Performance Identification of Different Heart Diseases Based On Neural Network Classification

Transcription

1 Performance Identification of Different Heart Diseases Based On Neural Network Classification I. S. Siva Rao Associate Professor, Department of CSE, Raghu Engineering College, Visakhapatnam, Andhra Pradesh, India. T. Srinivasa Rao Associate Professor, Department of CSE, GIT, GITAM University, Visakhapatnam, Andhra Pradesh, India. Abstract The heart diseases are the most widespread induce for human dying. Every year, 7.4 million deaths are attributed to heart diseases (cardiac arrhythmia) including 52% of deaths due to strokes and 47% deaths due to coronary heart diseases. Hence identification of different heart diseases in the primary stages becomes important for the protection of cardiac related deaths. The existing conventional ECG analysis methods like, RR interval, Wavelet transform with classification techniques, such as, Support Vector machine K-Nearest Neighbor and Levenberg Marquardt Neural Network are used for detection of cardiac arrhythmia Using these methods large number of features are extracted but it will not identify exactly the problem. Studies conducted in this paper work to carry out these individual optimization techniques which did not give the desired identification accuracy. It is also proposed to modify these techniques as the system is advised based on universal filtering, pulse identification, clustering, categorization of signal with small delay can be done to identify the life threatening arrhythmia. Double differentiation with multi- discrete wavelet transforms which protect time changing QRS complex as well as noise. The proposed multi wavelet transform modeling provides temporal and spectral information coincidental, and extends reach stretch ability with a potential of multi wavelet functions of dissimilar signal properties. The present development implicated with the origin of QRS complexes of ECG signals using a new finite set of algorithms constructed based on ECG Q R S T waveform shaping it is possible to find out whether the person is in normal or abnormal and identification of different diseases with possibility effect, existing standard Pan-Tompkins method and multi wavelet transform techniques evaluated using MATLAB. Then removing discrepant wavelet transform coefficients and denoising is done in ECG signal. In addition to that QRS complex will be detected and each complex will be used to find the peaks of the individual waves like, P and T, and their derivatives. At the end we are going to generate different disease inputs for the ECG signal by providing the heart beat rate, generate p, r waves. The values can be altered by observing the heart diseases and based on that real time inputs we can build the neural network and check the efficiency of the system. Kurtosis, neural network classification to achieve significant identification accuracy. This paper mainly aims at improving the identification and performance of the system to detect heart disease based on neural network classification. Keyword: Electrocardiogram (ECG), QRS Detection, Wavelet Transform, kurtosis, neural network classification Introduction Globally heart diseases are the most widespread induce for human mortality. Every year, 7.4 million deaths are attributed to high Blood Pressure (BP) including 52% deaths due to strokes and 47% deaths due to the coronary heart diseases. Most cardiac diseases are due to risk factors, such as, diet, maximum blood pressure, tobacco usage, obesity, diabetes and physical inactivity. Electrocardiogram (ECG) represents the electrical activity of human heart. The changes in the voltages during re-polarization and depolarization of the heart fibers are recorded by placing electrodes on the surface of the chest and on the limb (limb leads). The ECG waveform is either printed on to graph paper that runs at a constant speed or displayed on a computer screen. The benefits of ECG are its portability, prompt accessibility and flexibility. Computerized ECG classification can also help reduce health care costs. In the biomedical instrumentation industry there is an ongoing quest for the early detection of heart abnormalities using ECG signals. The objective of this paper is to develop new methods for different diseases in order to identify the exact problem. The present method includes signal processing and feature extraction techniques to obtain the discriminative features of the ECG signals that correspond to various cardiac minor and major conditions. These features are classified using neural network classification techniques. Fig 1 shows the Normal ECG graphical representation. ECG is huge amount of data that provides relevant information about the hearts pathological and physiological condition for different people so that it is easy to identify the different types of diseases based collection of data. Because of non-stationary nature of ECG signals, it is an arduous task to analyze them manually. Hence there is a need for the automatic detection of heart ailments. The morphology of ECG changes due to the abnormalities in the heart. By having a glance at the ECG, an expert physician can easily detect heart diseases. 3859

