INTERNATIONAL JOURNAL OF COMPUTER ENGINEERING & TECHNOLOGY (IJCET)

Transcription

1 INTERNATIONAL JOURNAL OF COMPUTER ENGINEERING & TECHNOLOGY (IJCET) International Journal of Computer Engineering and Technology (IJCET), ISSN (Print), ISSN ISSN (Print) ISSN (Online) Volume 3, Issue 2, July- September (2012), pp IAEME: Journal Impact Factor (2011): (Calculated by GISI) IJCET I A E M E PERFORMANCE EVALUATION OF MULTILAYER PERCEPTRON NEURAL NETWORK BASED CARDIAC ARRHYTHMIA CLASSIFIER Shivajirao M. Jadhav 1, Sanjay L. Nalbalwar 2 and Ashok A. Ghatol 3 1 Information Technology Department, Dr. Babasaheb Ambedkar Technological University Lonere, Maharashtra , India smj33@rediffmail.com 2 Electronics and Telecommunications Department, Dr. Babasaheb Ambedkar Technological University Lonere, Maharashtra , India slnalbalwar@dbatu.ac.in 3 Campus Director, K. J. Educational Institutions Pisoli, Kondahwa-Saswad Road, Near Bopdeo Ghat, Pune, Maharashtra , India vc_2005@rediffmail.com Abstract Today heart disease is the most common cause of death in the world so the detection and treatment of arrhythmias has become one of the biggest challenges for cardiac care unit. Artificial neural networks have been successfully applied to classify medical or diagnostic data. The main objective in this paper is to distinguish between the normal and abnormal subject cases of cardiac arrhythmia. A Multilayer perceptron (MLP) model with momentum learning rule approach is applied to classify ECG arrhythmias after replacing missing values by closest column method. We used UCI repository benchmark data for this classification purpose. The total numbers of samples are 452 in data. Five data sets are formed by varying training and testing percentages and separate experiments are performed on each data set by varying hidden layers from one to three. In this study, we are mainly interested in producing high confident arrhythmia classification results to be applicable in diagnostic decision support systems. The classifier performance is evaluated using various measures such as sensitivity, specificity, 1

2 classification accuracy, mean squared error (MSE), receiver operating characteristics (ROC) and area under curve (AUC). The experimental results presented in this paper are very attractive in terms of classification accuracy. Keywords: Accuracy, ECG arrhythmia, machine learning, momentum learning rule, sensitivity, specificity. 1. Introduction Cardiac arrhythmia, disorders of cardiac rhythm, may indicate the susceptibility of serious heart disease, stroke or sudden cardiac death. Early diagnosis of cardiac arrhythmia makes it possible to choose appropriate anti-arrhythmic drugs, and is thus very important for improving arrhythmia therapy. Various Machine learning (ML) and data mining methods have been applied to improve the accuracy for the detection of ECG arrhythmia. Once a data mining task is identified, appropriate methods have to be selected for execution of this task. Method selection depends highly on the application context as given by initial task analysis, on the properties of the data on which the analysis is being performed, on previous experience with similar domains, and on user-specified requirements for the results [1]. Electrocardiogram records electronic activities of the heart, and has been widely adapted for diagnosing cardiac arrhythmia [2]. By far, a number of signal processing [3], pattern recognition [4, 5], and machine learning [6] methods had been proposed. The publications of several generally available arrhythmia data sets also played an important role in stimulating research on cardiac arrhythmia diagnosis [7, 8]. In this paper, we proposed an AAN based approach, which can classify ECG arrhythmia into normal and abnormal classes i.e. distinguish between presence and absence of cardiac arrhythmia patient cases. We used Multilayer perceptron (MLP) neural network model with static backpropagation algorithm. We have varied one to three numbers of hidden layers while designing various network topologies. The proposed method first cleans the data set by replacing missing values by closest column values of the concern class. 2. Related Research Work Gao and Madden [1] developed an arrhythmia detection system with ECG signals based on a Bayesian ANN Classifier and its performance is compared with that of other classifiers, specifically Naive Bayes, Decision Trees, Logistic Regression and RBF Networks. Zuo et al. [2] proposed a kernel difference weighted k-nearest neighbor classifier (KDF-WKNN) for the diagnosis of cardiac arrhythmia based on the standard 12 lead ECG recordings. They have used a modified principal component analysis (PCA) approach to cope with the missing attribute values. This approach is different from classical K- nearest neighbor (KNN) classifier. Several methods for automated arrhythmia detection have been developed in the past few decades to attempt to simplify the monitoring task [9]. These include Wavelet transformation 2

