Adaptive Filtering and Artificial Intelligence Methods on Fetal Ecg Extraction

Journal of Critical Reviews

ISSN- 2394-5125 Vol 7, Issue 7, 2020

ADAPTIVE FILTERING AND ARTIFICIAL INTELLIGENCE METHODS ON FETAL ECG EXTRACTION

Dr. M. Pradeepa1,Dr. S. Kumaraperumal2

1Assistant Professor (Sr.), School of Information Technology, Vellore Institute of Technology, Vellore, India. Email: [email protected] 2Sr. Assistant Professor , Xavier Institute of Management & Entrepreneurship, Bangalore, India. Email: [email protected] Corresponding Author Email ID [email protected]

Received: 11.02.2020 Revised: 19.03.2020 Accepted: 23.04.2020

Abstract Above 30 percent of infant’s death occur due to heart problem like congenital heart disease during 2004 in United States of America. Every year, one in 125 infants is born with heart imperfection. To address these problems, early identification of cardiac anomalies and consistent monitoring of fetal heart can support Pediatric Cardiologist and Obstetrics to take necessary care on time to prescribe medicines and take precautionary measures during gestation period, delivery and/or after birth. Majority of cardiac abnormalities contain some symptoms in the cardiac electrical signal morphology. Electrocardiography gives more information in measuring cardiac signals compare to sonographic measurement. However, in non-invasive heartbeat recording by fetal Electrocardiogram (ECG) application Electrocardiography has its limitation due to low signal- noise ratio where impeding bio-signals are too stronger than fetal electrocardiogram signals. Various adaptive filtering and Artificial intelligence techniques are applied to solve this complex problem. The complex real world problems need a combination of knowledge, skills, and techniques from various sources as an intelligent system. That intelligent system should possess expertise of human, adjust itself to changing environment and learn to improve on its own.

Keywords: Artificial Intelligent, Adaptive Filter, Adaptive Neuro-Fuzzy Inference System (ANFIS), Kalman filtering (KF).

INTRODUCTION signals from more than one channels utilized to extract the fetal Fetus health condition is monitored by many methods where ECG components recorded at noisy environment. To extract Electrocardiography is one of the frequently used methods which desired signal many types of adaptive filters are used, in which shows the fetus heart’s electrical activities. some of them are presented below. Generally, an invasive or non-invasive method of recording of Time sequenced adaptive filtering has been recommended by Fetal ECG (FECG) is performed. In invasive method of recording, Ferrara & Widrow [4] for FECG enrichment. They identified the the electrode has to be placed on the scalp of the fetus to non-stationary fetal ECG signal having recurring statistical measure the ECG but the electrode has to be passed through characteristics. The Least Mean Square (LMS) adaptive filter can mother’s womb which creates difficulties to the mother [1] and able to follow up such fast changing non stationarities, hence an also possible only at the later stage of pregnancy period. The adaptive filter have to be designed with rapidly varying impulse non-invasive method of recording does not provide any trouble response to improve the performance of the extraction. The to the mother because the electrode has to be placed on mothers’ method uses many sets of hyper parameters to adapt for fast abdomen to measure the ECG of the fetus. changing impulse response. In order to adapt for fast changing impulse response, the method requires more abdominal signals There are several approaches proposed to record the fetal ECG and also timely identification of estimated fetal pulse. Apart from under non- invasive method which uses either a single lead or the above requirements the technique needs prior information of two leads or multiple leads. For a single lead method of fetal ECG positions. The time sequenced adaptive filtering recording, only one electrode is positioned on the mothers’ provides more accurate results compared to classical LMS abdomen, two lead systems uses two electrodes which have to be adaptive filter. The overall performance of the adaptive filter is positioned on the chest and abdomen and multiple lead systems increased when the number of channel input is increased. The require multiple electrodes to record the fetal ECG. main advantage of this approach is that the prior knowledge of There are several complications in non-invasive method of signals’ power spectrum is not required. But, the time sequenced recording fetal ECG, because the recording is not directly taken method need the estimation for the timely identification of the from the fetus which is measured on the abdomen, hence the pulse, to synchronize the filter regeneration and the fetal cardiac fetal ECG is to be extracted from signal contaminated by multiple cycles. They stated the future direction to enhance the results by sources of interferences. Apart from these sources of finding better method to locate the fetal pulse positions in order interferences the low signal level of fetal ECG [2] and the spectral to make this approach with recordings having lower SNR. overlapping of mother ECG and fetal ECG [3] makes the Kam & Cohen [5] identified a method to find the fetal ECG using extraction more critical. Infinite Impulse Response (IIR) filtering technique and Genetic Algorithm (GA). The hybrid IIR-GA approach on fetal ECG MATERIALS AND METHODS extraction, the adaptation rule is combined with GA, whenever Adaptive Filter based Methods the estimated gradient stuck with local extremum. Hybrid IIR-GA Generally, an adaptive filter has the ability of self-adjusting its provide best with simulation compared to FIR LMS based weight towards minimization of error. In this technique recorded method but with real data, the method fails to show the