2 reliable and a different QRS detection, sophisticated algorithm is not applied to get complete details of the disease problem not solved till today. Based on these drawbacks, a novel hybrid algorithm is developed, which incorporates assorted wavelet coefficients and Pan-Tompkins' method for extracting complete and appropriate features [10] and also possibilities of enhancement with the computational cost by using different parameters like uniformity, entropy, etc, to analyze ECG signal measurement statistically. Figure 1: ECG Signal Waveform Computer based automatic detection is one of the ways that can help doctors reduce their work load. This paper consists of five major steps for the detection of heart diseases, namely, Signal preprocessing, multi wavelet transform and neural network Classification. In the preprocessing phase, data can be collected from real diseases but in this paper it is collected from on-line MIT-BIH database. In this step the obtained ECG signals are filtered using high pass to remove noise contained in the ECG signal. In the second step, features are extracted using multi wavelet techniques. In general, the extracted features contain some redundant and non-discriminative features, which lead to computational burden and the performance loss especially in the case of ECG signal. Hence the above techniques do not give satisfactory detection accuracy. The classification accuracy of AF using AR coefficients is 74.5% and the classification accuracy of MI using WTC technique is 83.1% while the neural network classification accuracy of using kurtosis technique is 94.5%. Our goal is to predict the optimum features to overcome this problem for a selected neural network classification technique. Hence in the second step, optimization techniques are added to find the lower set of features in order to maximize the classification performance. The final step is the classification of the signal into ten different cardiac arrhythmias. Different techniques, such as, kurtosis and different wavelet transforms are used for ECG classification An improper, a series of discrete algorithms were previously implemented, to be a major advantage of the linear filtering and wavelet transform. Large extent, QRS detection is a challenge task to identify the different diseases and conditions of the patient by considering with different data with the same person major task. When the detection of the ECG signals waveform having some noise created by the electrode artifact and the placement of the electrode, baseline drift, and power line interference [9]. In most of the cases, the ECG signal may suddenly change with different shapes to the pathological or step changes, e.g. signals with very low QRS complexes or abrupt variable levels. So, to apply a Normal Heart Diseases Heart Disease refers to a common disorder or conditions. It concerned with a group of diseases or common problems that imply contracted or frozen blood vessels that can direct to a heart attack, chest pain (angina) or stroke. There are several types of heart diseases are identified depends on peoples ages, the general common type that affects the electrical system is known as arrhythmias. They can cause the heart to beat very fast (Tachycardia) or very slow (Bradycardia), or unexpectedly (Atrial fibrillation). Some of the heart diseases are discussed in the following section: Dextrocardia, Tachycardia, Bradycardia, Hyperkalemia, Myocardial ischaemia, Hypercalcaemia Sinoatrial block, sudden cardiac death, atrial fibrillation, Ventricular Fibrillation Proposed QRS Detection Algorithm The difficult operation during the detection of the QRS complex despite detecting the peak of the QRS complex or R wave, as in an electrocardiogram (ECG), the signal has a time-varying morphology [6]. This event takes place outstanding to the cause that an ECG signal is unprotected to physiological abrupt changes induced by the patient and the detected waveform having different shapes which are considered and represented by P Q R S T and it is corrupted due to noise. The QRS complexes are having a time varying morphology, and are not suitable to identify the disease and not the authentic signal portion in an ECG signal. To solve this, Q R S T waves with characteristics standardized to that of the basic QRS complex, as well as spikes from high frequency pacemaker s compromise the detection of the QRS complex. The input will be generated by providing the random values to the every wave. The program will generate the simulated diseased heart wave. The peak signals are obtained using the threshold the value of threshold is set by learning process by observing various ECG signals. Write the data into excel sheet for every iteration of program. After collecting all data sets paste it one excel sheet and this will be the input data set for the neural network. Read the excel sheet into Matlab and save those by the variable name. 3860

