CHAPTER 5 CANCELLATION OF MECG SIGNAL IN FECG EXTRACTION

Transcription

1 84 CHAPTER 5 CANCELLATION OF MECG SIGNAL IN FECG EXTRACTION 5.1 INTRODUCTION The analysis of the fetal heart rate (FHR) has become a routine procedure for the evaluation of the well-being of the fetus. Factors affecting FHR are uterine contraction, baseline variability, hypoxia and oxygenation. This method has many drawbacks such as position-sensitivity, signal drop out, frequent confusion between maternal heart rate and FHR, failure in obese patients (which in turn increases the rate of cesarean), and misinterpretation of cardiotocogram traces and failure to act in time. The use of FECG for monitoring the fetus overcomes all these limitations (Zarzoso et al 2001). However, FECG is mixed with interferences. In this chapter a brief introduction about fetal monitoring techniques, applications of FECG, and various interferences in FECG are discussed. Experimental results of interference (MECG) cancellation using BPN, CCN, ANFIS, and ANFIS- FCM are also presented. 5.2 FETAL MONITORING TECHNIQUES The most popular techniques available for noninvasive antepartum fetal monitoring systems are fetal phonography, ultrasonography and antepartum FECG.

2 Phonography and Phonocardiography Fetal phonography is the transcription of fetal sounds. Phonocardiography specifically involves the transcription of fetal heart sounds. Both cases are achieved by sensing fetal vibrations incident on the maternal abdomen. The clinical results obtained using phonocardiography for fetal monitoring are poor. The lack of success of phonographic monitoring systems is attributed to inappropriate transducer design; typically the transducers used for fetal heart monitoring are variants of those used for adults. The resulting signal has a poor SNR. This requires heavy filtering, which in turn leads to attenuation of potentially valuable signal information Ultrasonography Fetal surveillance has relied heavily on ultrasound imaging since the mid 1970s. Ultrasonic images are formed by the selective reflection of acoustic energy in soft biological tissue. In medical imaging, ultrasound in the frequency range of 2 to 20 MHz is coupled to the body by means of a piezoelectric transducer. The available systems for fetal monitoring are divided into those providing one and two dimensional image data. The one dimensional system employs a narrow ultrasound beam, which is used to illuminate specific fetal structures. The resulting ultrasound reflections are detected and the motion of the reflecting structure is quantified by using the Doppler principle. Doppler ultrasound is routinely used to visualize the motion of fetal structures such as heart valves. It is used as the basis for estimating heart rate. Portable (bedside) Doppler ultrasound systems to monitor the FHR, e.g. the cardiotocograph, are in common use. However there are objections to their routine use in long term fetal monitoring. Firstly, there is some concern amongst clinicians on the

3 86 safety to the fetus of prolonged exposure to ultrasound radiation. Secondly, the naturally occurring fetal movements must be tracked by the ultrasound beam. This requires close operator supervision because gross fetal movements are often sporadic. The two dimensional systems generate an array of ultrasound beams. The beam array can be swept or scanned across the fetus and a two dimensional image is formed using the detected ultrasound reflections. This is referred as B-mode (brightness mode) ultrasonography. The systems providing two dimensions of image data enable the routine characterization of many physical fetal activities. However, their high capital and running costs limits their use to short term monitoring in a clinical environment Fetal Electrocardiography The FECG signal yields information about the condition of the child during pregnancy. It is recorded non-invasively from the belly of the pregnant woman. But it is mixed with noise like the MECG, respiration baseline wander, power line interference etc. The low fetal SNR makes it impossible to analyze the FECG. Attenuating the noise by classical filtering techniques is not satisfactory due to an overlap in spectral content with the FECG. The characteristics of each noise signal determine which signal processing technique should be applied in order to achieve its removal. 5.3 APPLICATIONS OF FECG FECG is used to determine several abnormalities of fetus such as fetal asphyxia, congenital heart diseases (CHD), arrhythmias, etc. Fetal asphyxia results from insufficient uterine blood flow and decreased maternal arterial oxygen content. The diagnoses of asphyxia based on FHR gives an

