ECG Diagnostics based on the Filter-Bank Signal Processing and ANN/SVM Classification

Transcription

1 ECG Diagnostics based on the Filter-Bank Signal Processing and ANN/SVM Classification EMIR TURAJLIC, DZENAN SOFTIC, EHSAN EYDI Department of Computer Science University Sarajevo School of Science and Technology Hrasnicka cesta 3a, 7 Sarajevo BOSNIA AND HERZEGOVINA emirturajlic@sssteduba dzenansoftic@stusssteduba ehsaneydi@sssteduba wwwsssteduba Abstract: - Processing and classification of electrocardiogram (ECG) recordings are some of the most challenging fields of biomedical signal processing owing to the fact that ECG signals commonly exhibit complex temporal morphology and contain numerous artifacts of data collection process This paper presents a comparative analysis between the Artificial Neural Networks and Support Vector Machines classification performances based on the feature vectors developed from the Filter-Bank processing of ECG signal The system is evaluated in the context of Supraventricular Arrhythmia diagnostics FIR Filter-Bank decomposes the ECG waveform into a various frequency components and enables independent temporal and spectral processing of ECG signal The feature vectors are developed as a set of statistical measures that describe the energy distribution in the individual sub-bands The considered statistical descriptors include, variance, ness and kurtosis In this paper, a systematic study of diagnostic performance is imposed on the choice of feature vector An optimal Filter-Bank size is ascertained and the relevance of individual frequency bands is evaluated Furthermore, the diagnostic relevance of statistical descriptors is assessed The experimental results demonstrate that optimization of feature vectors, in terms of sub-band selection and statistical descriptor choices, leads to a considerable reduction in the feature vector size and to an improvement in the classification accuracy rate Key-Words: - Biomedical signal processing; ECG diagnostics; filter-banks; support vector machines; ANN Introduction Electrocardiogram (ECG) is a time-varying signal corresponding to the electrical activity of cardiac muscle and is readily obtained as a measurement of the potential difference between electrodes placed on a surface of the skin The number of people suffering from cardiac illnesses is on the rise, as well as the cost associated with their management and treatment Thus, it is increasingly important to develop a reliable, an objective and a fully automatic method for ECG monitoring and diagnostics However, even after decades of research, ECG signal analysis remains one of the most challenging undertakings in modern biomedical signal processing ECG signal belongs to a category of non-stationary, quasi-periodic waveforms It frequently exhibits complex nonlinear temporal morphology and contains artifacts of data collection process, such as baseline wander (caused by respiration) and high-frequency electromyography noise A number of approaches for ECG signal analysis have been proposed, including Gauss curve modeling via nonlinear optimization algorithms [], Hilbert Transform based modeling [2], Mealy and Moore automata model [3], threshold methods [4], wavelet transform and principal component analysis [5], Archetypical Analysis [6], Hidden Markov modeling [7], [8], and Filter Bank approach [9], [] Filter bank (FB) approach is used to decompose an ECG signal into multiple frequency sub-bands and thus, enable independent processing of temporal and spectral domains As such, filter bank signal processing methods have been successfully employed on range of ECG applications, including beat detection, beat classification, ECG enhancement and noise alert [9-2] This paper presents a study of filter bank based processing of ECG signal for the purpose of feature vector development for Supraventricular Arrhythmia database diagnostics The proposed method relies on the FIR filter bank to decompose an ECG signal into a number of signals ISBN:

2 corresponding to uniformly distributed frequency bands A set of statistical measures are employed to describe the energy distribution in each sub-band and to form a feature vector for the ECG signal Support Vector Machines (SVM) and Artificial Neural Networks (ANN) are considered as the principal algorithms for ECG-feature-vector classification In this paper, the effect of filter bank size on the diagnostic accuracy is examined The diagnostic relevance of each statistical parameter and individual sub-bands is studied, in isolation and in combination A systematic study of feature vector candidates is proposed in an attempt to reduce the feature vector size and improve diagnostic accuracy 2 Classification Algorithms In this section, a review of the considered classification algorithms is presented Figure SVM hyperplane U wlj U2 l = L+ UL- UL UBias= j = J+ wjk k = K S SK- SK 2 Support Vector Machines The basic principle behind SVM is to map the input vector to a higher dimensional space and to construct a hyperplane to classify the training data [3] [4], as illustrated in Fig In order to maximize the distance between the two classes and to get the optimal hyperplane, two parallel