, p. bxab170, 2021. [26] N. K. Rathore, N. K. Jain, P. K. Shukla, U. Rawat, and R. Dubey, “Image forgery detection using singular value decomposition with some attacks,” National Academy Science Letters, vol. 44, no. 4, pp. 331–338, 2021. [27] A. Kumar, M. Saini, N. Gupta et al., “Efficient stochastic model for operational availability optimization of cooling tower using metaheuristic algorithms,” IEEE Access, vol. 10, p. 1, 2022, https://ieeexplore.ieee.org/ stamp/stamp.jsp? tp�&arnumber�9682734&isnumber�6514899. [28] D. Singh, V. Kumar, M. Kaur, M. Y. Jabarulla, and H.-N. Lee, “Screening of COVID-19 suspected subjects using multi- crossover genetic algorithm based dense convolutional neural network,” IEEE Access, vol. 9, pp. 142566–142580, 2021. [29] H. Kaushik, D. Singh, M. Kaur, H. Alshazly, A. Zaguia, and H. Hamam, “Diabetic retinopathy diagnosis from fundus images using stacked generalization of deep models,” IEEE Access, vol. 9, pp. 108276–108292, 2021. [30] D. B. Saadi, G. Tanev, M. Flintrup et al., “Automatic real-time embedded QRS complex detection for a novel patch-type electrocardiogram recorder,” IEEE Journal of Translational

Journal

Journal of Healthcare EngineeringHindawi Publishing Corporation

Published: Apr 13, 2022

References