3 Q R S T to identify the diagnosis purpose. In this paper, we have successfully enforced ECG signal on various wavelet transforms techniques to extract sensitive abnormalities in the modest stages. The multi-wavelet transform applied to carry out the work are Haar wavelet, Daubechies wavelet, Bior 3.5 wavelet etc to enhance the precision of the feature of the ECG signal. The Fig.4 is mathematical function of a continuous variable into steps of coefficients. Decomposing the wavelet in the multi spectral using multi wavelet is very concept is useful to enforce the sensitivity. Figure 2: Basic Neural Network Architecture Fig 2 represents the basic Neural Network Architecture. Open neural network toolbox by typing nftool select the pattern recognition and imported data into tool and train the system. In program output is obtained and for testing purpose the procedure is same. Obtain the values and feed it into toolbox and import in the MATLAB before execution of the program. Figure 4: Wavelet function with P Q R S T Waveform Thresholding From the wavelet analysis techniques the signal is decomposed into near standard coefficients- which mapped the smoothed signal, and the particularization coefficients - that distinguish the noise content in the ECG signal. Such waveform portions can be forced out by carried out the process of thresholding that is by eliminating the coefficients whose values are less than the value threshold. Thresholding, at present is a view tool for the handling of cardio noise (high frequency components). While here we had used soft thresholding. Threshold process is achieved based on entropy and uniformity. Figure 3: Flow chart for the proposed work Wavelet Transform A wavelet can be termed as a small wave with less energy condensed in time. It is one of the sophisticated tools for the analysis of both frequency and time domain. There are different wavelets Transform techniques which are used to evoke important features from the ECG signal to represent P Double Differentiation An initial filter phase is normally used by all QRS detection algorithms since the typical frequency components of QRS complex ranges from around 5Hz to 25Hz. This process is done before the actual QRS detection to dominate the remaining attributes in the ECG signal which are the P, T waves, noise and baseline drift. Low pass filters are used to restrain the noise and the baseline drifts, while the other components like P and T waves are controlled by high pass filters. Hence the combinations of both the low pass filter and high pass filter yields the application of a band pass filter with cut-off frequencies of 5Hz and 25Hz meant for QRS detection. For many algorithms, the high pass filtering and 3861

4 the low pass filtering are segregated and are distinctly carried out. The QRS complex is detected using the comparison with the threshold using the filtered signals when the algorithms use only the high pass filters. Some other decision rules are employed to mitigate the false positives. Commonly in the older algorithms, the high pass filter was identified as a differentiator, due to which the QRS complex feature of having a large slope was used for its detection. It is observed that double differentiation of wavelets increases positive predictability which solves most of the problems. In this proposed algorithm, multiwavelets have been used to detect exactly the QRS complex with more sensitivity. The differentiator has the following difference equations The typical features of such algorithms is given by z(n).the contrast between the feature in the ECG and the threshold value gives the QRS complex. The selection of the threshold levels must be adaptive and flexible in nature and depend on varying signal morphology. When considering feature in equation, the threshold is proposed. Θx = 0.3 to 0.4*max[x] (9) where, the x is the signal segment and its maximum value is determined. This method of getting the threshold value is implemented in almost all QRS detectors [4], following which, various decision rules are applied to avoid false positives by using various peak detection logics represented. Thus, derivative detection method is used for identification of QRS in ECG signal using cumulative differentiation technique. Figure 5: Error estimation Squaring Function After differentiation, the signal is squared point by point, which makes all data points positive [11]. The equation for this is y(nt) = [x(nt)]2 (10) that is, predominantly the ECG frequencies does nonlinear amplification of the output of the derivative emphasizing higher frequencies is highlighted. Peak Detection For the peak detection, specific points of the signal are opted. Among all waves, the R peak [8] has the largest amplitude. The components of decomposed signals are kept and the others are discarded. So, in precise the QRS complex detection consists of determining the R point of the beat and it is squared. At the end the number of beats is calculated to know the time interval between successive heartbeats [12]. Figure 6: Performance Results The Multiwavelet transform is applied to ECG signal to initiate the undesired frequency. The fourth scale of Daubechies wavelet (db4) is used to attain this. After denoising, QRS Complexes are determined. This is obtained by implementing on MIT-BIH arrhythmia database [7]. It is clearly understood from the Table.1, that there is significant improvement in the error reduction using proposed algorithm. The tool will generate the neural model, training efficiency, error of the classification etc. the testing performance of the system will be displayed on the MATLAB command. The graphs shows how the system have learned through the training of input and based on the learning how much efficiently it has classified all these information is displayed on the command window. 3862

5 Figure 7: Testing and Validation Table I: Identification of Error when Wavelets are used S. No. Wavelet mutual group Error 1 Db, Haar 8% 2 Db,coefficient 7% 3 Haar coefficient 15% Table II: Identification of Abnormality when Kurtosis is used. S No KURTOSIS NAME OF ABNORMALITY ERROR 1 K1 Dextrocardia 0.01% 2 K2 Tachycardia 0.01% 3 K3 Bradycardia 0.02% 4 K4 Hyperkalemia 0.02% 5 K5 Myocardial ischaemia 0.02% 6 K6 Hypercalcaemia 0.05% 7 K7 Sinoatrial block 0.06% 8 K8 Sudden cardiac death 0.04% 9 K9 Atrial fibrillation 0.06% 10 K10 Ventricular Fibrillation 0.08% Table III: Comparision of Heart beat rate for all the points Heart P-Wave Q-Wave QRS wave S-Wave T-wave U-wave Beat Rate Conclusion In this study our aim is to automate the above procedure so that it leads to correct diagnosis. Early diagnosis and treatment is of great importance because immediate treatment can save the life of the patient. The morphology of ECG changes due to the abnormalities in the heart. This paper consists of three major steps for the detection of cardiac arrhythmia, Preprocessing, Feature extraction and Neural Network Classification. In the preprocessing, the data has been collected from MIT-BIH AF data base consisting normal sinus rhythm database (10 patients). The present research work proposed two efficient approaches for ECG classification. The first approach uses transform techniques for feature extraction: and detection accuracy using the above techniques is not satisfactory. To increase the detection accuracy and identification, in the second approach, multiwavelet techniques have been used to find the smallest set of features that maximize the classification accuracy. The result is analyzed statistically by taking different kurtosis methods for the identification of different diseases. The error percentage for the proposed algorithm when compared to the existing algorithm is low and same for few data records, but for some data records it is little bit high. Since the analysis is taken from different records of heart rate data, the comparison is done for all the points like p q r s u v with different heart beats. By taking mean, it is directly showing that the proposed algorithm is having less error compared to the existing algorithm. Due to this, the proposed algorithm is more efficient. References [1] B. U. Kohler, C. Henning and R. Orgelmeister, The principles of software QRS detection, IEEE Engineering in Medicine and Biology Magazine, Vol. 21, No. 1, pp , [2] J. Pan and W. J. Tompkins, A real-time QRS detection algorithm, IEEE Transactions on Biomedical Engineering, Vol. BME-32, No. 3, pp , [3] P. S. Hamilton and W. J. Tompkins, Quantitative investigation of QRS detection rules using the MIT/BIH arrhythmia database, IEEE Transactions on Biomedical Engineering, Vol. BME-33, No. 12, pp , [4] N. M. Arzeno, Z. D. Deng and C. S. Poon, Analysis of first derivative-based QRS detection algorithms, IEEE Transactions on Biomedical Engineering, Vo1. 55, No. 2, pp , [5] F. Zhang and Y. Lian, QRS Detection Based on Multiscale Mathematical Morphology for Wearable ECG Devices in Body Area Networks, IEEE Transactions on Biomedical Circuits and Systems, Vol. 3, No. 4, pp , [6] C. Li, C. Zheng and C. Tai, Detection of ECG characteristic points using wavelet transforms, IEEE Transactions on Biomedical Engineering, Vol. 42, No. 1, pp ,

6 [7] MIT-BTH Arrhythmia Database, 2nd edition, Available at: /physiobank /database /html/, [8] Cuiwei Li and Chongxun Zheng, QRS detection by wavelet transform, Proceedings of Annual Conference on Biomedical Engineering, Vol. 15, pp , [9] V. R. Sarma Dhulipala and G. R. Kanagachidambaresan, Cardiac Care Assistance using Self Configured Sensor Network a Remote Patient Monitoring System, Journal of The Institution of Engineers (India): Series B, Vol. 95, No. 2, pp , [10] S. Alavi and M. Saadatmand-Tarzjan, A new combinatorial algorithm for QRS detection, IEEE 3th International econference on Computer and Knowledge Engineering, pp , [11] Abhilasha M. Patel, Pankaj K. Gakare and A. N. Cheeran, Real Time ECG Feature Extraction and Arrhythmia Detection on a Mobile Platform, International Journal of Computer Applications, Vol. 44, No. 23, pp , [12] Pooja Sabherwal, Wavelet Transform As Method for ECG Signal Analysis, International Journal of Emerging Science and Engineering, Vol. 2, No. 1, pp , Author s Profile: Mr. I.S Siva Rao has completed his M.Tech in Computer Science and Technology from Andhra University. He is working as faculty member in the Department of CSE, Raghu Engineering College, Visakhapatnam. He has 15 years of teaching experience. Presently pursing Ph.D. in computer Science and Engineering from GITAM University. Dr T.Srinivasa Rao, has received his Doctoral degree in Computer Science and Engineering from Andhra University. He is working as Associate Professor in GITAM University, Visakhapatnam. He has 17 years of teaching experience. 3864