3 [10-12], RBF Neural Networks [13], self- organizing map [14] and fuzzy c-means clustering techniques [15]. Multilayer neural networks are used to classify arrhythmia QRS complexes, and for ischemia detection [16-17]. Parts of this work using MLP and modular neural network is available in [18] and [19] respectively. An Artificial immune recognition system (AIRS) with fuzzy weighted pre-processing [20] is also used for arrhythmia classification. ECG arrhythmia classification and Fetal state classification using ANN models is also available in [21-26]. Various neural network models are used to classify ECG arrhythmia and classification accuracies are reported in [27-30]. 3. Methods 3.1 Description of data set The Cardiac Arrhythmia Database from the UCI Machine Learning Repository [8] is used. This data set is a good environment to test classifiers as it is incomplete and ambiguous bio-signal data collected from total 452 patient cases. The first class is Normal, and the other 15 classes are 15 kinds of arrhythmia. These 15 classes are merged into a single class called Abnormal class a representative 15 arrhythmia classes. For each sample, there are 279 attributes, where the first four, age, sex, height, and weight, are the general description of the participant, and the other 276 attributes are extracted from the standard 12 lead ECG recordings. For the details of the data set, please refer to [4, 6, 8]. There are two significant characteristics which should be noted for the UCI cardiac arrhythmia database. The entire database is first preprocessed to replace missing attributes. We have used closest column value of the concern class. And later all the records are randomized. 3.2 Data set groups The original data set grouped into different data sets as shown in the table 1 and each group is partitioned into two subsets viz. training set and testing set except the last group labeled as DSMains in which all 452 instances are used for training purpose only. Table 1: Data Set Group Partitions Data Set Name Training % age Testing % age Training instances Testing Instances Data set 1 (DS1) Data set 2 (DS2) Data set 3 (DS3) Data set 4 (DS4) Data set 5 (DS5) Main Data Set (DSMains) Training set itself all 452 instances For Training only. 3

4 3.3 Multilayer Perceptron Network Model Artificial neural network (ANN) is a data processing mathematical model. It consists of a number of units or elements called nodes or neurons. These nodes or neurons are arranged in layers and are interconnected by weights and biases between the layers. The first layer is known as input layer and the last layer is known as output layer. The layers between input layer and output layer are known as hidden layers. The number of input neurons depends on the number of input variables to the ANN model while the number of output neurons depends on the number of output desired form the model. There is no hard and fast rule to determine the number of neurons in the hidden layer. It depends on the complexity of the problem sought. Usually three stages are considered in ANN applications, viz., (i) training, (ii) validation and (iii) testing. During training and validation stages both input and target data is introduced to the network. But in testing stage a different set of data containing only input values is fed to the network. Figure 1 illustrates architecture of a simple multilayer perceptron (MLP) neural network model with two hidden layers. Fig. 1: Multilayer Perceptron Neural Network Model This is a feed forward network and in this networks all the information is transferred in the forward direction only and there is no loop or cycle in the network. In Fig. 1, the symbols i, j and k represent the i th (1<i<P), j th (1<j<L) and k th (1<k<M) neuron in input, hidden and output layers, respectively. P, L and M are the total number of neurons in input, hidden and output layers, respectively. The input value to individual input neurons are denoted by the letter x, y is the output from output neurons and z is the target value introduced to the network during training. The purpose of training a network is to minimize the error between outputs of the network and target values. Training algorithm reduces the error by adjusting weights and biases of the network. In training, input values are multiplied by respective connection weights and then biases are added. The same process is repeated for output layer where output of hidden layer is used as input for output layer. The combination of net weighted input and biases net j to the j th neuron of the hidden layer can be expressed as [31]: where, j=x +b (1) 4

