Classification of electrocardiographic ST-T segments human expert vs artificial neural network

Transcription

1 European Heart Journal (1993) 14, Classification of electrocardiographic ST-T segments human expert vs artificial neural network L. EDENBRANDT, B. DEVINE AND P. W. MACFARLANE University Department of Medical Cardiology, Royal Infirmary, Glasgow, Scotland KEY WORDS: Electrocardiography, artificial intelligence, computer assisted diagnosis. Artificial neural networks, which can be used for pattern recognition, have recently become more readily available for application in different researchfields.in the present study, the use of neural networks was assessed for a selected aspect of electrocardiographic (ECG) waveform classification. Two experienced electrocardiographers classified 1000 ECG complexes singly on the basis of the configuration of the ST- T segments into eight different classes. ECG data from 500 of these ST-T segments together with the corresponding classifications were used for training a variety of neural networks. After this training process, the optimum network correctly classified 399/500 (79-8%) ST-T segments in the separate test set. This compared with a repeatability of 428j500 (85-6%) for one electrocardiographer. Conventional criteria for the classification of one type of ST-T abnormality had a much worse performance than the neural network. It is concluded that neural networks, if carefully incorporated into selected areas of ECG interpretation programs, could be of value in the near future. Introduction The ECG is still one of the most valuable and widely used diagnostic methods in cardiology, despite its recent centenary, but its usefulness depends on the quality of its interpretation, a facility for which can only be gained with experience that, for non-cardiologists, can be difficult to acquire. Against this, the introduction of computers for ECG analysis offered the possibility of improving the diagnostic method. Computerized ECG recorders which produce interpretations of high quality 1 ' 1 are widely used and are of value not least for non-cardiologists. At present, the best computer interpretation programs perform better than the average physician' 2 ', and almost as well as experienced cardiologists' 1 '. Although some ECG analysis programs use statistical methods in order to make interpretations' 1 ', most commercially available ECG analysis programs use a deterministic approach, i.e. they use criteria with specified cutpoints against which the ECG data are tested. These criteria are often based on results from the literature allowing advantage to be taken of experience from many studies, including different subject material. A complete ECG interpretation program is complex, consisting of many rule based criteria, but the reason why a particular ECG interpretation was made can still be determined in such a program. This is an important aspect of improving the performance of program interpretation. The cardiological community has also appreciated that reasons printed out for a particular ECG interpretation are of value, not least from an educational point of view. However, the deterministic approach is not infallible, and a Submitted for publication on 5 May 1992, and in revised form 25 September Correspondence Lars Edenbrandt. Department of Clinical Physiology. University of Lund, University Hospital, S Lund, Sweden. cardiologist can often present a more correct interpretation, though not detailed reasons, e.g. a cardiologist may report 'left ventricular strain' without being able to state precise amplitude criteria. Neither statistical nor deterministic ECG analysis programs use 'pattern recognition' in the literal sense. Artificial neural networks, which can be used for pattern recognition, recently became widely available for applications in different research fields. The introduction of a mathematical algorithm' 3 ' made it possible to develop more advanced software implementations of neural networks. Early research in this field was inspired by the similarities with neural transmission in the human brain hence the term neural networks. The purpose of the present study was to assess the performance of a neural network in an ECG classification task. In addition, the performance of the network was compared to that of a human ECG interpreter and of conventional ECG criteria, respectively. Methods NEURAL NETWORK A neural network processes ECG data in the following way. During a training process, a fixed though unlimited number of measurements for each ECG, e.g. amplitudes and durations of one or more leads, may be input to the network together with the desired classification, which in this study is a type C diagnosis' 41 such as one of several types of ST-T morphology. The network, which 'learns' to associate the training examples with the given classification for each case, will have a fixed number of outputs corresponding to the number of different classifications used for training. After training is completed, the same set of measurements from any new ECG can be fed to the network and a classification produced on the basis of the X $ The European Society of Cardiology