hyperplanes are constructed on each side of the separating hyperplane The larger distance between these two hyperplanes denotes better classification performance Based on the sample class, parallel hyperplanes are constructed in form of: wx b Where w represents a p-dimensional vector and b is the offset parameter that controls the separating hyperplane margin The points, x, along the hyperplane represent the supporting vectors The optimal separating hyperplane can be found by using Lagrange multipliers, as in (2) T L ( w, b, a) w ai ( yi ( w xi b) ) 2 Here, a i represents a Lagrange s multiplier Lagrange multiplier are minimized with respect to w and b and maximized with respect to a i for ( a i >) Primal or dual form can be used for solving this problem Nonlinearity of classification process is addressed via a kernel function n i INPUT LAYER HIDDEN LAYER Figure 2 ANN Structure OUTPUT LAYER In this paper, it is assumed that linear separation of features is not possible and thus, the Radial Basis Function (RBF) kernel method is employed, with the following feature space mapping: 2 j K( x, x ) exp( x, x ) i j Here, denotes the kernel parameter In comparison to the polynomial kernel, RBF has a reduced number hyperparameters and fewer numerical difficulties 22 Artificial Neural Networks Fig 2 presents a schematic diagram of a general artificial neural network architecture consisting of L inputs, J neurons in a hidden layer and K outputs Each neuron computes the weighted sum of its inputs and subsequently, passes the output through a activation function to obtain a neuron response The most important features of neural network classifier are related to its architecture, the choice of activation function and the choice of training method Under the proposed classification scheme, the informal experiments have shown that a 6- neuron hidden layer structure with a single output neuron, in conjunction with a sigmoid activation function and the Levenberg-Marquardt training i ISBN:

3 algorithm constitute the optimal ANN design The Levenberg-Marquardt method is, effectively, a hybrid algorithm that combines the advantages of the second order Gauss-Newton method and the first order Gradients Descent [5] The Levenberg- Marquardt learning method adopts a general second order learning formulation: m m m m m w w Δw w H g, m Here, w m, gm denote the ANN weight values and the error derivative at iteration step m, respectively However, in Levenberg-Marquardt learning method, Hessian matrix H, is modified to include a conditioning term, e I, which ensures that the approximated Hessian matrix is readily invertible The Levenberg-Marquardt weight update takes the following form: Figure 3 A proposed system for ECG diagnostics m m m m m m, T w w J J e I J e m Here, I represents identity matrix, e denotes natural exponential, J denotes the Jacobian matrix, and λ corresponds to an automatically evaluated constant that ensures the stability of solution Filter Coeffic ient V alue Filter-Bank 4: Hz 3 ECG Classification System Fig 3 presents a block diagram of the proposed ECG diagnostic system The system relies on SVM or ANN algorithm to classify feature vectors derived from the Filter Bank processing of ECG signal An FIR Filter Bank, decomposes the ECG signal into N uniformly distributed frequency bands that together encompass the range -64 Hz The upper limit of the considered frequency range corresponds to the maximum frequency represented by a signal with the sampling rate of 28Hz, a Nyquist frequency This value is related to the lowest sampling frequency in the considered ECG signal database and it corresponds to the operating sampling rate of the system presented in Fig 3 The filter bank filters are designed as Finite Impulse Response (FIR) filters with 56 coefficients In Fig 4, the coefficient values corresponding to the fourth filter in a 6-filter-bank system are displayed This particular filter is a band-pass filter with the passband range Hz In general, all filters in the filter-bank exhibit linear phase property and have the identical group delay An example of Filter bank processing of ECG signal is illustrated in Fig 5 Figure 4 Am plitude FIR-Filter cooefficients for the 4 th filter in a six filter-bank system ECG signal output:filter output:filter output:filter output:filter Am plitude output:filter output:filter Figure 5 An example of an ECG wavform and the corresponding filter outputs for a six filter bak system ISBN:

4 Here, the filter outputs for a six filter bank system corresponding to a sample ECG waveform are presented For each sub-band, energy of the ECG signal is evaluated over the duration of 5 seconds, every 5 seconds Subsequently, the energy in each sub-band is scaled to represent the percentage of total energy found in the range of -64Hz Subsequently, four statistical parameters, namely, variance, ness and kurtosis are used to independently describe the energy distribution in each sub-band These parameters are concatenated to form a 4xN long ECG feature vector, where N denotes the total number of filters in a filter bank When a database of feature vectors representing the healthy and pathological ECG signals is formed, the feature vectors are normalized using z-score normalization method These normalized feature vectors are employed for training and testing of the considered classification algorithms The classification results are employed to assess the quality of various feature vectors and to ascertain the levels of diagnostic information contained in the individual frequency sub-bands in relation to the selected set of statistical descriptors 4 Results and Discussion The training and testing of the ECG diagnostic system is based on a database containing 73 onehour-long recordings of ECG signal [6], including 32 cases of healthy control recordings from the Fantasia database, sampled at 25 Hz, and 4 cases related to the Supraventricular Arrhythmia database, sampled at 28 Hz The dataset is randomly divided into training (5% of available data) and testing data (5% of available data) Prior to the experiments, the entire database of ECG recordings is re-sampled to 28 Hz and the baseline wander is removed from the signal with the linear phase, high-pass filter with the cut-off at 8 Hz Examples of healthy and pathological ECG signals are shown in Fig 6 The classification performance is evaluated in terms of sensitivity (Se), specificity (Sp) and accuracy (Acc), defined in (7), (8) and (9), respectively Se TP /( TP FN ) Sp TN /( TN FP) Acc TP TN /( TP TN FP FN ) Here, TP and TN denote the number of true positive and true negative cases, respectively, while FP and FN denote the number of false positive and false negative cases, respectively Figure 6 Examples of a healthy and a pathalogical ECG signal The results are reported as the average values obtained on an ensemble of 5 independent training and testing scenarios In the first experiment, the effect of filter bank number on the ability to differentiate between healthy signals from pathological ECG signals is examined The filter bank size is varied in the range between 4 to 2 filters In each instant, the statistical description of the sub-band energy distribution is characterized by all four considered statistical measures:, variance, ness and kurtosis The experimental results are reported in Table The table reports the feature vector size, ANN accuracy and SVM accuracy results for each of the considered filter bank sizes It should be noted that the principal repercussion of having a greater number of filters in a filter bank are an increase in the feature vectors size, improvement in spectral resolution and a reduction in the pass-band range of individual filters The results clearly indicate that the classification performance is strongly affected by the choice of filter bank size The highest classification accuracy levels, 949% for SVM and 933% for ANN, are attained when ECG signal is decomposed into 6 uniformly distributed frequency intervals This optimal filter bank size corresponds to a feature vector with 24 parameters It can be observed that both, ANN and SVM results display similar trends across the filter bank size values In general, the further the filter bank size is increased or lowered from the optimal value, the lower the quality of classification performance However, for each of the considered filter bank sizes, SVM, if only but marginally outperforms the artificial neural network classifier Thus, SVM performance is examined in more detail and the corresponding sensitivity and specificity results are reported The sensitivity and specificity results closely follow the classification accuracy levels and are similarly affected by the filter bank number The reported sensitivity levels are slightly higher ISBN:

5 than specificity levels which is to be expected for a database containing a greater number of pathological Note that in all experiments to follow, SVM offers a superior performance to ANN and thus, sensitivity and specificity results are reported for SVM only In the next experiment, an attempt is made to establish the diagnostic relevance of individual subbands in isolation from other sub-bands The experiment is based on the optimal filter bank size of six filters Again, the feature vector is constructed using all four statistical measures The classification results are reported in Table 2 Clearly individual sub-bands have very different capacity levels to discriminate between healthy and pathological signals The highest classification accuracy levels, 886% for SVM and 884% for ANN, are attained for the 5 th sub-band This band is associated with a frequency range of Hz In general, the three sub-bands at a higher end of spectral range give rise to better classification performance results compared to those at the lower end of the spectral range In each case, feature vector has only four elements Compared to the previous experiments, this constitutes a significant reduction in the feature vector size and thus, the corresponding drop in the classification performance is observed The following experiment aims to address the effect of using information from multiple sub-bands on the quality of ECG diagnostics In fact, an attempt is made to ascertain the specific combination of sub-bands that would lead to the highest classification accuracy rate, while still using all four statistical