Journal of critical reviews 295

ADAPTIVE FILTERING AND ARTIFICIAL INTELLIGENCE METHODS ON FETAL ECG EXTRACTION

significant difference between them. This may be because of the results than EKF. The performance evaluation was performed by body transfer function acts as a simple low pass filter so that a calculation accuracy, sensitivity, and positive predictive value. lower order FIR adaptive filter is sufficient, and the authors Single channel is used for the extraction which requires needs suggested further studies are required to analyze this few electronic components and suits for portable monitoring assumptions. system. Talha et al. [6] also presented similar approach of GA based Khamene & Negahdaripour [10] developed a method to extract Finite Impulse Response (FIR) filter for extraction where Genetic desired signal by wavelet transform using the modulus maxima algorithm is used as a optimizer for FIR filter and the results are which is in the wavelet domain and singularity detection, compared with the other approaches of adaptive filters like acquired from the abdomen signal. To differentiate the maternal wiener filter, Recursive Least Mean Square (RLMS) and NLMS and fetal ECG signals, abdominal signal’s modulus maxima filters. The NLMS approach provide better results in terms of locations are used. The authors projected two different reliability and speed of convergence but provide divergence approaches for implementation of the algorithm, in the first results when the adaptation is too large which have been approach, to carry out the classification minimum one thoracic overcome by the method of GA based FIR filter. GA with eight bits signal is used, but in the approach no thoracic signal is required. and ten iterations provide better quality compared to other A reconstruction algorithm is applied to extract the desired algorithms and an improvement may be provided by changing signal from the identified fetal modulus maxima. The developed the order of the filter. procedure varies from the traditional time domain approaches. In this method, the significant features of the signal provide high The adaptive filtering approach may be combined with other performance against other signal disturbances is studied. approaches to provide enhancement in extraction. Kholdi et al. [7] identified a GA based adaptive filter which uses LMS based Elloumi et al. [11] projected a fetal Electrocardiogram extraction adaptive filter for extracting fetal ECG, where the best filter through two stages Pitch Synchronous Wavelet Transform coefficients are calculated based on genetic algorithm which (PSWT). This implementation is based on a modeling idea which makes the adaptive filter response converge into global is capable of capturing the signal and its fluctuations by the use extremum. The random search nature of GA find the optimum of basic elements. In the two stage method, first the maternal filter coefficients even the structure of nonlinear transformation component estimation from a contaminated abdominal signal by is unknown or the structure may vary for different person. This a pitch synchronous decomposition is made. From the result of adaptive process does not need the knowledge in advance about first stage estimation, the next iteration is carried out to extract the signal and noise statistics, only assumes the signal is the required FECG signal. The pitch synchronous wavelet uncorrelated with noise. The input and output SNR are calculated decomposition permits to identify the FECG from a composite with different delay. Even, changes in the amplitude of the MECG abdomen signal by finding its spectral components and eliminate signal and rate of delay, the output SNR does not have much the further components within the harmonic bands. The results change, along with it provides fast response for extraction of demonstrate the accurateness of the technique particularly at FECG. low power fetal ECG component. Niknazar et al. [8] presented an extended Kalman filtering Bhoker & Gawande [12] also introduces wavelet based methods technique to extract FECG. The method utilizes single lead for the extraction. Wavelet transform is the effective method for recording and applied nonlinear Bayesian filtering framework detecting ECG signal characteristics in which abdominal ECG for extraction. This Kalman filtering (KF) framework considered signal is decomposed using Daubechie wavelet to estimate as a type of adaptive filter, and provides an impressive method maternal components. Later, the FECG signal is extracted by for noise elimination and quality extraction of FECG. Extended deducting MECG signal from AECG signal and fetal R peaks are Kalman Filter (EKF) with a backward recursive smoothing phase, calculated from extracted FECG signal for detecting heart rate. design the Extended Kalman smoother (EKS). The approach first This algorithm is implemented on 15 recorded signals and heart extracts the dominant ECG and considering other sources are rate is estimated. Gaussian signal. Once detecting the predominant component of maternal ECG in the actual signal, then the fetal ECG extraction Wu et al. [13] has ratified an integrated approach, which takes place from the lingering signal. The process is called as combines wavelet based approach, LMS based adaptive filtering sequential EKF and the extended state Kalman filtering linearizes and Spatially Selective Noise Filtration (SSNF). Stationary the mean and covariance is named as parallel EKF or EKS where Wavelet Transform (SWT) is applied on the recorded signal and the parallel EKS produce more accurate extraction compared to the coefficients are obtained and given as input, to the adaptive sequential EKS. This model effectively separates the FECG even if filter. The correlation between the recoded signals is evaluated the signal is embedded with mother components and permits to by the adaptive filter. Based on the threshold value, noises were identify R peaks. The sensitivity analysis is performed under eliminated using SSNF method from the results obtained by dissimilar cases; demonstrate the success of the algorithm. This adaptive filter. The method was examined by visual comparison approach is also tested with real abdominal recordings and twin and quantitative analysis where quantitative performance is magnetocardiograms. measured by SNR based on eigen value and cross correlation. The method provides improvements in identification of R peaks, Panigrahy et al. [9] projected an efficient method using single removal of noise components by SSNF algorithm and higher SNR channel abdominal ECG, in that PSO with EKS algorithms are values. applied. The abolition of MECG signal and other noise components in the abdomen ECG signal are essential to detect Sharma & Suji [14] proposed a method based on Weiner filter the fetal ECG signal. During pre-processing power line and adaptive LMS algorithm for denoising the ECG signals. The interference and baseline wander are cancelled from AECG signal performance of the filters is evaluated using signal to noise ratio after that, phase assignment is made by detecting the first R peak and power spectral density. The result shows that the Weiner of maternal component. PSO framework was used to compute filter produces good signal to noise ratio compared to LMS the optimal parameters of EKS. EKS template applied to find the adaptive filter but the power spectral density of both the filter MECG component AECG and then the MECG component is are same. deducted to extract the fetal ECG. EKS exploit future data to Satija et al. [15] presented a low complexity method to detect and provide enhanced approximation of the present state. Because of classify ECG noises such as baseline wander, muscle artifact, flat non-causal behavior, the EKS is projected to provide improved line and time varying noise. The method is implemented using