2 Classification of electrocardiographic ST-Tsegments 465 -J D. Table 1 Definitions of group A to H A normal ST-T segment B ST-T segment with ST elevation C ST-T segment with terminal T wave inversion D flat ST-T segment E downward sloping ST segment F ST depression with downward sloping ST segment and asymmetric T wave inversion (left ventricular strain pattern) G downward sloping ST segment and positive T wave H symmetric T wave inversion E F G H Figure 1 Examples of ST-T segments from group A to H. Group A represents the normal ST-T segment and the groups B to H represents different types of ST-T changes to which a verbal definition is given in Table 1. previous training. More complete details are available in the Appendix. ECG DATA The present study was based on resting ECGs recorded from patients at the University Department of Medical Cardiology, Glasgow Royal Infirmary. The 12-lead ECGs were processed with the Glasgow ECG analysis program' 51, which includes routines for analogue-digital conversion of the ECG signals (sampling rate 500 Hz) and calculation of a median beat from which measurements of the waveforms can be produced. The interval between the ST-J point and the end of the T wave is divided into eight parts of equal duration. For the purposes of the present study, the amplitudes at the end of each interval, designated the ^, ^ etc timenormalized ST-T amplitudes were fed to a number of neural networks together with measurements of the ST-J amplitude, ST slope, positive and negative amplitudes of the T wave, i.e. there were 12 input measurements. The ECG complexes were output on a laser printer for visual inspection. For the purpose of this study, a classification task suitable for application of a pattern recognition method was defined. ECG complexes in the lateral leads V 4 -V 6 were classified singly on the basis of the configuration of the ST-T segment into the following eight groups; one group for normal ST-T segments (A) and seven groups for defined pathological ST-T changes (B-H). Examples of ST-T segments from the groups A-H and the corresponding definitions are presented in Fig. 1 and Table 1, respectively. ST-T classification was made independently by two experienced electrocardiographers. The class was assigned by consensus in cases in which the reviewers initially differed. This classification was used as the gold standard. The two electrocardiographers agreed initially on the eight groups of ST-T pattern to be included in the study. Thereafter, the first 1000 ST-T patterns from leads V 4 -V 6 in a consecutive series of ECGs from a local database, were used. Technically unsatisfactory complexes were excluded as were those which clearly did not fit into one of the eight groups. The material was divided into one training group and one test group, each of which consisted of 500 ST-T segments with the same relative distribution between the groups as for the total group. The distribution of ST-T segments among the eight groups is shown in Table 2. STUDY DESIGN ECG data of 500 ST-T segments with corresponding classification A-H were used in the process of training a number of artificial neural networks with different structures (see Appendix) whose performance was assessed using the separate test set. Each ST-T segment was analysed separately without taking into account the lead from which it came or the configuration of the other ST-T segments of the same ECG. The performance of the best network was compared to the repeatability of one of the two electrocardiographers, who made a second classification of the ST-T segments in the test set more than 2 weeks after the first classifications were assigned by the two reviewers together. The second classification was made in a blinded fashion. The performance of the best network was also compared to that of a conventional criterion. Group B was the only group in which ST elevation was part of the definition (Table 1). Therefore, the possibilities for separating ST-T segments of group B from other ST-T segments using a conventional criterion seemed promising. The criterion was developed using the training set, the goal being to find a criterion with the same number of true positive group B ST-T segments, as was found by the best network in order to make the comparison between the methods easier using the test set. At first, the ST-J amplitude alone was tested but improved performance was reached using combinations of ST-J amplitude, ST slope and the positive amplitude of the T wave. The best result in the training set was found using the following criterion: ST-J amplitude > positive T amplitude/5 and [ST-J amplitude >0-4 mv or (ST-J amplitude >005 mv andst slope >0 )] The performance of this criterion was compared to that of the best network using the test set.

3 466 L. Edenbrandt et al. Table 2 Number of correctly classified ST- T segments in the test set. Result of a neural network and the corresponding result of the repeatability study in which one of the two eleclrocardiographers in a blinded fashion made a second classification of the ST-T segments in the test set. For definition of group A to H see Table 1 and Fig. 1 A B C Class D E F G H Total Number of cases Correctly classified by Network Human expert (79-8%) 428 (85-6%) Results A number of networks with different structures were tested with the best result being achieved using a network with 12 input measurements, consisting of the ST-slope, the ST-J amplitude, the positive and negative amplitude of the T wave and eight time normalized amplitudes of the ST-T segments. After the training process, this network correctly classified 399 (79-8%) of the 500 ST-T segments in the test set. The result was best for group A and worst for group C (Table 2). These were the classes with the greatest and smallest number of examples in the training set, respectively. The performance of the network was compared to the repeatability of one of the electrocardiographers. A total of 428/500 (85-6%) ST-T segments in the test set was classified in the same way the first and the second time by the electrocardiographer. Forty-three of the 72 ST-T segments incorrectly classified by the electrocardiographer were correctly classified by the network. The opposite was true in 72 out of 101 cases. Both the electrocardiographer and the network correctly classified 41 of the 47 ST-T segments of group B in the test set. The conventional criterion for group B developed in the training group, correctly classified 40/47 group B cases and 307/453 non-group B cases in the test set. The specificity for the conventional criterion, the network and the electrocardiographer were 68% (307/453), 97% (438/453) and 98% (444/453), respectively. Thus, with a similar sensitivity, the specificity for the conventional criterion was significantly lower than that of the