5 x i is the input value to the i th neuron of input layer, while w ji is the weight of j th neuron of hidden layer connected to the ith neuron of the input layer and bj is the bias of jth hidden neuron. The net value netj is passed through a transfer or activation function in the hidden layer to produce an output from the hidden neuron. The output from the hidden layer can be expressed as [36]: = ( )= x w +b (2) where, y j - the output from j th hidden neuron. The output from the hidden layer y j is used as an input to the output layer and the same process as in hidden layer is repeated in the output neurons to produce output from output layer. The net weighted input to the output neuron can be represented by [31]: = + = ( + + (3) where, is the weight of k th neuron of the output layer connected to the j th neuron of the hidden layer and b k is the bias of the k th neuron in the output layer. The output from the k th neuron in the output layer is given by [31]: =( )= ( + ) + (4) The error e k of the k th neuron in the output layer, which is the square of the difference between the target and output, can be written as [31]: =( ) (5) If there are L numbers of input data pairs such that (1 l L), the global error of the network in terms of mean squared error is written as [36]: (6) ( ) Equation 6 is called the error function of the network and is a function of network connection weights. During training the global error function is minimized by the training algorithm which enables updating (adjusting) the connection weights through successive iterations such that the difference between target and output values for the whole dataset is within a predefined tolerance limit [31]. 4. Experimental Results Experiments are performed on Neuro Solutions (version 5.0) software simulation tool [32]. We have varied number of hidden layers (HLs) from 1 to 3 for each data set. Testing 5

6 classification results for all data sets are shown in given in table 2. Resultss shown in table 2 against data set DSMains are obtained using all 452 instances for training only. Observing classification results it is clear that the data set 4 gives better performance in terms of training sensitivity, specificity and accuracy. Training MSE is shown in figure 2. For two to three number of HL design models training is very good as MSE for these designs is low as compared to one number of hidden layer design model. Figure 3 shows training data set specificity, sensitivity and accuracy. Training data sets have given very attractive performance in terms of these measures. Figure 4 gives performance against testing for all the data sets. Testing accuracy of arrhythmia classifier for data set 5 is very good and it is above 80% for all the three designs of ANN models. From figure 4 it is clear that testing accuracy is very attractive (86.67%) for network design with two numbers of HLs for this data set 5. Testing specificity and sensitivity is shown in figure 5. Sensitivity for MLP design of two numbers of HLs is very attractive and it is 93.75%. Data set DS5 has given best classification results therefore for this data set ROC matrix is graphed as an ROC curve as shown in figure 6. Area under ROC curve is higher as compared with ANN models with 1 and 3 numbers of hidden layers as shown in figure 7 for this data set 5. Table 3 gives the comparison of classification accuracy for ECG classification problem we have considered and existing classifiers designed by other researchers. Table 2: Arrhythmia Classification Results Data Set No. of Hidden Sensitivity Specificity Classification Name Layers (%) (%) Accuracy (%) DS DS DS DS DS DSMains Fig. 2: Training MSE 6

7 Fig. 3: Training Specificity, Sensitivity and Accuracy Fig. 4: Testing Accuracy for all Data Sets Fig.5: Testing Sensitivity and Specificity Fig. 6: Testing ROC Curve for Data Set 5 Fig. 7: Area under ROC Curve for Data Set 5 7

8 Table 3: MLP model s classification accuracy for ECG classification problem with classification accuracies obtained by other methods in literature A novel pruning method [35] KDFW KNN [2] SVM with Gaussian Kernel HLVQ [33] Fuzzy weighted AIRS NEW -FM [32] Modular NN (Our earlier work) [24] MLP with two HLs (our work) [34] [25] % CONCLUSION With the improvements in expert systems and ML tools, the effects of these innovations are entering to more application domains day-by-day and medical field is one of them. Decision-making in medical field can be a trouble sometimes. When investigators design neural networks for the application of arrhythmia classification presented in this paper, there are many ways used to investigate the effects of various ANN MLP models which refer to the specification of network size (i.e. number of HLs) when the number of inputs and outputs are fixed. When we design MLP network analyze the performance of arrhythmia classification we reach to the following facts: MLP with two numbers of HLs with momentum learning rule gives better classification accuracy as compared to other ANN MLP topologies. Our experimental results strongly suggest that MLP model can aid in the diagnosis of cardiac arrhythmias. Our experimental results on the UCI cardiac arrhythmia database show classification accuracy of for data set 5. It is also proved that to evaluate the performance of classifier almost all different performance measures are required to evaluate the performance of neural network based classifier. We hope that this system can be further developed and fine-tuned for practical applications. References [1] Dayong Gao and Michael Madden (2003), Bayesian ANN Classifier for ECG Arrhythmias Diagnostic System, Proceeding of IEEE International Joint Conference on Neural Network pp [2] Zuo, W.M. et. al.(2008), Diagnosis of cardiac arrhythmia using kernel difference weighted KNN classifier, Computers in Cardiology, pp [3] Thakor NV and Zhu Y. S. (1991), Applications of adaptive filtering to ECG analysis: Noise cancellation and arrhythmia detection, Biomedical Engineering, pp [4] Coast D.A. et. al. (1990), An approach to cardiac arrhythmia analysis using hidden Markov models, Biomedical Engineering, 37(9), pp [5] Lima C.S. and Cardoso M.J. (2007), Cardiac arrhythmia detection by parameters sharing and MMIE training of hidden Markov models, Engineering in Medicine and Biology Society, pp [6] Guvenir H.A. et. al., (1997), A supervised machine learning algorithm for arrhythmia analysis, Computers in Cardiology, pp; [7] Moody G.B. and Mark R.G. (2001), The impact of the MIT-BIH arrhythmia database, Engineering in Medicine and Biology Magazine, 20(3), pp

