Acute Myocardial Infarction Detected in the 12-Lead ECG by Artificial Neural Networks
Background The 12-lead ECG, together with patient history and clinical findings, remains the most important method for early diagnosis of acute myocardial infarction. Automated interpretation of ECG is widely used as decision support for less experienced physicians. Recent reports have demonstrated that artificial neural networks can be used to improve selected aspects of conventional rule-based interpretation programs. The purpose of this study was to detect acute myocardial infarction in the 12-lead ECG with artificial neural networks.
Methods and Results A total of 1120 ECGs from patients with acute myocardial infarction and 10 452 control ECGs, recorded at an emergency department with computerized ECGs, were studied. Artificial neural networks were trained to detect acute myocardial infarction by use of measurements from the 12 ST-T segments of each ECG, together with the correct diagnosis. After this training process, the performance of the neural networks was compared with that of a widely used ECG interpretation program and the classification of an experienced cardiologist. The neural networks showed higher sensitivities and discriminant power than both the interpretation program and cardiologist. The sensitivity of the neural networks was 15.5% (95% confidence interval [CI], 12.4 to 18.6) higher than that of the interpretation program compared at a specificity of 95.4% (P<.00001) and 10.5% (95% CI, 7.2 to 13.6) higher than the cardiologist at a specificity of 86.3% (P<.00001).
Conclusions Artificial neural networks can be used to improve automated ECG interpretation for acute myocardial infarction. The networks may be useful as decision support even for the experienced ECG readers.
A significant number of patients attending the emergency department have chest pain and/or other symptoms concomitant with acute myocardial infarction. Early diagnosis of acute myocardial infarction is important because of the benefits of immediate and correct treatment. The diagnosis can be difficult, and different diagnostic methods have been studied. Elevated cardiac muscle enzymes are important for a correct diagnosis but are not useful when there is a short duration of symptoms. Together with patient history and clinical findings, the 12-lead ECG is still the most readily available and best method for the early diagnosis of acute myocardial infarction.1 2 3 4
It has been indicated that ≈25% of the patients sent home with an acute myocardial infarction had ST elevations that were misjudged or overlooked by the physician.5 Also, an estimated 80% of the patients admitted to the coronary care unit for suspected acute myocardial infarction are discharged without having this diagnosis confirmed.6 7 Computer-based ECG interpretation programs may be of help for the early diagnosis of acute myocardial infarction, but the performance of these interpretation programs could still be improved.
Artificial neural networks represent a computer-based method8 9 that has shown high performance in ECG analysis.10 11 12 13 14 Neural networks achieve their performance during training sessions in which a number of measurements for each example of a training set and the desired classification are fed to the network. The networks learn to associate the training examples with the given classification for each case. When used for diagnosing healed myocardial infarction in the 12-lead ECG, networks have demonstrated significantly higher performance than both a widely used interpretation program and an expert electrocardiographer.10 12 A recent study also showed higher accuracy in a neural network than in the physician’s diagnosis of acute myocardial infarction based on clinical data collected by the physicians, including 12-lead ECG classifications.15 The purposes of this study were to develop artificial neural networks that detect acute myocardial infarction in the 12-lead ECG and to compare the performance of the networks with those of conventional rule-based criteria, a widely used ECG interpretation program, and an experienced cardiologist.
The study was based on all ECGs recorded and stored at the emergency department of the University Hospital in Lund, Sweden, from July 1990 through June 1995. Only ECGs with severe technical deficiencies and pacemaker ECGs were excluded. A total of 1120 ECGs were recorded on patients who were admitted to the coronary care unit after the recording and were discharged with the diagnosis “acute myocardial infarction.” These ECGs were defined as the “acute infarction group.” The remaining ECGs were defined as control ECGs. Several patients contributed more than 1 ECG; eg, one patient presenting to the emergency department on three different occasions contributed 3 ECGs. If acute myocardial infarction was diagnosed on one of these three occasions, that ECG was part of the acute infarction group, and the other two were control ECGs.