measures to construct the feature vector Although, exhaustive evaluation of sub-band combinations is performed, only the most relevant results are reported in Table 3 The results demonstrate that a direct combination of sub-bands with the highest diagnostic relevance, as evaluated in the previous experiment, does not necessarily lead to a significant increase in the classification accuracy rate Individually, 5 th and 6 th bands have the highest diagnostic relevance, but the combined set of features from the two bands does not lead to notable improvement in the accuracy rate compared to the case when the fifth sub-band is used on its own (889% vs 886% for SVM) To put these values in a context of the first experiment, SVM classification accuracy rate is almost 5% lower compared to the instant when all six sub-bands are employed TABLE : PERFOMANCE VS FILTER BANK SIZE Number of Filter Banks in the range -64 Hz N=4 N=5 N=6 N=7 N=8 N= N=2 No Features ANN Acc [%] SVM Acc [%] SVM Se [%] SVM Sp [%] TABLE 2: PERFOMANCE FOR INDIVIDUAL SUBBANDS Sub-bands Banks Fr range (Hz) No Features ANN Acc [%] SVM Acc [%] SVM Se [%] SVM Sp [%] TABLE 3: OTIMIZING SUBBAND SELECTION Sub-bands Banks No Features ANN Acc[%] SVM Acc[%] SVM Se [%] SVM Sp [%] TABLE 4: PERFOMANCE FOR VARIOUS CHOICES OF STATISTICAL MEASSURES Banks Statistical Measures var var kurt var var kurt No Features ANN Acc[%] SVM Acc[%] SVM Se [%] SVM Sp [%] TABLE 5: OPTIMIZING SUB-BAND SELECTION AND THE CHOICE OF FEATURE VECTOR MEASSURES Sub-bands Banks Meassures var kurt No Features ANN Acc[%] SVM Acc[%] SVM Se [%] SVM Sp [%] ISBN:

6 A similar observation is made when the features from spectrally adjacent sub-bands, eg 5 th and 4 th, are combined Combining the adjacent sub-bands did not lead to an improvement the overall accuracy rate On the other hand, when non-adjacent subbands are combined, higher improvements in the classification performance are observed, eg combination of 4 th and 6 th sub-bands and 5 th and st sub-band pairing These observations suggest that there exists a significant amount of redundancy associated with diagnostically relevant information between the individual sub-bands and especially, between the adjacent sub-bands A similar observation is made when combinations of three sub-bands are used to form ECG feature vectors The highest classification accuracy rate of 978% is reported for a SVM classifier and a feature vector involving a combination of st, 3 rd, and 5 th In comparison to the case when all sub-bands are used, this result constitutes an increase of 3% in the accuracy rate and a 5% reduction in the feature vector size In a persistent attempt to reduce the feature vector size and to improve the classification performance, the diagnostic relevance of individual statistical parameters, in isolation and combination is considered in the following experiment The experiment is based on all six sub-bands being used to construct the ECG feature vectors The results are reported in Table 4 When individual statistical measures are considered in isolation, attains the highest accuracy rate, 957% for SVM, which is slightly higher than when all four statistical measures are used in combination, see Table This observation could be explained by the high capacity of this particular statistical measure to describe the diagnostically relevant information and an associated significant decrease in the feature vector size, which allows improved classifier training Other statistical descriptors attain a similar accuracy rate of about 9% for the SVM classifier and individually they show a significant capacity to represent the diagnostically relevant information in the ECG signal In an exhaustive search where the statistical descriptors are used in combination for all six sub-bands, the highest classification accuracy rate, 976% for SVM, is attained for a 2-parameter long feature vector involving the and pairing In the final experiment, an attempt is made to find the optimal combination of the statistical descriptors and frequency sub-bands that gives rise to the highest classification accuracy rate All combinations of sub-bands and statistical measures are evaluated, but the most significant results are reported in Table 5 The highest classification accuracy rate, 988% for SVM, is obtained for the combination of st, 3 rd, and 5 th frequency sub-bands, and with and as the principal statistical descriptors The corresponding feature vector corresponds to the optimal frequency sub-band combination and the optimal statistical descriptor combination evaluated independently in the previous experiments In comparison to the case when all six sub-bands are used together with all four statistical descriptors, the optimized feature vector denotes an improvement in the classification accuracy rate, from 949% to 988%, and a significant feature vector size reduction, from 24 parameters to 6 parameters The decrease in the dimensionality of pattern representation has implications in the classification speed increase and lowering of computational cost 5 Conclusion This paper presents a systematic study of feature vectors derived from the filter bank processing of ECG signal for the diagnostic