9 [8] Blake C.L. and Merz C.J. (1998), UCI Repository of Machine Learning Databases, Available online from: (Downloaded date: 25th January, 2012 ) [9] Raut R.D. and Dudul S.V. (2008), Arrhythmias Classification with MLP Neural Network and Statistical Analysis, First IEEE International Conference on Emerging Trends in Engineering and Technology, pp [10] G. Selvakumar, and K. Boopathy Bagan (2007), Wavelet Decomposition for Detection and Classification of Critical ECG Arrhythmias, Proceeding of the 8th WSEAS International Conference on Mathematics and computers in Biology and Chemistry, pp [11] G. Selvakumar, K. Boopathy Bagan (2006), An Efficient QRS Complex Detection Algorithm using Optimal Wavelet, WSEAS Transactions on Signal Processing, Volume 2, Issue 8, pp [12] Sung-Nien and Yu,Ying-Hsiang Chen (2007), Electrocardiogram beat classification based on wavelet transformation and probabilistic neural network, Pattern Recognition, Elsevier Science Inc New York, NY, USA, Volume 28, Issue 10, pp [13] Hafizah Hussain and Lai Len Fatt (2007), Efficient ECG Signal Classification Using Sparsely Connected Radial Basis Function Neural Network, Proceedings of the 6th WSEAS International Conference on Circuits, Systems, Electronics, Control and Signal Processing, pp [14] Rahime et. al. (2007), Comparison of FCM, PCA and WT techniques for classification ECG arrhythmias using artificial neural network, Elsevier s Expert Systems with Applications Volume 33, Issue 2, pp [15] Labib Khadra, et. al. (2005), A quantitative analysis approach for cardiac arrhythmia classification using higher order spectral techniques, IEEE Transactions on Biomedical Engineering, Vol. 52, No. 11, pp [16] Yang Wang et. al. (2001), A Short Time Multifractal Approach for Arrhythmia Detection Based on Fuzzy Neural Network, IEEE Transactions on Biomedical Engineering, Vol. 48, No. 9, pp [17] Bortolan et. al. (2005), Comparison of four methods for premature ventricular contraction and normal beat clustering, Computers in Cardiology, September pp [18] S. M. Jadhav et. al. (2010), Artificial Neural Network Based Cardiac Arrhythmia Classification Using ECG Signal Data, in Proc. Int Electronics and Information Engineering (ICEIE) Conf. On, vol. 1 [19] Shivajirao Jadhav et. al. (2010), Modular Neural Network based Arrhythmia Classification System using ECG Signal Data, International Journal of Information Technology & Knowledge Management (ISSN ) Vol-IV, Issue-I 9