During the study period, the policy of the diagnosis of acute myocardial infarction was as follows. At least two of the following three criteria must have been fulfilled: characteristic chest pain lasting >20 minutes, elevated creatine kinase levels, and characteristic serial ECG changes. Creatine kinase–B values >0.23 μkat/L with a typical rise and fall were used as diagnostics for acute myocardial infarction. ECG evidence of acute myocardial infarction included new Q waves in at least two adjacent leads and/or persistent T inversions in more than two adjacent leads after a newly developed ST elevation in those leads. Each discharge diagnosis was confirmed by a senior cardiologist at the department of cardiology.
The acute infarction group consisted of 699 ECGs recorded on men and 421 ECGs recorded on women. The mean (±SD) age was 70.9±12.4 years. The control ECGs recorded during 1992 and 1993 constituted the control group (no acute myocardial infarction), which consisted of 10 452 ECGs. There were 5275 ECGs recorded on men and 5177 ECGs recorded on women in the control group. The mean age was 64.0±18.5 years.
The 12-lead ECGs were recorded by use of computerized electrocardiographs (Siemens-Elema AB). The ST-T measurements used as input to the artificial neural networks were obtained from the measurement program of the computerized ECG recorders. The following six ST-T measurements from each of the 12 leads were considered: ST-J amplitude, ST slope, ST amplitude 2/8, ST amplitude 3/8, positive T amplitude, and negative T amplitude. The ST amplitude 2/8 and ST amplitude 3/8 were calculated as follows. The interval between the ST-J point and the end of the T wave was divided into eight parts of equal duration. The amplitudes at the end of the second and the third intervals were denoted ST amplitude 2/8 and ST amplitude 3/8. No QRS measurements were used as input to the neural networks.
Artificial Neural Networks
Artificial neural networks with a multilayer perceptron architecture were used.16 A more general description of neural networks can be found elsewhere.8 The neural networks consisted of one input layer, one hidden layer, and one output layer. There were 72 neurons in the input layer, one for each of the input variables, ie, six ST-T measurements from each of the 12 leads. The hidden layer contained 15 neurons. The output unit encoded whether the ECG was classified as acute myocardial infarction or not.
During a training process, the connection weights between the neurons were adjusted by use of the Langevin extension of the back propagation updating algorithm.17 The learning rate (η) had a start value of 0.5. During the training, η was decreased geometrically every epoch using the following equation: η=kη, with k=0.998. The momentum (α) was set to 0.7. Updating occurred after every 10th pattern. The Langevin noise was chosen to decrease geometrically from 0.005, with k=0.993 during the training process. The network weights were initiated with random numbers between –0.025 and 0.025.
To decide when to terminate the training process to achieve optimum performance and to avoid overtraining, a stopping criterion was established. This criterion was calculated by use of a threefold cross-validation procedure. The data set was randomly divided into three equal parts. One part was used as a test set, and training was performed on the remaining two parts. This procedure was repeated three times so that each part was used once in a test set. Each time all ECGs in the training set were presented to the network, the performance was evaluated with respect to the error obtained in the training and test sets. This evaluation did not alter the connection weights. The error in the training set decreased with an increased number of training cycles, whereas the error in the test set reached a minimum, after which it increased despite the further decrease in training error. Network training beyond the minimum error in the test set is called overtraining. The error in the training set, which corresponds to the minimum error in the test set, was assessed. The mean of the three training errors calculated in the threefold cross validation procedure was defined as the stopping criterion in the training procedure.
In the final training procedure, an eightfold cross-validation procedure was used to obtain as reliable a performance as possible. Each of the eight different networks was trained until the error in the training set reached the stopping criterion. The test results of the eight different networks were combined in the calculations of neural network performance. All calculations were done with the JETNET 3.0 package.18
The output values for test ECGs were in the range of 0 to 1. A threshold in this interval was used above which all values were regarded as consistent with acute myocardial infarction. By varying this threshold, a receiver-operating characteristic (ROC) curve was obtained.
The performance of the artificial neural networks was compared with that of an experienced cardiologist who was the head of the coronary care unit. All the ECGs of the acute infarction group and the same number of ECGs from the control group were presented to the cardiologist in random order. The 1120 control ECGs were selected at random from the control group. The cardiologist classified each of the ECGs into one of the following four classes: definite acute myocardial infarction, probable acute myocardial infarction, probable non–acute myocardial infarction, and definite non–acute myocardial infarction. Personal data, clinical findings, and the results from the neural networks were not available at the classification procedure.
The performance of the neural networks was also compared with those of two sets of conventional rule-based criteria. Criteria A were as follows: ST-segment elevation >1 mm in two or more adjacent extremity leads or >2 mm in two or more adjacent precordial leads. Criteria B represented a more complex set of criteria. The ECG interpretations presented by the computerized electrocardiographs at the emergency department were used as criteria B.19 The criteria were considered positive if a statement indicating the possibility of acute myocardial infarction was present in the computer-based interpretation, eg, “acute infarction,” “recent infarction,” or “repeat if myocardial infarction is suspected.”
Sensitivities, specificities, and differences in sensitivities are presented with 95% confidence intervals (CIs). Because of the reciprocal relation between sensitivity and specificity, the comparisons of networks versus criteria A, networks versus criteria B, and networks versus cardiologist were performed as follows. The threshold applied to the network outputs was chosen so that the specificity of the neural networks was the same as that of the criteria or cardiologist. Thereafter, the corresponding sensitivity of the networks was compared with the sensitivity of the criteria or cardiologist, and the significance of the difference in sensitivity was tested by paying attention to the fact that the same ECGs were used; ie, a McNemar type of statistic was used.
The discriminant power of a test was also calculated to facilitate a comparison between criteria and cardiologist. This measure was calculated by use of the log of the likelihood ratios of each approach, with values between 2 and 3 indicating a high discriminant power and values of ≈1 consistent with low performance.20
Fig 1⇓ shows the ROC curve of the neural networks and the results of criteria A, criteria B, and the cardiologist. The Table⇓ gives the sensitivities and specificities with 95% CIs and the discriminant power of the criteria and the cardiologist. The sensitivity and discriminant power of the neural networks compared at the same levels of specificity as the criteria and cardiologist are also presented in the Table⇓.
The neural networks showed higher sensitivities and discriminant power than both the criteria and cardiologist. The sensitivity of the neural networks was 18.3% (95% CI, 15.1 to 21.5) higher than that of criteria A compared at a specificity of 95.2% (P<.00001) and 15.5% (95% CI, 12.4 to 18.6) higher than that of criteria B compared at a specificity of 95.4% (P<.00001). The difference in sensitivity between neural networks and cardiologist was 10.5% (95% CI, 7.2 to 13.6) at a specificity of 86.3% (P<.00001).
The cardiologist classified 22.3% and criteria B classified 3.8% of the ECGs in the acute infarction group as definite acute myocardial infarction. A false classification as definite acute myocardial infarction was made in only 0.2% and 0.4% of the ECGs in the control group by the cardiologist and criteria B, respectively. At these high levels of specificity, the neural networks had 13.0% (95% CI, 10.7 to 15.3; P<.00001) higher sensitivity than criteria B and 8.9% (95% CI, 6.7 to 11.1) lower sensitivity than the cardiologist (P<.00001).
Fig 2⇓ shows one of the ECGs from the acute infarction group that was classified as definite acute myocardial infarction by the cardiologist. This ECG also had a very high output from the neural networks, indicating a high probability for acute infarction, but it was missed by both criteria. The appearance of the ST segment strongly suggests acute inferior myocardial infarction, but an ST elevation >1 mm was found only in lead III of the inferior leads.
The results show that artificial neural networks can be trained to detect acute myocardial infarction in the 12-lead ECG at a sensitivity much higher than two sets of conventional rule-based criteria and even better than an experienced cardiologist. This is in accordance with earlier ECG studies in which neural networks have shown high performance in the diagnosis of healed myocardial infarction and detection of ECG lead reversals.10 11 12 The accuracy of the networks in these studies was also better than that of rule-based criteria and an experienced ECG reader. These networks have one important characteristic in common: They are trained by use of large and well-validated databases. The performance of a neural network depends largely on the size and composition of the training database. ECG databases consisting of hundreds or thousands of recordings have made it possible to successfully apply neural networks in the automated ECG analysis.