Journal of critical reviews 296

ADAPTIVE FILTERING AND ARTIFICIAL INTELLIGENCE METHODS ON FETAL ECG EXTRACTION

temporal features and the use of derivative and moving average Mean Square Difference (PRD) parameter which indicates the filters. similarities between the actual signal and the resultant signal is Artificial Intelligent based Methods used to evaluate the performance of the algorithm which is Assaleh [16] proposed a method to separate the FECG by higher than ANFIS trained with Gradient Descent (GD) method. utilizing a thoracic and an abdominal ECG recording. To map nonlinearly the signals of thoracic ECG and abdominal ECG, the Nasiri et al. [21] presented an ANFIS trained with PSO method polynomial networks technique is utilized. Then, FECG is that boosts up the extraction technique which decomposes the separated by deducting the aligned thoracic ECG from the electrocardiogram signal to its Intrinsic Mode Functions (IMFs) abdominal ECG signal. The qualitative nature of the FECG by Empirical Mode Decomposition algorithm for the removal of extracted by this approach exceeds the results obtained from noise components and the other IMFs recreate the other methods like ICA. electrocardiogram signal without any Baseline. The Particle Swarm Optimization is presented to train and fine tune the Assaleh [17] introduced a procedure based on ANFIS for fetal parameters of ANFIS to model the path in which maternal ECG extraction using two ECG signals measured at thoracic and electrocardiogram signal travel to reach abdominal region. The abdominal surface of the mother. The author identified that the system has the capability to discriminate and eliminate maternal maternal component in the composite abdominal ECG is a signal components from the recorded abdomen signal and thus a nonlinear transformation of the MECG and proved superior approximation of fetal electrocardiogram signal was experimentally. ANFIS network was used to find the nonlinear obtained. The algorithm was tested with simulated and real relationship, and also used to align the MECG signal with the signals. PRD is utilized to quantitatively evaluate the maternal component in the composite ECG signal. Then the FECG performance of the projected algorithm against other algorithms. component was found by deducting the estimated maternal PRD parameter for wavelet transformation is 1.1279, for ANFIS component from the abdominal ECG signal. The effectiveness of with GD is 0.5320 and for proposed approach is 0.4734, so the this approach is that it can extract even at very low signal to method demonstrates considerable improvement. The noise ratios and also able to extract the Fetal ECG even if it is performance is also evaluated using two SNR values such as completely overlapped within the maternal QRS complex. The SNRsvd and SNRcor for the Physiobank database signals, where the performance was compared with the Normalized Least Mean Independent Component Analysis (ICA) method shows some Square (NLMS) technique and the polynomial networks based improvement over the proposed approach, but one of the methods where the quality signal to noise ratio have higher positive aspects of the proposed approach against ICA based values even at low fetal to maternal SNR. algorithm is that only two recorded signals are needed which in turn minimizes the evaluation time and makes the mother Jia et al. [18] implemented an adaptive linear neural network comfortable during recording but the ICA based method requires which is trained based on the input and target values for multiple signals. The advantage of using less recorded signal is extraction. The error signal produced by the network is that requires less recording electrodes, less noise sources and considered as fetal ECG. Training rule adopts the network less cost. weights according to W-H learning rule. The results indicate that the algorithm can extract the FECG even when small amount of