10 [20] Kemal Polat et. al. (2006), A new method to medical diagnosis: Artificial immune recognition system (AIRS) with fuzzy weighted pre-processing and application to ECG arrhythmia, Expert Systems with Applications, Vol. 31, Issue 2, pp [21] Shivajirao Jadhav et. al. (2011), Artificial Neural Network Based Cardiac Arrhythmia Disease Diagnosis, 2011 International Conference on Process Automation, Control and Computing Coimbatore India, Pages 1-6 [22] Shivajirao Jadhav et. al. (2010), Arrhythmia Disease Classification using Artificial Neural Network, 2010 IEEE International Conference on Computational Intelligence and Computing Research, Coimbatore India, pp. 1-4 [23] Shivajirao Jadhav et. al. (2010), ECG Arrhythmia Classification using Modular Neural Network Model, in Proc. IEEE EMBS Conference on Biomedical Engineering & Sciences, Kuala Lumpur, Malaysia, pp [24] Shivajirao Jadhav et. al. (2010), Generalized Feedforward Neural Network based Cardiac Arrhythmia Classification from ECG Signal Data, th International Conference on Advanced Information Management and Service (IMS) with ICMIA, Seoul South Korea, pp [25] Shivajirao Jadhav et. al. (2011), Foetal State Classification from Cardiotocography Signal Recordings Data using Multilayer perceptron Artificial Neural Network Model, International Joint Conference on Advances in Signal Processing and Information Technology, (Accepted) [26] Jadhav S. et. al. (2011), Modular neural network model based foetal state classification, IEEE International Conference on Bioinformatics and Biomedicine Workshops pp [27] Sang-Hong et. al., Extracting Input Features and Fuzzy Rules for Detecting ECG Arrhythmia Based on NEWFM, International Conference on Intelligent and Advanced Systems, Division of Software, Kyungwon University, Korea [28] Alaa M. Elsayad (2009), Classification of ECG arrhythmia Using Learning Vector Quantization Neural Networks, International Conference on Computer Engineering & Systems, pp [29] Uyar A., and Gurgen F. (2007), "Arrhythmia Classification Using Serial Fusion of Support Vector Machines and Logistic Regression," 4th IEEE Workshop on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications,, vol., no., pp [30] Ali Mirza Mehmood and Mrithyunjaya Rao Kuppa (2012), A novel pruning approach using expert knowledge for data-specific pruning, Engineering with Computers pp [31] Bishop, C.M. (1995), Neural Networks for Pattern Recognition, Oxford University Press. [32] (Downloaded date: 25th January, 2012) 10

11 Shivajirao M. Jadhav has received B.E. (Computer Science & Engineering) in 1993 from SGGS College of Engineering and Technology, Nanded India and M.E. (Computer Engineering) in 2003 from VJTI, Mumbai India. He is presently working as Associate Professor of Information Technology Department at Dr. Babasaheb Ambedkar Technological University, Lonere, Raigad, India. He is doing his Ph.D. study in Computer Engineering Department in the same University through UGC s Faculty Improvement Program (FIP. His area of interest includes signal data analysis using machine learning and soft computing techniques, real time distributed databases, computer networks, artificial neural networks and its applications to medical decision support systems. Dr. Sanjay L. Nalbalwar has received B.E. (Computer Science & Engineering) in 1990 and M.E. (Electronics Engineering) in 1995 from SGGS Institute of Engineering and Technology, Nanded, India. He has completed Ph.D. from IIT Delhi in He has around 21 years of teaching experience and is working as Associate Professor of Electronics & Telecommunication Engineering Department at Dr. Babasaheb Ambedkar Technological University, Lonere, District Raigad, India. His area of interest includes Multirate signal processing and Wavelet, stochastic process modeling. Dr. Ashok A. Ghatol born on 29th August, 1948 is the Campus Director of K. J. Educational Institutions, Kndhwa, Pune and Former Vice-Chancellor of Dr. Babasaheb Ambedkar Technological University, Lonere-Raigad, India. Before Joining as Vice-Chancellor, he was Principle at College of Engineering, Pune during and Principle at Government College Engineering, Amravati during He holds a B.E (Electrical), Nagpur University, M. Tech. (Electrical) and Ph.D. in Electrical Engineering from I.I.T., Mumbai in the field of High Power Semiconductor Devices. Over the last 33 years, he has been actively involved in the field of Technical Education as Academician, Researcher, Teacher, Planner and Administrator. Under his able and effective leadership, the Dr. Babasaheb Ambedkar Technological Lonere-Raigad University researched to a newer height and gained a feather in cap by receiving the UGC recognition under section 12(B), started receiving grants from UGC under various schemes. Taking into consideration his active involvement in technical education, AICTE also entrusted a responsibility to Act as Chairman of Western Regional Council. He was Vice-Chairman of ISRE for advising and assisting the council on issues and strategies on quality and quantity in the technical education. He was adjusted as Eminent Executive Member of Indian Society of technical Education, New Delhi. 11