The neural networks showed a higher sensitivity than the cardiologist, but the cardiologist was better at finding ECGs with clear-cut changes of acute infarction, ie, definite acute myocardial infarction. A possible explanation of these results could be the fact that a cardiologist is used to focusing his attention on symptoms and clear-cut ECG changes that qualify the patient for thrombolytic treatment. In contrast, the networks were trained to separate the ECGs of the acute infarction group from those of the control group, not to find candidates for thrombolysis.
The sensitivity and specificity of a diagnostic method depend on the composition of the population studied. To facilitate a comparison with other studies, a simple set of criteria was also included in this study (criteria A). The sensitivity for these criteria was 28.8% in the present study but was much higher (68%) in a study by Lee and coworkers.3 This difference in sensitivity, using the same criteria, demonstrates that the materials studied are different. For example, in the Lee et al study, only patients with chest pain were included, whereas all ECGs recorded in the emergency department were included in this study. Furthermore, only ECGs with severe technical deficiencies and pacemaker ECGs were excluded from this study. Pathological QRS complexes, eg, conduction defects, which often affect the ST-T morphology, were not excluded. This approach was used because the aim was to develop a method that can be applied to all types of ECGs in an emergency department.
The results show that artificial neural networks can be used to improve automated ECG interpretation for acute myocardial infarction and that even an experienced cardiologist could use these networks as decision support. This improvement could lead to a more accurate early diagnosis of acute myocardial infarction in patients attending an emergency department.
The neural networks of the present study could be incorporated in computer-based ECG interpretation programs and could detect acute myocardial infarction in the 12-lead ECGs by use of input variables from a measurement program; ie, no data would need to be fed manually to the network. The advantage with this type of input is that the same performance could be expected when the network is used in other emergency departments. The disadvantage is that this type of decision support helps the physician with only a limited part of the diagnostic decision, the ECG interpretation. In a recent study by Baxt and Skora,15 a network identified acute myocardial infarction using a set of clinical variables entered into a computer by the physician. The accuracy was higher for the networks than for the physicians. A problem with the use of inputs that are preclassifications by different physicians is the intraobserver and interobserver variabilities of the preclassification. The neural network could give different answers in the same case if the users make different preclassifications.
Artificial neural networks were trained to detect acute myocardial infarction in the 12-lead ECG at a sensitivity much higher than that of conventional rule-based criteria. These results show that the networks can be used to improve automated ECG interpretation for acute myocardial infarction. The neural networks also performed higher than an experienced cardiologist, indicating that they may be useful as decision support even for the experienced ECG readers. The potential for neural networks as decision aid is probably high.
This study was supported by grants from the Swedish Medical Research Council (B95-14X-09893-04B), Swedish National Board for Industrial and Technical Development, and the Faculty of Medicine, Lund University, Sweden.
- Received December 22, 1996.
- Revision received April 23, 1997.
- Accepted April 28, 1997.
- Copyright © 1997 by American Heart Association
Timmis AD. Early diagnosis of acute myocardial infarction. BMJ. 1990;301:941-942.
Timmis AD. Will serum enzymes and other proteins find a clinical application in the early diagnosis of myocardial infarction? Br Heart J. 1994;71:309-310.
Lee HS, Cross SL, Garthwaite P, Dickie A, Ross I, Walton S, Jennings K. Comparison of the value of novel rapid measurement of myoglobin, creatine kinase, and creatine kinase-MB with the electrocardiogram for the diagnosis of acute myocardial infarction. Br Heart J. 1994;71:311-315.
Roberts R, Kleiman NS. Earlier diagnosis and treatment of acute myocardial infarction necessitates the need for a ‘new diagnostic mind set.’ Circulation. 1994;89:872-881.
Bortolan G, Willems JL. Diagnostic ECG classification based on neural networks. J Electrocardiol. 1993;26(suppl):75-79.
Rumelhart DE, McClelland JL, eds. Parallel Distributed Processing, Volumes 1 and 2. Cambridge, Mass: MIT Press; 1986.
Rögnvaldsson T. On Langevin updating in multilayer perceptrons. Neural Comput. 1994;6:916-926.
Macfarlane PW, Lawrie TDV. Comprehensive Electrocardiology. Oxford, UK: Pergamon Press Inc; 1989;3:1554-1561.