Azar et al. [22] and Inbarani et al. [23] are proposed a method for data are presented and the outcome is higher than the adaptive feature selection using unsupervised PSO based relative reduct filter technique which is ensured by visual comparison. technique. The approach is experimented with different classification algorithms and its performance is analyzed. The Swarnalatha & Prasad [19] provided a method for maternal ECG method reduces the number of features to evaluate the heart cancellation using ANFIS and wavelets with three different rate. approaches namely ANFIS alone, wavelets preprocessing followed by ANFIS and then ANFIS followed by wavelet post Ma et al. [24] ratified a hybrid nonlinear Adaptive Noise processing. In wavelet processing, either preprocessing or post Canceller (ANC) uses single or multiple reference signals to processing technique involves decomposition and reconstruction perform the task of extracting fetal electrocardiogram. A Volterra process were done using coiflets since it decreases the noise filter without 2nd- and higher-order exponential terms and a components and afford high resolution output and also it is Functional Link Artificial Neural Network (FLANN) possessing related shape of fetal ECG. The method proposed by them is only exponential term are equipped along with every reference evaluated using SNR, correlation coefficient and other channels to estimate both the linear and nonlinear maternal performance indices. ANFIS followed by wavelet post processing component and its vague version within the abdominal ECG. The approach produce best extraction compared to other two FECG signals obtained by this approach provide best visual approaches and also advantage of ANFIS based method is a less quality as compared with other ANC schemes but still few noise computational analysis because of the qualitative aspects of elements are present in the extracted FECG. The author artificial intelligence. suggested more powerful nonlinear filters are designed to improve the extraction accuracy and introducing some post- Sargolzaei et al. [20] applied ANFIS network trained with Particle denoising technique to enhance the result. Swarm Optimization (PSO) method for FECG extraction from the two recorded ECG signals. The non-linear transformation of the Ayat et al. [25] presented an extraction method using single maternal component in abdomen ECG against the thoracic ECG is maternal abdominal measurement which combines savitzky- identified using ANFIS algorithm. The training algorithm golay filter and polynomial networks. First the estimation of the contributes a primary role in identification and affects the MECG by executing the abdominal signal via a savitzky-golay competence of the ANFIS, hence they applied PSO algorithm act filters. Then, using polynomial network, the estimated as a training procedure to tune ANFIS model. The application of component is aligned to extract the fetal ECG signal. Application PSO produces good convergence and also provides less of the algorithm on synthetic and real data exhibits the proposed computational difficulties when compared with other training method is efficient than multichannel based methods. procedures. After finding the non-linear transformation, the method estimates the maternal component in the composite Lei et al. [26] presented a deep learning feature representation abdominal ECG signal. Later, the FECG signal is extracted by for ECG identification. In this approach, a deep fusion features deducting the estimated maternal signal from the abdominal ECG are extracted through convolutional neural network. Then, the signal. The main advantage of this approach is that it can extract neural network and support vector machine are used to Fetal ECG signal even in early gestational period. Percent Root- categorize the ECG signals. Casas et al. [27] projected three machine learning procedure to identify the Premature

Journal of critical reviews 297

ADAPTIVE FILTERING AND ARTIFICIAL INTELLIGENCE METHODS ON FETAL ECG EXTRACTION

Ventricular contractions from the ECG signals. Neural network, Signal Processing. Proceedings. ICASSP99 (Cat. No. logistic regression and support vector machine are applied for 99CH36258) (Vol. 4, pp. 2335-2338). IEEE. classification of arrhythmias. 6. Talha, M., Guettouche, M. A., & Bousbia-Salah, A. (2010, November). Combination of a FIR filter with a genetic CONCLUSION algorithm for the extraction of a fetal ECG. In 2010 Signal processing importance is increasing in many applications Conference Record of the Forty Fourth Asilomar Conference that includes biomedical field, in which the enrichment of quality on Signals, Systems and Computers (pp. 1756-1759). IEEE. in application of signal processing is prime apprehension. Since 7. Kholdi, E., Bigdeli, N., & Afshar, K. (2011). A New GA-Based human body signals are measured through biomedical Adaptive filter for fetal ECG extraction. World Academy of instruments which are contaminated by signal recording Science, Engineering and Technology, 54. procedure and/or emerge from the environment where the 8. Niknazar, M., Rivet, B., & Jutten, C. (2012). Fetal ECG signals are recorded. The major contamination sources are extraction by extended state Kalman filtering based on electrical power interference, random interference of electronic single-channel recordings. IEEE Transactions on Biomedical noise, recording instrument noise and baseline wandering Engineering, 60(5), 1345-1352. interference. So, the extraction’s precision and enhancement of 9. Panigrahy, D., Rakshit, M., & Sahu, P. K. (2015, May). An signal through signal processing algorithm is prime disquiet in efficient method for fetal ECG extraction from single treatments, since clinical treatment by the physicians are made channel abdominal ECG. In 2015 international conference based on the results obtained from the signal processing. The on industrial instrumentation and control (ICIC) (pp. 1083- determination of an algorithm in the biomedical application 1088). IEEE. varies based on extraction features of signals such as positioning 10. Khamene, A., & Negahdaripour, S. (2000). A new method for of electrodes, recording techniques, instruments physiological the extraction of fetal ECG from the composite abdominal aspects and other procedural differences. Noise spectrum signal. IEEE Transactions on Biomedical Engineering, 47(4), imbricates with the signal and increases the complexity of signal 507-516. extraction. Signal extraction need higher care in algorithm 11. Elloumi, A., Lachiri, Z., & Ellouze, N. (2004, March). Pitch implementation where actual characteristics of signal of clinical synchronous wavelet based fetal ECG extraction. In First interest should not be compromised during extraction. International Symposium on Control, Communications and Signal Processing, 2004. (pp. 239-242). IEEE. The time-sequenced adaptive technique needs information about 12. Bhoker, R., & Gawande, J. P. (2013). Fetal ECG extraction the location of the fetal ECG [4]. The adaptive filter may produce using wavelet transform. ITSI Transactions on Electrical divergence results when the adaptation is too large and and Electronics Engineering, 1(4), 19-22. improvement may be provided by changing the order of the filter 13. Wu, S., Shen, Y., Zhou, Z., Lin, L., Zeng, Y., & Gao, X. (2013). [6]. When the adaptive filter estimates the gradient which may Research of fetal ECG extraction using wavelet analysis and stuck with local extremum [5]. When the spectrum of mother and adaptive filtering. Computers in biology and fetal ECG overlapped, the method fails to extract and also when medicine, 43(10), 1622-1627. the noise amplitude is greater than fetus then the error signal of 14. Sharma, B., & Suji, R. J. (2016, May). ECG denoising using the adaptive filter contains the noise instead of fetal ECG [28]. weiner filter and adaptive least mean square algorithm. The ANN shows guaranteed performance using two channel In 2016 IEEE International Conference on Recent Trends in recordings, but may produce number of minima and bear the Electronics, Information & Communication Technology over fitting risk [29]. (RTEICT) (pp. 53-57). IEEE. To address the critical issues in fetal care different studies 15. Satija, U., Ramkumar, B., & Manikandan, M. S. (2016, June). utilized different techniques and achieved reasonable success Low-complexity detection and classification of ECG noises with some constrains. Recently, the complicated real world for automated ECG analysis system. In 2016 International issues are solved through intelligent learning systems that Conference on Signal Processing and Communications integrate elements like knowledge, techniques and process from (SPCOM) (pp. 1-5). IEEE. multiple sources. 16. Assaleh, K., & Al-Nashash, H. (2005). A novel technique for the extraction of fetal ECG using polynomial networks. IEEE Source of Funding Transactions on Biomedical Engineering, 52(6), 1148-1152. The Authors declares that there is no source of funding. 17. Assaleh, K. (2006). Extraction of fetal electrocardiogram using adaptive neuro-fuzzy inference systems. IEEE Conflict of Interest Transactions on Biomedical Engineering, 54(1), 59-68. The Authors declares that there is no conflict of interest. 18. Jia, W., Yang, C., Zhong, G., Zhou, M., & Wu, S. (2010, October). Fetal ECG extraction based on adaptive linear REFERENCES neural network. In 2010 3rd International Conference on 1. Sargolzaei, S., Faez, K., & Sargolzaei, A. (2008, May). Signal Biomedical Engineering and Informatics (Vol. 2, pp. 899- processing based for fetal electrocardiogram extraction. 902). IEEE. In 2008 International Conference on BioMedical 19. Swarnalatha, R., & Prasad, D. V. (2010). Maternal ECG Engineering and Informatics (Vol. 2, pp. 492-496). IEEE. cancellation in abdominal signal using ANFIS and 2. Kimura, Y., Sato, N., Sugawara, J., Velayo, C., Hoshiai, T., wavelets. Applied Science, 10, 868-77. Nagase, S., ... & Yaegashi, N. (2012). Recent advances in fetal 20. Sargolzaei, A., Faez, K., & Sargolzaei, S. (2011, May). A new electrocardiography. The Open Medical Devices method for Foetal Electrocardiogram extraction using Journal, 4(1). Adaptive Nero-Fuzzy Interference System trained with PSO 3. Fanaswala, M. (2005). Fetal Electrocardiogram Extraction: algorithm. In 2011 IEEE INTERNATIONAL CONFERENCE A Case-Study in Non-linear System Identification. ON ELECTRO/INFORMATION TECHNOLOGY (pp. 1-5). IEEE. 4. Ferrara, E. R., & Widraw, B. (1982). Fetal electrocardiogram 21. Nasiri, M. (2012). Fetal electrocardiogram signal extraction enhancement by time-sequenced adaptive filtering. IEEE by ANFIS trained with PSO method. International Journal of Transactions on Biomedical Engineering, (6), 458-460. Electrical and Computer Engineering, 2(2), 247. 5. Kam, A., & Cohen, A. (1999, March). Detection of fetal ECG 22. Azar, A. T., Banu, P. N., & Inbarani, H. H. (2013, August). with IIR adaptive filtering and genetic algorithms. In 1999 PSORR-An unsupervised feature selection technique for IEEE International Conference on Acoustics, Speech, and fetal heart rate. In 2013 5th International Conference on

Journal of critical reviews 298

ADAPTIVE FILTERING AND ARTIFICIAL INTELLIGENCE METHODS ON FETAL ECG EXTRACTION

Modelling, Identification and Control (ICMIC) (pp. 60-65). beats. In 2016 Global Medical Engineering Physics IEEE. Exchanges/Pan American Health Care Exchanges 23. Inbarani, H. H., Banu, P. N., & Azar, A. T. (2014). Feature (GMEPE/PAHCE) (pp. 1-6). IEEE. selection using swarm-based relative reduct technique for 28. Jafari, F., Tinati, M. A., & Mozaffari, B. (2010, May). A new fetal heart rate. Neural Computing and Applications, 25(3- fetal ECG extraction method using its skewness value which 4), 793-806. lies in specific range. In 2010 18th Iranian Conference on 24. Ma, Y., Xiao, Y., Wei, G., Sun, J., & Wei, H. (2015, December). Electrical Engineering (pp. 30-34). IEEE. A hybrid nonlinear adaptive noise canceller for fetal ECG 29. Han, L., Pu, X. J., & Chen, X. J. (2015). Method of fetal extraction. In 2015 Asia-Pacific Signal and Information electrocardiogram extraction based on ν-support vector Processing Association Annual Summit and Conference regression. IET Signal Processing, 9(5), 430-439. (APSIPA) (pp. 811-814). IEEE. 25. Ayat, M., Assaleh, K., & Al-Nashash, H. (2015, February). Extracting fetal ECG from a single maternal abdominal record. In 2015 IEEE 8th GCC Conference & Exhibition (pp.

1-4). IEEE.

26. Lei, X., Zhang, Y., & Lu, Z. (2016, October). Deep learning

feature representation for electrocardiogram identification.

In 2016 IEEE International Conference on Digital Signal

Processing (DSP) (pp. 11-14). IEEE.

27. Casas, M. M., Avitia, R. L., Reyna, M. A., & Cárdenas, A. (2016,

April). Evaluation of three machine learning algorithms as

classifiers of premature ventricular contractions on ECG

Journal of critical reviews 299