4 87 inaccurate result because there is no precise relationship that exists between asphyxia and FHR. CHD refers to a problem with the heart's structure and function due to abnormal heart development before birth. Cardiac arrhythmia is a group of conditions in which the muscle contraction of the heart is irregular or is faster or slower than normal. FECG is also used to diagnose fetal position, twins and fetal well-being. 5.4 MEASUREMENT OF FECG The diagnostic tests of fetal well-being are categorized as invasive and noninvasive. During delivery, accurate recordings can be made by placing an electrode on the fetal scalp. However, as long as the membranes protecting the child are not broken, one should look for noninvasive techniques. Further, the use of noninvasive techniques enables the monitoring of well-being of fetus from the early period of pregnancy onwards. The amplitude of FECG increases during the first 25 weeks, mini towards the 32 nd week and increases again afterwards. Some of the problems that limit the extraction of FECG using noninvasive technique are the presence of background noise leading to poor SNR, no standard electrode positioning for optimizing acquisition, and the shape of the signal that depends on the position of the electrodes and the gestational age. Moreover, electrode output from the abdomen is hampered by many artifacts. Therefore, it is required to remove these artifacts from FECG. The types of artifacts are discussed in detail in section ARTIFACTS IN FECG SIGNAL The artifacts in FECG signal are power line interference, electrode contact noise, motion artifact, muscle contraction, baseline drift, instrumentation noise generated by electronic devices, MECG, and EMG due to uterus contraction and respiration.

5 Power Line Interference The cables carrying ECG signals from the examination room to the monitoring equipment are susceptible to electromagnetic interference of power line frequency. It consists of Hz pick up and harmonics, which are modeled as sinusoids. The amplitude varies up to 50 percent of the peakto-peak ECG amplitude. Many researchers have done work to cancel this interference from the ECG signal. Specifically, adaptive noise cancellation addresses this problem Electrode Contact Noise It is a transient interference caused by loss of contact between the electrode and the skin that effectively disconnects the measurement system from the subject. The loss of contact can be permanent, or can be intermittent as would be the case when a loose electrode is brought in and out of contact with the skin as a result of movements and vibration. This switching action at the measurement system input can result in large artifacts since the ECG signal is usually capacitive coupled to the system. It can be modeled as randomly occurring rapid baseline transition, which decays exponentially to the baseline value and has a superimposed 50 Hz component Motion Artifact Motion artifact is transient baseline changes caused by changes in the electrode-skin impedance with electrode motion. As this impedance changes, the ECG amplifier sees different source impedance, which forms a voltage divider with the amplifier input impedance. Therefore, the amplifier input voltage depends upon the source impedance, which changes as the electrode position changes. The usual cause of motion artifact is assumed to

6 89 be vibrations or movements of the patient. This type of interference represents an abrupt shift in baseline due to movement of the patient while the ECG is being recorded Electromyogram The muscle contractions generate millivolt level potentials. EMG noise has a frequency range of Hz. EMG is measured from within the muscle or from the skin surface overlying the muscle. It has a spatial and temporal interference pattern of the electrical activity of the activated motor units located near the detection surfaces Baseline Drift The baseline drift is low frequency interference in ECG signal caused due to respiration. It is nothing but change in the d.c component of the ECG signal. The drift of the base line with respiration is represented by a sinusoidal component at the frequency of respiration (less than 0.5 Hz) added to the ECG signal. The amplitude and the frequency of the sinusoidal component are variables. The variations can be represented by amplitude modulation of the ECG by the sinusoidal component added to the baseline Maternal Electrocardiogram The main source of interference in FECG is MECG. It is a quasiperiodic signal with a fundamental frequency of about 1 Hz. The fundamental frequency of the FECG is about twice as that of MECG. The amplitude of MECG is much higher (5 to 100 times) than that of the FECG. Hence, it is difficult to recognize the FECG unless the MECG is cancelled. The main aim of this work is therefore to eliminate the MECG.

7 CANCELLATION OF MECG BY AIC The mother and fetal hearts behave independently, which means that both are uncorrelated. MECG is mainly recorded on the mother s chest. The effect of FECG on MECG is negligible due to the large distance between the small fetal heart and the mother s chest electrodes, also the magnitude of FECG is less compared to that of MECG. But, when FECG is measured at the abdomen of the mother, MECG is a significant interference. The problem in hand deals with the suppression of MECG component in FECG. The block diagram of AIC for FECG extraction is shown in Figure 5.1. MECG FECG n(k) x(k) Nonlinear passage d(k) ECG recorded from Mother s Abdomen + + Σ Measured signal y(k) + Σ - d ˆ( k ) Estimated FECG xˆ( k) x( k) d( k) dˆ( k) x( k) e( k) x( k), as e( k) 0 ECG recorded from mother s chest as the reference Signal Adaptation Techniques Estimated Figure 5.1 Block diagram of AIC for FECG extraction In Figure 5.1, the signal measured from mother s abdomen y(k) is inevitably noisy due to the mother s heartbeat signal n (k), which is measured clearly via sensor at the thoracic region. However, n (k) does not appear directly in y (k). Instead, it travels through the nonlinear passage (mother s

8 91 body) and appears as delayed and distorted version of n (k) called as interference i.e. d (k). A clean version of the noise signal (MECG) that is independent of the required FECG signal is considered in order to estimate the interference. It acts as a reference signal for the adaptation. As long as x(k) is not correlated with n (k), y (k) can be used as the desired output for training. In this report BPN, CCN, ANFIS and ANFIS-FCM are used as the adaptation techniques to estimate the MECG interference in the measured signal y (k). When the estimated interference is very close to the actual interference in y (k), these two get cancelled and the required FECG signal is obtained as the output. The implementation phases for AIC in real biosignals are given in the form of a flowchart in Figure 5.2. Noise Signal Delayed noise signal Adaptation Techniques Measured signal Estimated interference - Σ + Estimated required signal Figure 5.2 Flowchart for implementation of AIC in real biosignals In FECG extraction, proposed techniques take the MECG as the reference signal and the measured signal (abdominal signal), y(k) as the target

9 92 signal and try to estimate the MECG present in the measured signal. Once the designated epoch is reached or the goal (minimum value of x ˆ( k) ) is reached (whichever is earlier), it stops training and gives the estimated interference d ˆ( k ). Then FECG is extracted by simply subtracting d ˆ ( k) from y (k). The flowchart for finding the noise in the estimated signal is shown in Figure 5.3. The noise is obtained by passing the estimated signal (FECG) through a Butterworth filter. The cut off frequency is selected based on the frequency component of the estimated signal. Estimated Signal obtained from AIC Butterworth Filter - + Σ Noise Figure 5.3 Flowchart for finding the noise in the estimated signal The signals recorded at a sampling frequency of 500 Hz for about 5 seconds from 8 electrodes located on a pregnant woman s body (5 electrodes on the abdomen and 3 electrodes on the chest) are considered to extract the FECG. These signals are taken from the website mentioned by Jafari et al (2005).

10 93 The real cutaneous electrode recordings for a duration of 2 seconds (1000 samples) are plotted in Figures 5.4 and 5.5. Figure 5.4 shows the abdominal signals recorded in 5 locations. They have both MECG and FECG along with some high frequency noise. Figure 5.5 shows the signals recorded in the mother s thoracic region (MECG). Due to the longer distance between the thorax electrodes and the fetal heart, no FECG heartbeat component can be perceived in Figure 5.5. The abdominal signal, abd1 and thoracic signal (MECG), thr3 are used in this work for AIC as these signals posses rich information. Figure 5.4 Abdominal signals from different electrode positions

11 94 Figure 5.5 MECG signals recorded at the thorax region of a pregnant woman 5.7 IMPLEMENTATION OF ADAPTATION TECHNIQUES FOR AIC in this section. The implementation of proposed adaptation techniques is explained Implementation of BPN The steps used are: The software used for the implementation of BPN is MATLAB. 1. Specifying the inputs and targets to the BPN.

12 95 2. Giving the minimum and maximum values of the input ranges. 3. Specifying the number of layers and the number of units in each layer. 4. Mentioning the activation functions for each layer. TANSIG is a hyperbolic tangent sigmoid transfer function which is used as the activation function for the hidden layer to calculate a layer's output from its net input PURELIN is a linear transfer function which is used as the activation function for the output layer 5. Creating a feed forward back propagation network model by using the command called NEWFF. 6. Simulating the network for plotting the network output. 7. Specifying the training parameters like learning rate momentum performance goal number of epochs etc. 8. Training the network using the training function called TRAINLM. 9. Giving the conditions to stop training the network The maximum number of EPOCHS (repetitions) is reached or The maximum amount of TIME has been exceeded or Performance has been minimized to the GOAL or The performance gradient falls below MINGRAD.

13 Implementation of CCN The software used for the implementation of CCN is same as that of BPN. The CCN procedure includes the following steps. 1. Specifying the inputs and targets to the CCN. 2. Giving the minimum and maximum values of the input range 3. Mentioning the number of neurons randomly in different layers 4. Forming a rough architecture with the randomly specified neurons using NEWCF. 5. Refining the architecture by finding the covariance between the input and the rough architecture output. 6. Obtaining the final architecture by including the maximum covariance unit. 7. Remaining steps are same as that of BPN Implementation of ANFIS The basic steps used in the computation of ANFIS are given below: 1. Specifying the inputs and targets 2. Generating an initial Sugeno type FIS system using the MATLAB command GENFIS1. It goes over the data in a crude way and finds a good starting system. 3. Giving the parameters like number of iterations (epochs), tolerance error etc for learning.

14 97 4. Leaning process starts using the command ANFIS and stops when goal is achieved or the epoch is completed, whichever is earlier. 5. The EVALFIS command is used to determine the output of the FIS system for given input Implementation of ANFIS-FCM ANFIS-FCM uses the implementation steps similar to that of ANFIS except the data clustering operation. The steps for ANFIS-FCM include: 1. Initializing the membership matrix with random values between 0 and 1 2. Calculating the fuzzy cluster centers 3. Computing the cost function (or objection function) 4. Stopping the training if either it is below a certain tolerance value or its improvement over previous iteration is below a certain threshold. 5. Compute a new membership matrix 5.8 EXPERIMENTAL RESULTS Experiments are carried out to cancel the MECG in FECG using different AI techniques. The techniques along with their results are discussed in this section.

15 Results with BPN The BPN used for MECG cancellation consists of an input layer with two neurons, hidden layer with 35 neurons and an output layer with single neuron. Its structure is shown in Figure 5.6 (a), where IW {1, 1} = Initial weights connecting the layer inputs to the hidden layer, LW {2, 1} = Layer weights connecting the output of hidden layer to the inputs of output layer b {1} = Layer 1 bias values b {2} = Layer 2 bias values. Figure 5.6(b) shows the flow diagram from the input to the output of the network, where p {1} represents the input layer, a{1} is the hidden layer, a{2} is the output layer and y{1}is the output. Figure 5.6(c) gives the details between p {1} and a{1}. The details of weights between p{1} and a{1} are shown in Figure 5.6(d). It shows only 5 neurons weight out of 35 neurons in the hidden layer for simplicity. The connections between the hidden and output layer are shown in Figure 5.6(e). The weights between a{1} and a{2} are shown in Figure 5.6(f). (a) BPN structure with single input layer, single hidden layer and an output layer

16 99 (b) Flow diagram from the input to the output of the network (c) Connection between the input layer and the hidden layer (d) Weights between the input layer and the hidden layer

17 100 (e) Connection between the hidden layer and the output layer (f) Weights between the hidden layer and the output layer Figure 5.6 BPN structure for FECG extraction The parameters used for training BPN are epochs = 10, learning rate = 0.5, parameter goal set =0.65, minimum time to train in seconds = infinity and momentum =0.9. The performance criteria used in BPN is Mean Square Error. The training result is shown in Figure 5.7 which gives the relationship between the epochs and the MSE. Training, Goal Figure 5.7 Training results of BPN

18 101 The results of AIC using BPN are shown in Figure 5.8. Only 350 samples of data are shown in this Figure for clarity purpose. In Figure 5.8 (a) is the abdominal signal (which contains both MECG and FECG), (b) is MECG alone (c) is the estimated thoracic signal determined by BPN, (d) is the estimated FECG after cancellation and (e) is the noise present in the estimated FECG after AIC. Figure 5.8 Results with BPN (a) Abdominal signal (b) MECG (c) Estimated MECG in Abdominal signal (d) Estimated FECG (e) Noise after AIC In this work, a Butterworth filter of order 5 and normalized frequency of 0.7 is considered. Arrow in Figure 5.8 (d) shows the presence of MECG in estimated FECG even after AIC. Based on this signal, it is difficult

19 102 to differentiate the FECG component from the MECG. It is, hence, observed that the cancellation of MECG is not perfect when BPN is used Results with CCN Another AI technique namely CCN is used for AIC. The structure of CCN for FECG extraction is same as that of BPN except the way in which the weights are adjusted. The results of AIC with CCN are shown in Figure 5.9. They give all the information which are same as in Figure 5.8. In Figure 5.9 (d), arrow indicates the presence MECG component which is less than that in the BPN method. But, significant noise is present between 50 and 100 samples Figure 5.9 Results with CCN (a) Abdominal signal (b) MECG (c) Estimated MECG in Abdominal signal (d) Estimated FECG (e) Noise after AIC

20 Results with ANFIS Matlab version 7.5 is used for software implementation of ANFIS. Since no idea is available about the initial membership functions, the command called genfis1 is used to examine the training data set and generate a single output Sugeno type FIS that is used as the starting point for ANFIS training. Fuzzy model with 2 inputs and one output generated by this command is shown in Figure 5.10, where delayed thoracic and thoracic represent the inputs to the fuzzy model. Each input contains 3 MFs. Infismat represents the system name and has 9 fuzzy rules. Estimated thoracic represents the system output. Since the Sugeno type FIS is used in this application, defuzzification is not required at the output. Figure 5.10 Fuzzy model generated by GENFIS After generating the fuzzy model, ANFIS requires a good number of epochs, training pair, and MFs for training. The function used for training is anfis, and generalized bell shape MF (gbellmf) is used for ANFIS training. The structure of ANFIS used for the extraction of FECG is shown in Figure Two nodes are present in the input layer and the inputs are MECG and the delayed MECG. Fuzzification is done by layer 1 (inputmf) which allocates 3 MFs to each input. Totally 9 rules are used in layer 2 (rule). Normalization layer (layer 3) is not included in this architecture. Layer 4 is the defuzzification layer (outmf). Layer 5 performs summation operation. After

21 104 training, the estimated MECG is obtained using the command evalfis. The results obtained through ANFIS are shown in Figure Figure 5.11 ANFIS structure Figure 5.12 Results of AIC in FECG with ANFIS (a) Abdominal signal (b) MECG (c) Estimated MECG in Abdominal signal (d) Estimated FECG (e) Noise after AIC

22 105 It may be observed that the magnitude of noise that is present between 50 and 100 samples in Figure 5.12 (d) is very much less than that in Figure 5.9 (d). By comparing the Figures 5.8 (d), 5.9 (d) and 5.12 (d), it is inferred that ANFIS gives better cancellation of MECG without degrading the FECG. Three more cases are considered to demonstrate the power of ANFIS in FECG extraction. Figure 5.13 (a) shows a case in which the measured signal contains 3 non-overlapping FECG beats and 2 MECG beats. Figure 5.13 (b) shows the output of ANFIS which is the estimated MECG present in the abdominal signal and Figure 5.13 (c) shows the estimated FECG. Figure 5.13 ANFIS output (a) Abdominal signal with non-overlapping FECG beats and MECG beats (b) Estimated MECG (c) Estimated FECG Figure 5.14 (a) shows the second case where the abdominal signal consists of partially overlapping FECG beats and MECG beats. Figures 5.14 (b) and (c) show the MECG and estimated MECG in abdominal signal respectively. Figure 5.14 (d) shows the estimated FECG.

23 106 Figure 5.15 (a) shows the third case (same as Figure 5.12) in which the abdominal signal has full overlap between the first FECG and the MECG beats. This represents the extreme case where the FECG is completely masked by the MECG component to the extent that the FECG beat is no longer visually distinguishable. The three arrows in Figure 5.15 (a) indicate the location of FECG signal in the abdominal signal. The three arrows in Figure 5.15 (c) indicate the location of the estimated FECG signal. Figure 5.15 (c) shows that the ANFIS technique is successful in extracting the FECG signal in this case also. However, the extracted FECG in the overlapping region is slightly distorted compared to the FECG in other locations. Figure 5.14 ANFIS output (a) Abdominal signal containing partially overlapping FECG beat and two MECG beats (b) MECG (c) Estimated MECG (d) Estimated FECG

24 107 The 2 second signals are divided into different frames in order to explain the effectiveness of the proposed technique for the extraction of FECG. Practically, if the signals are divided into many frames, then there is a possibility of losing some important data. It also increases the processing time. But ANFIS can process the entire data without frames which in turn decreases the processing time. In order to find out the SNR of the estimated FECG signal Butterworth filter is used to separate the amount of noise present in the estimated FECG. (a) (b) (c) Figure 5.15 ANFIS output (a) Abdominal signal containing full overlap between the first FECG beat and the MECG (b) Estimated MECG (c) Estimated FECG Results with ANFIS-FCM Though ANFIS yields better performance than BPN and CCN, the use of Fuzzy C Means clustering method is also investigated to explore the possibility of reducing mean square value of the estimated FECG signal and convergence time and increasing SNR.

25 108 In this application of ANFIS-FCM, the ANFIS structure is chosen as explained in section Four clusters are considered in order to cluster the input data. Based on the cluster centroid and the nature of the signal characteristics, the number of samples in each cluster varies. It is shown in Figure Cluster diagram varies from time to time depending on the selection of the centroid. Clustering takes the advantage of less data searching, fast convergence time and less mean square value of the estimated FECG. The results of AIC using ANFIS-FCM are presented in Figure Centroid values obtained for 2 inputs and a target data are given below. Delayed Thoracic signal = [ ] Thoracic Signal = [ ] Target = [ ] Figure 5.16 Four clusters of ANFIS-FCM

26 109 Figure 5.17 ANFIS-FCM output (a) Abdominal signal (b) MECG (c) Estimated MECG in Abdominal signal (d) Estimated FECG (e) Noise after AIC If the Figures 5.15(c) and 5.17(d) are compared, visually it is difficult to make any difference. But the quantitative comparison in terms of performance criteria shown in Table 5.1 clearly indicates the improved performance of ANFIS-FCM over ANFIS. 5.9 PERFORMANCE COMPARISON In order to make a comparative study of the four AI techniques used for the cancellation of MECG in the abdominal signal, three performance criteria namely mean square value of the estimated FECG signal, SNR and convergence time are computed in each case as discussed in