applications of heart conditions, and in particular Supraventricular Arrhythmia Here, the considered ECG feature vector is based on a set of statistical descriptors (, variance, ness and kurtosis) of the energy distribution in the individual filter banks The quality of feature vector, in terms of its capacity to differentiate between pathological from healthy ECG signals, is evaluated using the Artificial Neural Networks and Support Vector Machines The results have shown that the classification accuracy is strongly affected by the choice of filter bank number A filter bank involving six filters that uniformly partition the spectral range -64Hz is shown to be the optimal filter bank size The diagnostic relevance of each statistical parameter and individual sub-bands is studied, in isolation and in combination The results show that the individual frequency sub-bands have very different capacity levels to discriminate between healthy and pathological signals The highest discriminatory capacity is found in the band associated with the Hz range It is demonstrated that a direct combination of sub-bands with the highest individual diagnostic relevance does not necessarily lead to a significant increase in the classification accuracy rate A significant amount of redundancy in diagnostically relevant information is found to exist between the individual sub-bands and especially between spectrally adjacent sub-bands In a study of the diagnostic relevance of the ISBN:

7 considered statistical parameters, it is ascertained that of sub-band energy distribution is associated with the highest diagnostic accuracy rate The optimal feature vector is found, to be a combination of st, 3 rd, and 5 th frequency sub-bands with and as the principal statistical descriptors In comparison to the case when all six sub-bands are used together with all four statistical descriptors, the optimized feature vector denotes an improvement in the classification accuracy rate, from 949% to 988%, and the 75% feature vector size reduction, from 24 parameters to 6 parameters The decrease in the dimensionality of pattern representation has implications in the classification speed increase and lowering of computational cost Classification accuracy, specificity and sensitivity rates attained on the proposed ECG feature vector, under the support vector machine classification framework, indicate that the proposed method could be used as a part of the processes for monitoring and diagnosing heart conditions References: [] PE McSharry, GD Clifford, L Tarassenko, LA Smith, A Dynamical Model for Generating Synthetic Electrocardiogram Signals, IEEE Trans On Biomed Eng Publishing House, 5(2), 23, pp [2] CN Jean, NA Amine, Hilbert Transform Based ECG Modeling, Biomedical Engineering, Springer Link, vol 39(3), 25, 36 4 [3] A Martusevičienė, Z Navickas, A Vainoras, ECG Data Analysis Using the Convolution of Mealy and Moore Automata, Electronics and Electrical Engineering Kaunas: Technologija, 4(), 2, pp 3 6 [4] R Jan e, A Blasi, J Garc ia, and P Laguna, Evaluation of an automatic threshold based detector of waveform limits in Holter ECG with QT database, In Computers in Cardiology, IEEE Press, 997, pp [5] H Chaouch, KOuni, L Nabli, Segmenting and supervising an ECG signal by combining the CWT PCA, IJCSI International Journal of Computer Science Issues, vol 9(), 22, pp [6] MD Ortigueira, Archetypal ECG Analysis, Proceedings of the th Portuguese Conference on Pattern Recognition, Proc RECPAD'98, March 998, pp [7] S Graja and JM Boucher, Multiscale hidden Markov model applied to ECG segmentation, In WISP 23: IEEE International Symposium on Intelligent Signal Processing, Budapest, Hungary, 23, pp 5 9 [8] A Koski, Modelling ECG signals with hidden Markov models Artificial Intelligence in Medicine, 8, 996, pp [9] U Qidwai, and M, Shakir, Filter Bank Approach to Critical Cardiac Abnormalities Detection using ECG data under Fuzzy Classification, International Journal of Computer Information Systems Industrial Management Applications (ISSN: ), vol 5, 22, pp [] U Qidwai, and M, Shakir, Embedded System Design with Filter Bank and Fuzzy Classification Approach to Critical Cardiac Abnormalities Detection, IEEE Symposium on Industrial Electronics and Applications, Indonesia, 22 [] AG Ramakrishnan, S Saha, ECG compression by multirate processing of beats, Computers and Biomedical Research Journal, vol 29(5), 996, pp [2] VX Afonso, WJ Tompkins, TQ Nguyen, K, Michler, and S Luo, Comparing stress ECG enhancement algorithms: with an introduction to a Filter Bank based approach, IEEE Eng in Med and Biol Mag, vol 5, no 3, 996, pp [3] BE Boser, I Guyon, and V Vapnik, A training algorithm for optimal margin classifiers, In Proceedings of the Fifth Annual Workshop on Computational Learning Theory, 992, pp [4] V Vapnik, The Nature of Statistical Learning, Theory NY: Springer-Verlag, 995 [5] S Haykin, Neural Network and Learning Machines, Ed 3, Prentice Hall, 29 [6] Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PCh, Mark RG, Mietus JE, Moody GB, Peng CK, Stanley HE PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals Circulation (23):e25-e /e25]; 2 (June 3) PMID: 8528; doi: 6/CIR23e25 ISBN: