A Bottom-Up Design of Neural Network Based ECG Beat and Rule-based Rhythm Classifier

T Srikanth^a, SA Napper^b, H Gu^c

^aCardiac Science Inc., Irvine, CA-92606, USA
^bDepartment of Biomedical Engineering, Louisiana Tech University, Ruston, LA-71270, USA
^cDepartment of Mathematics/Statistics, Louisiana Tech University, Ruston, LA-71270, USA

Correspondence: T Srikanth, Cardiac Science Inc., 16931 Millikan Avenue, Irvine, CA-92606, USA.
E-mail: sthiagarajan@cardiacscience.com, phone +1 949 567 9066, fax +1 949 247 4013

Abstract. The focus of neural network based ECG classifiers has been on narrow clinical problem domains. An optimal utilization of frequency domain information was missing. The objective of the present work is to improve the accuracy of neural network-based single-lead (lead-II) ECG beat and rhythm classification. A bottom-up approach defined in terms of perfecting individual sub-systems to improve the over-all system performance is used. Sub-systems include pre-processing, QRS detection and fiducial point estimations, feature calculations, and pattern classification. Inaccuracies in time-domain fiducial point estimations are overcome with the derivation of features in the frequency domain. The entire data and problem are divided into four major groups, each group with inter-related beat classes. Classification of each group into related sub-classes is performed using smaller feed-forward neural networks. Optimal implementations of feed-forward neural networks provide an accuracy of more than 85% for all 13 classes included in the study. The system shows a graceful degradation in performance with increasing noise, as a result of the noise consideration in the design of every sub-system. Results indicate a neural network-based bottom-up design of single-lead ECG classification is able to provide very high accuracy, even in the presence of noise, flutter, and fibrillation.

Keywords: Back Propagation; Feature Set; Neurons; Pattern Classification; Electrocardiogram; Feed-forward

1. Introduction

The electrocardiogram (ECG) has been the major diagnostic tool for cardiologists and the ECG signal provides almost all information about the electrical activity of the heart. Three major processes in any automated ECG interpretation system involve ECG signal pre-processing, feature extraction, and pattern classification. An optimal feature extraction methodology described in the earlier article is followed up in the present article with a bottom-up approach to the design of neural network based ECG classifier [Srikanth et al., 2002].

Early developments in pattern classification were dominated mainly by statistical and syntactic approaches. Elghazzawi and Geheb [Elghazzawi and Geheb, 1997] pointed out a major problem associated with heuristic rule-based syntactic approaches in an important critique. They found that the problem of distinguishing between normal (N) and ventricular (V) beats belongs to the non-linearly separable class. When AND-OR binary structures with hand-tuned thresholds or linear-separation techniques are used to separate N and V distributions in a K-dimensional feature space, errors are guaranteed because these distributions are not linearly separable. As a result, these algorithms have a limited dynamic range. The designers of these techniques always faced the dilemma of tuning the algorithms to be more sensitive at the expense of achieving less positive predictivity, resulting in a high rate of ventricular false-positives and vice versa. Such problems exist through out the ECG interpretation task [Rautaharju et al., 1992].

Compared to the disadvantages of syntactic rule-based programs, systems based on statistical methodologies present a different set of problems. They start with the assumption of Gaussian characteristics for the signal as well as the features derived from ECG signal. Discrete ECG features may not always obey Gaussian distribution as suggested in a review of bio-signal interpretation by Ciaccio et al. [Ciaccio et al., 1993]. Another problem in statistical pattern classification relates to the complexity of the problem domain. With more output classes and input features, linear discriminant rules begin to fail. Non-linear statistics are not yet fully understood, and this fact is especially true in a clinical environment.

In the last two decades, a third methodology of pattern recognition, namely artificial neural networks (ANN), has become quite popular, and several successful implementations with respect to ECG pattern recognition have been reported [Baxt, 1992; Silipo et al., 1995]. Unique applications such as detection of lead reversals in ECG recordings have been tried successfully using ANNs [Heden, 1996]. Heden et al., 1997, have demonstrated the performances of ANNs in predicting heart attack to be at least 10% superior to an individual expert.

So far, implementations of ANN-based ECG classification schemes have been restricted to problems of narrow clinical domains, and the focus has been on applications rather than on design concepts or choice of better features. With newer knowledge available from the theory of ANNs and with the variety of accurate features, the stage is now set for moving towards a bottom-up approach in designing artificial neural network-based clinical ECG interpretation systems. A bottom-up approach is defined in terms of need-based implementation of component sub-systems; i.e., QRS complex detection, fiducial point estimation, feature extraction, beat classification, and rhythm identification. In this article, the focus is on the final classification of beats and rhythms using the feature extraction approach suggested in the previous article.

2. Methods

The overall block diagram for the entire schematic is provided in Fig. 1 of the earlier article [Srikanth et al., 2002]. In the present work, normal arrhythmia conditions, and morphological variations like inverted T waves, noise, and artifacts are identified. ECG beats from the MIT-BIH database and from clinical recordings were classified into one of the following classes; normal (NOR) beats, premature ventricular contraction (PVC), atrial premature contraction (APC), left bundle branch block (LBBB), right bundle branch block (RBBB), nodal (JUN), premature beat (NODE), paced beat (PCE), ventricular escape beat (VEB), noisy beat (NOI), inverted T-wave (INVT), and ST segment slope (ST), atrial fibrillation (AFIB) and atrial flutter (AFL). Features are calculated from each beat in both time and frequency domains and the entire feature extraction methodology is explained earlier in [Srikanth et al., 2002]. Definitions for all the new features in time and frequency domains are described in the same article.

2.1. Data Set and Input Feature Set

Input features are organized as a matrix of the order [M´N] where, M= number of beats and N= number of features/beat. For each beat, the organization of the features and organizational details are as follows:

F=[Feature(t); Feature(f); Data index; Beat-index; Group index; Sub-class index]

(1)

F=[1´16; 1´11; 1´1; 1´1; 1´1; 1´4]

(2)

where
                Feature(t) = time-domain feature set,
                Feature(f) = frequency-domain feature set.

The overall number of columns thus equals 34. The entire set of features may not be useful for classifying all beat classes chosen, and only a sub-set is chosen for an individual group of classes. Table 1 in [Srikanth et al., 2002] provides the details regarding the final data set chosen for analysis. Data segments include 430 segments from MIT-BIH data set and 65 from clinical data set. A total of 3854 beats were analyzed from the total of 495 data segments. ECG data segments varied in duration from six to ten seconds.

2.2. Neural Network Implementation

Multi-layer feed forward neural networks (perceptrons) are the most used artificial neural networks for pattern recognition tasks. In the present work, perceptrons are used for classifying an ECG beat and a rule-based system is designed for relating the classified beats to the closest rhythm statements. Two options are possible in solving the classification problem. The first option involves using a single feed-forward neural network, which should be able to classify the chosen thirteen ECG classes, as listed in Table 1. However, this specification leads to a large neural network with more hidden neurons and internal connections proportional to multiples of number of inputs. The major problem in using a single feed-forward network seems to be its inability to adjust to wide variations in feature values. Another peculiar problem in ECG feature classification is the accuracy of features and their relation to exact location of fiducial points. Exact location of fiducial points is difficult in beats belonging to atrial fibrillation, flutter, ventricular beats, and in noisy beats. Hence, an accurate set of time-domain features is not available for all thirteen classes.

In the present work, the second strategy is followed. The problem is designed in such a way that an optimal combination of smaller feed-forward networks may work on groups of smaller set of similar beats. A strategy is employed where time-domain features are used for classifying mostly supra ventricular arrhythmia, and a combination of time and frequency domain features are used for other classes including noisy beats. Such an arrangement of dividing the problem into smaller sub-problems helps to utilize the inherent ability of feed-forward networks to learn the minor variations of feature values and generalize.

Sub-Problems and Smaller Feature Sets

The whole set of beat classes is divided into four groups as shown in Table 1. Four different sets of features are chosen for identifying each of three/four beat classes among each group. The indices for features are as indicated in earlier article on time and frequency domain features [Srikanth et al.,2002]. The initial sets of features had eleven features for each of four groups, selected based on the Minnesota Coding by Blackburn [Blackburn, 1999], Novocode by Rautaharju et al. [Rautaharju et al., 1998] as well as by clinical suggestions by Wagner [Wagner, 1994] are also used for the selection of appropriate features. Frequency domain features are selected based on Minami et al., [Minami et al., 1999] Strohmenger et al., [Strohmenger et al., 1997], and Schkurovich et al., [Schkurovich et al., 1998]. The final feature sets used to classify with in each group are shown in Table 2.

Table 1.    Division of data into four major groups with their sub-classes.

Groups 1 and 3 consist of beats with clearly identified fiducial points and intervals. Hence, the beats are classified purely based on time-domain features. On the other hand, groups 2 and 4 constitute beats with errors in the detection of all the fiducial points. Hence, frequency domain features are added in addition to those time domain features least affected by errors in fiducial point calculations. Accuracy of fiducial point calculations in group 4 is extremely low and hence frequency domain features play a dominant role. Values of features show clear variations across different groups. ST/T abnormalities are calculated only in clinical data with adequate bandwidth of (0.05-100) Hz and are neglected in MIT-BIH data set, which has a bandwidth of (0.1-100) Hz [Moody, 1990].

Table 2.    Set of features for classification inside each group.

Practical Considerations and Strategy

Selection of a final set of features, number of hidden neurons, and training goal in mean square error have always been an art in neural network implementations. Useful strategies included the selection of features, showing sufficient variation between three or four sub-classes inside each group (using t-test). Another approach followed throughout is the use of scatter plots. A sample scatter plot is shown for variations of the ratio of upward vs. downward slopes (T11) for the classes in Group 1 is shown in Fig. 1.

Initially, a set of eleven features is chosen for classifying within each group. The initial selection is based on the literature and the scatter plots. Smaller sets of 6-8 features are chosen from the initial eleven features, in iterative fashion and performance of the network in training is monitored with this smaller set. The sets of features providing better performance in training are chosen. Inputs to the final chosen networks contain the least number of features (Table 2). Features chosen are unique for each group and hence, a new set of feature values corresponding to an input ECG beat, will match only with a particular group.

Selection of the number of hidden neurons for a particular task requires practical experience in the absence of established theory. An ideal value is to choose the number of hidden neurons such that they satisfy the following two competing requirements: a very low mean square error (MSE) in training and a very high generalization ability for new set of test data.

Figure 1.   Scatter plot of the distribution of the values of the ratios of upward versus downward slopes of QRS complex in beat classes in Group 1– NOR, RBBB and LBBB.

An optimum mean square error target for training seems to be in the range 0.10-0.20, based on the considerations of over-fit and loss of generalization abilities. The network weights are sufficiently flexible and the testing performance is usually better [Haykin, 1999]. The neural network structure providing the lowest mean square error compared to the other structures is chosen as the final choice. Steps below indicate the procedure followed in the present work to design a perceptron.
      Step 1: Pre-process the input set of features (Normalization to the range [-1,1]).
      Step 2: Choose the initial set of features.
      Step 3: Select features -- essential and non-essential features based on t-tests and scatter plots for each feature.
      Step 4: Train with adequate data set and hidden neurons for different combinations of features.
      Step 5: Choose the best possible set of features based on the ability to reach the goal mean square error (MSE), by an iterative training process.
      Step 6: Vary the number of hidden neurons and hidden layers for the chosen feature set.
      Step 7: Choose the best network for each group in terms of low number of neurons, and an ability to reach the goal MSE.
      Step 8: Final choice is based on generalization ability of the network to the test data.
      Step 9: Test performance for noisy data.

The present study used the hyperbolic tangent sigmoid transfer function for hidden neurons and the linear transfer function for output layers. Levenburg-Marquadt training provides an adaptive optimization in back-propagation and the number of training steps reduces to one-tenth to one-hundredth of traditional procedures like conjugate gradient and steepest descent algorithms with fixed step-size [Hagan et al., 1995].

Training and testing data sets are chosen adequately large, so that each epoch contains a minimum of 5-10 times the number of connections between the input and hidden neurons. Some of the selected networks are found to lack generalization capability. The reason for such cases is that the training is based on a very small data set and hence, the network has to be re-trained with a much larger training set. An ideal training data set should contain enough representative beats from all the classes.

Rule-based System for Rhythm Classification

Detection of beat classes is followed by rhythm identification, which indicates the overall statement for the observed duration. For example, a rhythm statement for the following beats NHR-NOR-NOR-NOR-NOR-NOR-NOR should indicate normal sinus rhythm (NSR) and a data segment with beats NHR-NOR-PVC-NOR-PVC-NOR-PVC should indicate ventricular bigeminy (VBIG) and so on. The task is essentially associating the information from beat classifier output to the rhythm class output. Such a task takes into account two inputs: a code based on heart rate, and a set of codes corresponding to beats. Rhythms in all data segments were annotated by at least two independent cardiologists. Neural networks seem to be ill-equipped to solve such deterministic problems.

This problem is essentially a deterministic one with well-defined rhythm classes. Binary coding schemes are used for the individual beat classes and for indicating the heart rates. Outputs are designated in one of the fifteen rhythm classes. Any variation from one of the pre-defined classes is treated as unknown (UNK) class and the entire beat structure is given as output. A sample-coding scheme for a few rhythm classes is presented in Table 3. The first two bits indicate the information regarding the heart rate, normal (NHR), bradycardia (BC), and tachycardia (TC) and very low heart rates (VLHR). With thirteen beat classes detected in present scheme, four bits are required for representing them. The representations for beat classes, rhythm classes and heart rate classes are shown below. Sometimes, more than one input pattern can represent the same output rhythm, as shown for the rhythm classes ventricular bigeminy (VBIG) and ventricular tachycardia (VTACH) in Table 3. A set of forty different rules is used to relate the beat and heart rate information to one of the rhythm classes. This scheme is flexible and more rules can be added to indicate the precise relation between the beat classes and the rhythm class.

In implementation, an assumption is made about the availability of six beats. In cases of long data records or greater number of beats, the first six beats were taken. On the other hand, in cases of very slow bradycardia, a replication is done in order to get the six beat class information. For example, in sinus bradycardia, NOR-NOR-NOR-NOR is taken as NOR-NOR-NOR-NOR-NOR-NOR and in ventricular bigeminy, NOR-PVC-NOR-PVC is taken as NOR-PVC-NOR-PVC-NOR-PVC, a replication for last two beats.

Beat classes:

Heart rate classes:

Rhythm classes:

Table 3.    Sample input pattern for selected rhythm classes.

System Evaluation

Two different evaluation procedures are attempted. Around one quarter of the original data is chosen for testing. The initial testing is with respect to the training data and how well the network adjusts for the set. Testing is also continued further with a newer set of data, different from the training data. A few cases of simulated beats provide an ideal set of test data. Depending on the performance, some of the training steps are implemented again. In the present case, random noise is added to a collection of ECG signals and fed as input to the system for evaluation. Gaussian random noise is generated using Matlab®-Simulink and the amplitudes are normalized with respect to the amplitude of the ECG signal. Hence, three sets of evaluation results are available for the selected network, for different sets of inputs, namely (i) training data, (ii) test data, and (iii) noise data.

3. Results of Final Beat and Rhythm Classification

Feed-forward artificial neural networks (ANN) require an extensive and well-classified ECG database for both training and testing. Initial neural network implementations provided clues to the required training data set for adequate generalization. Implementation of ECG beat and rhythm classification with neural networks in present work involves analyzing four steps: choice of an optimal network, beat classification results, testing with noisy data and rhythm classification results.

3.1. Results in Beat Classification

Table 4 presents the performance of chosen networks for each group and their performances in testing. Performance is defined as percentage of the training and testing cases successfully classified. Initial attempt with training data revealed a couple of problems like (i) Inadequate number of samples for few beat classes, dominated by a major beat class in each group and (ii) Training did not involve a thoroughly mixed data in each epoch.

To overcome the above problems, additional beats belonging to minority classes were added in training and a thorough randomization is performed before each epoch. The total number of training and testing data is kept same for comparison. Beat classes with inadequate number of samples, i.e., ventricular escape beats (VES), Wolf-Parkinson-White (WPW) pre-excitation beats, Nodal beats (JUN) and atrial flutter (AFL) beats, are supplemented with 40, 20, 30 and 20 more beats and the cases belonging to the major beat classes were reduced correspondingly. These additional beats were extracted from MIT-BIH data records and used for training.

Table 4 indicates the performances of the neural networks before and after the adequate sampling of all the beat classes. Structure information indicates the number of neurons in the input, hidden and the output layers. Perf1 indicates the results from the first attempt and Perf2 indicates the results after adequate sampling of all the beat classes. Two neural network structures are chosen for the Groups 1, 3, and 4 and four neural net structures are chosen for classification of beat classes in the Group 2. Choices are based on the best performing networks with low number of neurons and an ability to reach the goal MSE.

Table 4.    Results for feed-forward neural networks in each group with inadequate minority sampling.

3.2. Testing with Noisy Data

Best performing networks from the noise-free testing are finally chosen for further testing with noisy beats. The feature extraction step acts as a buffer, eliminating the effects of non-stationarity of the raw signal and the noise characteristics. A total of 120 noise-free beats (10 for each class, other than noisy beats class) are chosen for the analysis. Each one of the beats is added with additive gaussian noise in steps of 5% with respect to ECG amplitude (2.5% with respect to power), after normalizing signal and noise amplitudes. Matlab-Simulink® toolbox is used to generate the noisy beats. Performance of the fiducial point detection algorithm has already tested for such cases is also noted. Visual criteria are used to identify the errors in fiducial point identification by the algorithm. Table 5 indicates the performance of beat classification neural network with increasing noise level. Performance of the algorithm with increasing noise power is compared with the performance of human experts on an undistorted data. Even with an additive noise power of 12.1% relative to the R point amplitude, the combined neural network system is able to detect and classify 90 beats in one of the twelve chosen beat classes.

Table 5.

3.3. Results of Rhythm Classification

Rhythm classification in the present work is viewed as a deterministic problem, since the rhythm classes essentially derive from heart rate class and the individual beat classes. A hundred percent accuracy is obtained in this scheme. Hence, the accuracy of overall rhythm classification in the present work is only dependent on the classification of individual beats. A variation of more than one beat in the input set of beats compared to one of standard rhythm patterns produce an unknown rhythm output. Such an approach enables the possibility of secondary opinion by human experts.

4. Discussion

The process of organization of data and features into an array structure enables easy implementation with neural networks. It is possible to select data belonging to any particular group/sub-group. Such a selection procedure increased the speed of implementation and enabled the evaluation of more than 200 neural networks with different configurations. Present classification system offers two major advantages compared to related efforts. The first advantage is the inclusion of fiducial point detection and feature extraction systems in implementation and the second one is the ability to include complementary information from frequency domain, a derivation of the first advantage. Feature extraction provides a very good buffer, with respect to the disturbance and noise in the input.

A third minor advantage may be the inclusion of noisy beats in implementation of every sub-system. Performance with respect to noisy data is evaluated at the level of fiducial point detection and feature extraction and also at final beat classification task. As the performance of the overall system mainly depends on the accuracy of the features, the presence of outliers in feature space warrants rejection in the present work.

The main reason assumed for the increased accuracy in the present results seems to be the use of a feature set highlighting each group. A weak link in the selection of features is the absence of comprehensive theory on the selection of features. However, the advantage of easy implementation of neural networks enabled all possible combinations of features to be tested in each group [Srikanth, 2000]. Such an approach seems to provide the neural network equivalent of regression analysis, and helps to determine the feature set with best possible training performance.

The selection of independent features for each group is another concern. In the present work, selection of features of each group is such that the outside beat classes produce completely different outputs for those features. This aspect provides another inherent statistical capability of training in neural networks [Schalkoff, 1992]. Feature sets are selected based on the standards suggested in literature and by regulatory authorities [Wagner, 1994; Rautaharju, 1998; Blackburn, 1999].

Type-C system performance in ECG requires manual feedback for evaluation. Usually a group of experts provide a better gold standard compared to an individual expert. Consideration of noisy data is uniformly discarded in many cases and carefully chosen data with homogenous characteristics are used. Two major advantages of the present implementation over such systems are
      i)        Increased accuracy of features in time and frequency domains.
      ii)       Ease of training and testing a neural network, compared to statistical and syntactic tasks of similar complexity.

Some beats usually have two different annotations. For example, a noisy atrial premature (APC) beat provides two beats for training, one for APC class and another for NOI class. Similarly, a fusion involving a normal and a PVC beat provides two beats for training. Hence, noise is taken into account in training and constitutes around 6% of the total number of beats.

Individual relations of performance for particular beats are not analyzed. However, one major problem is in distinguishing the APC beats compared with normal beats. Group 2 always provides the best performance, due to relatively clear distinction in feature space, i.e. VES, PVC and PACED beats. Hence, APCs are added to the group 2. Another major problem is in distinguishing the ST and T wave morphologies from normal beats. Mismatch among groups occurs for this classification.

5. Conclusions

Single lead ECG beat and rhythm classification constitutes a vital step towards the implementation of Type-B and Type-A systems. Speed and superior performance provided by neural networks make the online multi-lead applications possible. A bottom-up design to pattern classification using artificial neural networks is expandable to other signals also. Signals like electroencephalogram (EEG) from scalp electrodes and electromyogram (EMG) have been studied in frequency domain and features can be extracted to study multi-channel patterns.