electrocardiographer and the network. Discussion In the U.S.A., an estimated 100 million ECGs are now processed annually 16 ' with computerized recorders, which output the ECG signals together with an interpretation. Although most of these interpretations are satisfactory, even a 5% error represents 5 million misinterpretations. A cardiologist can often correct an erroneous computer interpretation but not explain his interpretation in terms of simple criteria which can easily be translated into new improved criteria for an ECG analysis program. This type of feed-back would otherwise be useful in order to improve the performance of the interpretation programs. However, the interpretation of a cardiologist is not based on a single observation in a particular ECG but on many findings and, most important, the relationships between them. The situation is often even more complex as the cardiologist may take other clinical data into account and may be inconsistent in his/her subjective diagnosis' 11. Because of the complexity of ECGs, it is appealing to try artificial neural networks, in order to improve computerized ECG interpretations. In a recent study by Bortolan et a/.' 7 ' a neural network was used to classify ECGs into seven diagnostic groups, including normal, ventricular hypertrophy and myocardial infarction of different location. After the training process, using 46 ECGs, the network correctly classified 80-9% of normal ECGs. The corresponding figure for the total test group was 66-3%. Even though the network produced only seven different outputs, which is much less than what is demanded from a complete interpretation program, the performance of the network, especially the specificity, is presently much too low to be clinically useful. The importance of the size and composition of the training set for the performance of a network has been demonstrated' 8 ' and the number of ECGs required to train a network to classify ECGs into all possible diagnoses demanded from a complete interpretation program, capable of analysing ECG from patients of all ages and including combinations of diagnoses, with a good performance, is likely to be very high. A different approach to using neural networks in ECG interpretation has been adopted in this study. Neural networks can be incorporated into conventional interpretation programs in order to improve those parts of the analysis which have proved to be difficult to handle using conventional criteria. ST-T segments seemed to represent an ideal testing ground. Most ST-T abnormalities are classified using Type C statements' 4 ', i.e. there are no ECG independent reference methods. The gold standard of this study was the classification of two experienced electrocardiographers. A problem can arise using this type of reference method if similar configurations are classified into different groups by the electrocardiographers. The performance of the method will be impaired with increasing inconsistencies of the gold standard. This situation is likely to be more common if the material used consists of many borderline cases. The repeatability of one of the two electrocardiographers was studied in order to assess the consistency of the reference method. The electrocardiographer classified 85-6% of the ST-T segments in the test

4 Classification of electrocardiographs ST-T segments 467 set in the same way the first and the second time. This result showed that many of the ST-T segments were difficult borderline cases. However, the overall performance of the network was only slightly worse than the human expert, which is a promising result. Furthermore, the network correctly classified 43 of the 72 ST-T segments which the electrocardiographer, whose repeatability was assessed, classified in a different way the second time, indicating that the network also classified difficult borderline cases relatively well. Input measurements to the neural network were taken directly from the measurement matrix of the Glasgow Analysis Program' 51. As such, they were subject to error in the reference used for amplitude measurement and the fiducial points used for defining the end of QRS and the end of T wave, i.e. the interval which was divided into eight equal time normalized segments. The present study did not address the question of noise affecting variation in measurements. It was the hope of the authors that neural networks would be more resilient to slight errors in measurement which possibly, with a first generation ECG program, might have been sufficient to cause an amplitude threshold to be crossed and a different diagnosis made. Neural networks do not work on this basis, but a separate study would be required to answer the question of the influence of noise on network classification. The selection of the eight ST-T morphologies, as described in Table 1, was based on a study of the varying ST-T configurations seen in routine practice. The actual interpretation which a cardiologist may place on these morphologies is not of any concern in the present study where the objective was to assess whether or not a neural network could differentiate one pattern from the others. I n other words, whether the T wave inversion of group H is due to myocardial ischaemia or was caused by pericarditis, is not of relevance in this study. What is important is whether the neural network could detect that pattern H was different from pattern F, as it is likely that most cardiologists would agree that these two patterns were of different aetiologies. In fact, of the 42 patterns classed as group H by the electrocardiographers, five were categorised as group F by the network. On the other hand, the electrocardiographer on repeat testing also misclassified three of the 42 group H patterns as group F, indicating the presence of borderline cases. This emphasises the fact, as is well known, that not only is there inter-observer variation but clearly, there is intraobserver variation as discussed above. This is what makes the task of those who design computer software for interpretation of the ECG so difficult because it is not possible to 'please all cardiologists all of the time'. Designers of deterministic computer logic cannot provide descriptions of an infinite number of ST-T patterns, although they may accommodate the different morphologies to a greater extent than is done by statistically based ECG analysis programs. In other words, despite the variation in the interpretation of ST-T changes by different physicians, there still remains the problem of providing a finite number of classifications by a computer program and it is this problem which the present study has addressed. The neural network has performed relatively well with respect to the gold standard given the difficulty of classifying the ST-T patterns as evidence by the 85-6% repeatability of the electrocardiographer. It should be said that the two electrocardiographers were from different centres and, therefore, were trained in different ways, but notwithstanding this, each was asked to classify an ST-T morphology into one of the eight categories A to H so that the question of interpretation of the pattern or indeed of the background of the experienced reviewer is essentially irrelevant. At present, conventional criteria are used in interpretation programs for the analysis of ST-T segments. Neural networks must prove to be better than such criteria to be of any interest for computerized ECG analysis. Therefore, the comparison between the network and the conventional criterion for the classification of ST-T segments of type B was of interest. The performance of the conventional criterion was much worse than both that of the neural network and that of the human expert. These results are promising, indicating that neural networks could prove to be useful tools if incorporated into conventional interpretation programs. Dr Edenbrandt was supported during his research work in Glasgow in part by grants from the Swedish Medical Research Council and from the Faculty of Medicine, University of Lund, Sweden. References [1] Willems JL, Abreu-Lima C, Arnaud P el al. The diagnostic performance of computer programs for the interpretation of electrocardiograms. N Engl J Med 1991; 325: [2] Jakobsson A, Ohlin P, Pahlm O Does a computer-based ECGrecorder interpret electrocardiograms more efficiently than physicians? Clin Physiol 1985; 5: [3] Rumelhart DE, Hinton GE, Williams RJ. Learning internal representations by error propagation. In: Rumelhart DE, McClelland JL, eds. Parallel Distributed Processing, Vol 1. Cambridge, MA: MIT Press, 1986: [4] Rautaharju PM, Anet M, Pryor TA el al. Optimal Electrocardiography. Task Force III: Computers in diagnostic electrocardiography AmJCardiol 1978; 41: [5] Macfarlane PW, Devine B, Latif S, McLaughlin S, Shoat DB, Watts MP. Methodology of ECG interpretation in the Glasgow program. Method Inform Med 1990; 29: [6] Macfarlane PW. A brief history of computer assisted electrocardiography. Method Inform Med 1990; 29: [7] Bortolan G, Degani R, Willems JL. Neural networks for ECG classification. In: Ripley KL, Murray A, eds. Computers in Cardiology Washington: IEEE Computer Society Press, 1991: [8] Edenbrandt L, Devine B, Macfarlane PW. Neural networks for classification of electrocardiographic ST-T segments. J Electrocardiol 1992; 25: Appendix Artificial neural networks are simple structures consisting of interconnected units called neurons as shown in Fig. 2. Each neuron is capable of receiving inputs, processing the data and producing an output. The output of a neuron (O in Fig. 2) is the input (I in Fig. 2) to other neurons, to which the first neuron is connected. Each connection between two neurons is characterized by a weight (w in

5 468 L. Edenbrandt et al. Input ST-J ampl., ST slope i ^ ^» _ ^^5c T+ ampl. i T-ampl. ' Neural network g^ ^ XT. Output class 1 class 2 Let X = S Ii-Wi 1=1 ThenO = 1 +e~ Figure 2 Schematic diagram illustrating the basic structure of a neural network and a neuron. For a complete explanation, see appendix. The neural network used in this study consisted of 12 input neurons, 10 hidden neurons and eight output neurons. For simplicity, only four of the 12 input measurements are shown. I = input value; H = hidden layer; O = output value; w = interconnection weight. Fig. 2), by which the normalized input value (0 < input < 1) is multiplied. The sum of the w x I values at a neuron in the hidden or output layer is then transformed to a value between 0 and 1 which is the output from the neuron. The neurons are arranged in different layers with each neuron being connected to all neurons in the following layer but not to other neurons in the same layer. One input layer, one hidden layer (H in Fig. 2) and one output layer were used in this study. The number of neurons in the input layer depends on the number of input variables used. In this study, there were 12 input neurons, namely the eight time normalized ST-T amplitude measurements as well as the ST-J amplitude, ST slope and positive and negative amplitudes of the T wave. The number of input variables can be varied as can the number of neurons in the hidden layer. The number of neurons in the output layer depends on the number of groups used in the classification. Thus, data for example from an ST-T segment are fed to the input neurons and the network produces a value for each output neuron. The classification of the network for a given set of input data is that which corresponds to the output neuron with the greatest output value between 0 and 1. The classification depends on the values of the interconnection weights, w, of the network. These weights are initially assigned values at random and thereafter are adjusted during a training procedure using the backpropagation algorithm' 31 in order to produce the correct output for a given input. A separate learning set of data is used to train the network. After the training is completed, the weights are kept constant and the network can be used to classify new input data.