27 110 section 4.6. The parameters used for the comparison are MF=3, Epochs=10, SS=0.5 and samples =350. The computed values are presented in Table 5.1. Table 5.1 Performance comparison of AI techniques in FECG Sl.No. Technique Mean Square value of the estimated FECG SNR (db) Convergence Time(s) 1 BPN CCN ANFIS ANFIS -FCM From Table 5.1, it is concluded that ANFIS- FCM produces least Mean Square value of the estimated FECG and convergence time and highest SNR among the four techniques. The effects of varying parameters like epochs, step size, membership function on the performance are discussed in the following section EFFECT OF VARYING TRAINING PARAMETERS IN ANFIS In order to choose the optimum values for the important training parameters like epochs, step size, and membership function, it is necessary to study the effect of variation on the parameters.

28 Effect of Varying Epochs The number of epochs is varied from 10 to 50 in steps of 10 and the performance criteria are computed for each value epoch. The results of variation of epochs are presented in Table 5.2. Table 5.2 Comparison of performance criteria by varying the epochs Sl.No Epochs Mean Square value of the estimated FECG SNR(dB) Convergence Time (s) It is observed from Table 5.2 that when the number of epochs increases, its effect on Mean Square value of the estimated FECG and SNR is very less, but it increases the convergence time. Hence epoch 10 is selected by a default for ANFIS training Effect of Varying Step Size The step size in ANFIS training is varied from 0.2 to 0.7 in steps of 0.1. The computed values of performance criteria for different step sizes are given in Table 5.3.

29 112 Table 5.3 Comparison of performance criteria by varying the step size Sl.No Step size Mean Square value of the estimated FECG SNR (db) Convergence Time (s) Average It is noted from Table 5.3 that variation in SS produces random variation in Mean Square value of the estimated FECG, SNR and convergence time. Hence, an average SS value of 0.5 is used for training Effect of Varying the Number of MF The number of MF used in ANFIS training is varied form 2 to 6 in steps of one. The performance criteria are calculated for each MF and are shown in Table 5.4. It is inferred from Table 5.4 that when the MF increases, Mean Square value of the estimated FECG decreases which in turn increases the SNR. However, it also increases the convergence time. Hence, an MF value 3 is chosen as a compromise between the Mean Square value of the estimated FECG and the convergence time.

30 113 Table 5.4 Comparison of performance criteria by varying the number of MF Sl.No MF Mean Square value of the estimated FECG Convergence Time (s) Choice of the Type of MF Five different types of MF (Gaussian two sided, Trapezoidal, Gaussian Single sided, Triangular, Gbell) are tried for ANFIS training. The variation of performance criteria for these types is given in Table 5.5. When the type of MF is changed, there is no significant variation in Mean Square value of the estimated FECG and convergence time. Since Gbell has a property of smoothness and also yields less Mean Square value of the estimated FECG, it is used to generate the fuzzy model. The shape of the MF changes during the process of ANFIS training. The shape of the MF before and after training is shown in Figure The x axis in Figure 5.18 represents the amplitude range of the input and the y axis gives the membership value. During training, ANFIS varies the MF parameters to map the reference inputs with the target.

31 114 Table 5.5 Choice of the type of MF on the performance criteria Sl.No MF Type Mean Square value of the estimated FECG Convergence Time (s) 1 Gaussian 2 sided Trapezoidal Gaussian Triangular Gbell Figure 5.18 MF before and after training

32 115 The ANFIS-FCM technique uses the same training parameters as that of ANFIS. Hence, separate study of variation in performance criteria for changes in training parameters is not required CONCLUSION The motivation for monitoring the fetus during pregnancy is to recognize pathologic conditions, typically asphyxia, with sufficient warning to enable intervention by the clinician before irreversible changes set in. However, the monitoring techniques in current practice have serious shortcomings. The noise free FECG is required to overcome these limitations. Measurement of FECG is affected with many artifacts. The major artifact called MECG is cancelled from the FECG using four AI techniques. Performance comparison of these techniques has been made in terms of Mean Square value of the estimated FECG, SNR, and convergence time.