A theoretical question that can be studied is the one related to the choice of features. A choice of features can be identified iteratively first and a theory can be developed based on the choice. Another development can be the integration of clinical data along with ECG data. Based on above two comparisons, a clear case exists for further expanding the approach to typical multi-lead problems, e.g. in distinguishing size and location of myocardial infarction or location of hypertrophy. Type-A and B systems are easy to develop from present systems, as the syntactic, rule-based systems provide the essential information on the choice of features, choice of leads and their inter-relation. Groups of dependent neural networks for each lead are an easy way to visualize the solution. The present bottom-up approach provides an extendable approach in terms of time of recording, number of leads and problem space.

Acknowledgements

The authors wish to acknowledge Drs. D. Prabhakar of Madras Medical College, Chennai, India and K.P. Mishra, Apollo Hospitals, Chennai, India for their help in ECG beat classification tasks and Ms. Nandhini for the documentation help.

References

    Baxt WG. Analysis of clinical variables driving decision in an artificial neural network to identify the presence of myocardial infarction. Annals in Emergency Medicine, 21: 1439-44, 1992.

  Blackburn H. History of ECG coding, http://www.epi.umn.edu/Ecg/Pages/history.html ,1999.

  Ciaccio EJ, Dunn SM, Akay M. Bio-signal pattern recognition and interpretation Systems. IEEE-EMBS Magazine, 89-97, Sept, 1993.

  Elghazzawi Z, Geheb F. Critique of arrhythmia detectors based on heuristic rules. Biomedical Instrumentation & Technology, 31: 263-271, 1997.

  Hagan MT, Demuth HB, Beale M. Neural network design, PWS Publishing Company, 1995.

  Haykin. S. Neural networks – a comprehensive foundation. 2nd Edition, Prentice Hall, NJ, 1999.

  Heden Bo, Ohlin H, Rittner R, Edenbrandt L. Acute myocardial infarction detected in the 12-lead ECG by artificial neural networks. Circulation, 96: 1798-1802, 1997.

  Heden Bo. Analysis of electrocardiograms using artificial neural networks. Doctoral Dissertation, Department of Clinical Physiology, Lund University, Sweden, 1996.

  Minami K, Nakajima H, Toyashima T. Real-time discrimination of ventricular tachyarrhythmia with fourier-transform neural network. IEEE Transactions on Biomedical Engineering, 46: 179-185, 1999.

  Moody GB. MIT-BIH Arrhythmia Database Directory, Harvard University-MIT Division of Health Sciences and Technology, 1992.

  Rautaharju PM, Calhoun HP, Chaitman BR. NOVACODE Serial ECG classification system for clinical trials and epidemiologic studies. Journal of Electrocardiology, 24(Supplement): 179-187, 1992.

  Rautaharju PM, Park LP, Chaitman BR, Rautaharju F, Zhang Z. The novacode criteria for classification of ECG abnormalities and their clinically significant progression and regression. Journal of Electrocardiology, 31: 157-187, 1998.

  Schalkoff R. Pattern recognition--statistical, structural and neural approaches. John Wiley & Sons, Inc., New York, 1992.

  Schkurovich S., Sahakian AV, Swiryn S. Detection of atrial activity from high-voltage leads of implantable ventricular fibrillators, using a cancellation technique. IEEE Transactions on Biomedical Engineering, 45: 229-234, 1998.

  Silipo R.,Gori M, Taddei A, Varanini M, Marchesi C. Classification of arrhythmic events in ambulatory electrocardiogram using artificial neural networks. Computers and Biomedical Research, 28: 305-318, 1995.

  Srikanth T, Napper SA, Gu H, Bottom-up Approach to Uniform Feature Extraction in Time and Frequency Domains for Single-lead ECG Signal, International Journal of Bioelectromagnetism, 3, 2002.

  Srikanth T. Bottom-up Design of Artificial Neural Network for Single Lead Electrocardiogram Beat and Rhythm Classification. PhD Thesis, Louisiana Tech University, Ruston, Louisiana, USA, 2000.

  Strohmenger HV, Lindher KH, Brown CG. Analysis of the ventricular fibrilation ECG signal amplitude and frequency parameters as predictors of countershock success in humans. Chest, 111: 584-589, 1997.

  Wagner GS. Marriot's practical electrocardiography. Ninth Edition, Williams & Wilkins, 1994.