Clin. Cardiol. 6, 447-455 (1983) © Clinical Cardiology Publishing Co., Inc. A New Microcomputer-Based ECG Analysis System M. C. KYLE, MS., J. D. KLINGEMAN, MS.*, J. D. CONRAD, B.S., E. D. FREIS, M.D., H. V. PIPBERGER, M.D.t Veterans Administration Medical Center, Washington, D.C., USA Summary: A new automated ECG system using ad- vances in microprocessor technology and computerized electrocardiography is described. This microcompu- ter-based system is self-contained and mobile. It acquires both the 12-lead and orthogonal lead (Frank) electro- cardiograms and analyzes the latter within minutes. Software includes the program developed in the Veterans Administration which uses advanced statistical classi- fication techniques and a large well-documented patient data base. Diagnostic probabilities are computed using a Bayesian approach. Diagnostic performance has been tested using independent clinical criteria and found to be quite accurate. This system enables the clinician to immediately review the computer’s identifications, measurements, and diagnostic classifications and quickly * The ECG system was conceived by Jack Klingeman, who since 1965, has specialized in diagnostic techniques for noninvasive evaluation of cardiovascular disease. In cooperation with the other authors he implemented the system and began demonstrating its usefulness to the medical community. He completed the initial manuscript describing the system shortly before his death on July 27, 1982. + The authors express their appreciation for the contribution of the ECG signal conditioning module and the software developed and supplied by Biometric Analysis Group Inc. Address for reprints: M.C. Kyle VA Medical Center (151E) 50 Irving Street, N.W. Washington, D.C. 20422, USA Received: October 19, 1982 Accepted with revision: June 23, 1983 use these results in clinical decision making. Serial comparisons are readily made since all previous re- cordings are stored on floppy diskettes. The use of mi- croprocessors in this system makes it economically fea- sible for practicing physicians. Key words: electrocardiography, microcomputers, computer-assisted diagnosis, cardiovascular diagnosis, Bayes’ theorem, probability, computers Introduction A new ECG system which combines the most ad- vanced ECG diagnostic software with hardware based on a microcomputer has now become available for clin- ical use. Software which has taken many years of re- search and much manpower to develop (Cornfield et al., 1973; Klingeman and Pipberger, 1967, 1970) has been incorporated into this system that is small, self-con- tained, and relatively inexpensive. It can be utilized as a powerful tool in a doctor’s office as well as a clinic or hospital. This system is a product of the vast amount of expe- rience accumulated in the Veterans Administration studies on computer analysis of electrocardiograms which began in the late 1950s. The software, which in- cludes the computer program developed in the Veterans Administration, uses advanced statistical techniques and the large, well-documented data base accumulated over many years. The hardware for the new system is mi- crocomputer based and represents a genuine advance in hardware design for ECG processing. 448 Clin. Cardiol. Vol. 6, September 1983 Materials and Methods Equipment and Software ECGs are acquired and analyzed in minutes using a microcomputer-based system (Fig. 1). This self-con- tained, mobile system is built around a Digital Equip- ment Corporation computer, the MINC (either the MINC-11 or MINC-23 can be used). It is on wheels and can be moved to the patient’s bedside. The MINC con- tains from 32,000 to 128,000 words of 16-bit memory, a 12-bit analog-to-digital converter, a clock, an isolation transformer, a videographing terminal, and two floppy diskette drives with a capacity of approximately one-half million bytes per diskette. One diskette contains the system software and ECG program and the other is used for storage of patient data. A graphics printer can be attached to the system to provide a hard copy of the current video display of the terminal at any time. In addition to this equipment the ECG system contains a module with patient signal conditioning amplifiers having bandwidths of 0.05-250 Hz. This module, which was designed and built especially for this system, makes it possible to record both the 12-lead and orthogonal lead ECGs and transmit them directly from the patient to the computer while providing complete electrical isolation. =e oe Fic. 1 Microcomputer-based ECG acquisition and analysis system. The ECG module plugs into a single-quad height slot of the MINC in the same manner as other MINC input/ output modules. (Another module is also available for use with this system which provides for fully automated systolic time interval analysis (Kyle et a/., 1980)). This module was designed and built to be easy to use, de- pendable, and to provide accurate results. The software is based on the computer program de- veloped in the Veterans Administration. It uses the Bayesian approach to diagnostic classification combining multivariate techniques with statistics from the large, well-documented data base accumulated over many years. The program for wave recognition and diagnostic classification has been extensively tested against inde- pendent evidence of diagnosis. Recording Procedure The patient’s diskette is placed in one of the diskette drives. The patient’s information is entered through the keyboard in response to requests for information speci- fied on the videographing terminal. The name and identifying number are entered first. Additional infor- mation such as the date, blood pressure, age, race, and sex are then entered in response to the prompts. A list of broad diagnostic categories (Table I) is then displayed for entry of prior information known about the patient. A special set of prior probabilities (priors) will be used in the diagnostic classification procedure depending on the choice. If two categories are selected a set of priors for the combination will be used. The 12-lead and Frank orthogonal lead (XYZ) ECGs are recorded with a standard Marquette Electronics, Inc. patient cable. Sets of 3 simultaneous leads are auto- matically switched under software control as follows: I M. C. Kyle et al.: Microcomputer-based ECG analysis 449 TABLE I Diagnostic categories used for selection of the patient’s tentative diagnosis for determination of prior probabilities Normal cardiovascular status Coronary artery disease Hypertensive cardiovascular disease Valvular or congenital heart disease Pulmonary disease . Heart disease of unknown etiology . Preoperative MAYER YWNO IL II], aVR aVL aVF, V; V2 V3, V4 V5 Vo, calibration signals and XYZ leads. Two seconds of each 12-lead ECG and 6 seconds of the XYZ leads are converted to digital form at a sampling rate of 500 points/s and im- mediately displayed on the graphing terminal with the appropriate labels for visual inspection. The XYZ leads are then analyzed by the microcomputer and displayed on the screen with markers at the places where the computer has identified the features for calculation of the ECG measurements (Fig. 2). They are the onset and end of the P wave, the beginning and end of the QRS, and the end of the T wave. This display allows the tech- nician or clinician to inspect the computer’s identification of these features. There is an option to proceed to the measurement and diagnostic phases or record new data. Then a report is displayed which includes the diagnostic classification, rhythm analysis, and selected measure- ments (Fig. 3). Vector loops (Fig. 4) are also available. The results and the ECG tracings are stored on the floppy diskette. Comparisons with data and results previously recorded can be obtained immediately. Wave Recognition The wave recognition program locates the beginning and end of the P wave, the onset and end of the QRS, and the end of the T wave. Spatial velocities and magnitudes are calculated from the simultaneously recorded XYZ leads to transform the three leads into a single signal accentuating the distinguishing features of the ECG. Statistics from a large data base of normal and abnormal groups of ECG, where the important features had been visually identified by trained readers, were calculated for extensive use in the wave recognition program de- velopment. Ranges of spatial velocity and magnitudes are used first for locating the ECG waveforms. After the waveforms are located statistics are used to obtain the allowable windows for their onsets and ends. The actual onsets and ends are then determined by comparing the ECG’s spatial velocity curves to templates for the be- ginning and end of the waveforms calculated from the large series of records. This program has been extensively tested and carefully checked. It has been shown that the variability of the computer wave recognition is reduced drastically as compared to human observers (Ishikawa et al., 1971). Measurements and Diagnostic Classification The measurement program uses the landmarks identified by the wave recognition program to analyze and measure the waveforms. Approximately 200 mea- surements are calculated for each cycle in 6 seconds of recording. Prior to diagnostic classification the mea- surements for all cycles are entered into a statistical al- ahah Ae lh LAL ' Fic. 2 Display of orthogonal lead electrocardiogram with markers at the places where the computer identified the features for calculation of the ECG measurements. 450 Clin. Cardiol. Vol. 6, September 1983 aay aaa LEFT VENTREL ae rane Ch vn) ae Ram ee AMPLITUDES ARE GIVEN IN HILLIVOLTS, TIME IN SECONDS Saat Tpit reer a ies naa 0.130 9.03 atl ae rz Q-DURATION R-AM bes: at CMe eC OU eee S180 at i. 4 O.3 -0.96 vt wet ees tt ae) ie aia mt ra RIL S| 5 imal mia Pei 4 XY pe Sagittal \ Horizontal KR - wo” FIG. 4 Vector loops for the P, QRS, and T waves in the frontal, sagittal, and horizontal planes. M. C. Kyle et al.:; Microcomputer-based ECG analysis 451 gorithm which tests for the reliability of the measure- ments. Measurements obtained from ectopic or aberrant cycles are eliminated and the remaining are averaged for use in classification. This procedure also eliminates measurements which are distorted by noise interference. Six-second record lengths allow efficient operation of the decision process with the number of beats depending on the heart rate. The distinctive feature of the ECG analysis program is the application of a multivariate procedure (Cornfield et al., 1973) in which large numbers of electrocardio- graphic measurements are used simultaneously for di- agnostic classification. The results of the classification are diagnostic statements with posterior probabilities computed using a Bayesian approach. These probabili- ties indicate the likelihood that an electrocardiographic abnormality is present in a given record. The sum of the probabilities for all diagnoses being considered at one time equals one. The posterior probability of a patient being in a par- ticular disease group i from the m possible groups is by Bayes’ theorem P(i|x) = le 8 gifxl pr" gif(x|i) where x is the vector of ECG measurements, g; is the prior probability of being in disease category i and f(x|/) denotes the probability density function which is calcu- lated from the patient’s ECG measurements and sta- tistics from the large well-documented patient data base. The prior probabilities are determined by asking the physician requesting the ECG to choose one or more of the broad diagnostic categories (Table I) as a tentative diagnosis. Different sets of priors are used for each choice or combinations of diagnoses. This Bayesian classification procedure is used by the microcomputer to compute the patient’s probabilities for the diagnostic categories listed in Table IT. For ECGs not exhibiting a conduction defect, 66 QRS-T mea- surements are used in combination to compute the pos- terior probabilities for the first seven categories. If the probability of either type of hypertrophy is 70 percent or more, a conditional probability of biventricular hy- pertrophy is computed. In cases with ventricular con- duction defect (QRS duration greater than 122 ms) records are classified in the categories 9-12 according to the ventricle involved and depending on whether a myocardial infarction is simultaneously present or not. The number of measurements used for this classification is 24, If a P wave is present and the measurements con- form to certain standards, a probability of normal, right (RAO) and left atrial overload (LAO) is computed. Sixteen measurements are used for P-wave analysis. A separate analysis is made of the ST segment and T wave, resulting in statements relating to left or right ventricular strain, digitalis effect and myocardial injury. Analysis of Cardiac Rhythm This section of the program, which employs a deci- sion-tree method for classification has been reported by Willems and Pipberger (1972) where the arrhythmia logic is described in detail. The most common arrhyth- mias are detected by simulating a cardiologist’s “logic tree” approach. Ectopic beats and atrial fibrillation are the most common arrhythmias. Results This self-contained, microprocessor-based ECG sys- tem makes it possible to acquire and analyze ECGs in minutes. The system is user-friendly and any technician trained to record ECGs can operate this system with minimal instructions. A key feature of the system is the display of the waveforms on the videographing terminal (Fig. 2). Vertical markers show where the computer has identified the important features. The accuracy of the identifica- tions can be immediately verified by the technician or clinician. Vector loops (Fig. 4) are also displayed. These loops are in steps of 2 ms for the P, QRS, and T waves. Plotting can also be done for each wave separately. The user has a choice of different plot rates and of having the dots connected. Time interruptions can also be made. The patient’s diskette is updated by adding the results from the current recording. The data stored include the information entered through the keyboard, the calibrated orthogonal ECGs, the wave recognition results, the re- port which was displayed on the terminal, and all other measurements which were calculated. Since the ECGs and analysis results are stored on the patient’s diskette, serial comparisons are readily available. Any results from a previous ECG recording can be retrieved for immediate visual inspection of any changes or trends. A hard copy of the waveforms or results displayed on the video screen can be obtained from a graphics printer. The most important output of the system is the ECG diagnostic report. This report contains the results of the computer analysis and is displayed for immediate viewing on the 12 inch graphing terminal (Fig. 3). The first two lines of the report contain the patient infor- mation that was entered through the keyboard. The re- mainder displays the results of the computer analysis of the ECG: heart rate and rhythm analysis, the QRS-T and P wave diagnostic statements, and selected time and amplitude measurements. The appropriate diagnostic statements (Table II) are displayed together with the posterior probability for each statement. Posterior probabilities are calculated by multivariate analysis for all diagnoses under consideration. For each diagnostic category the probability can range from 0 to 100%, and the sum of the probabilities for all categories equals 452 Clin, Cardiol. Vol. 6, September 1983 TABLE II Diagnostic categories considered in computer analysis based on morphology of ECG QRS-T wave categories if no conduction defect . Normal . Anterior myocardial infarct . Posterodiaphragmatic myocardial infarct . Lateral myocardial infarct . Chronic obstructive lung disease . Left ventricular hypertrophy . Right ventricular hypertrophy . Biventricular hypertrophy Com AM whe QRS-T wave categories if ventricular conduction defect (QRS duration > 122 ms) 9. Left ventricular conduction defect 10. Left ventricular conduction defect with myocardial infarct 11. Right ventricular conduction defect 12. Right ventricular conduction defect with myocardial infarct P wave categories 13. Normal 14. Left atrial overload 15. Right atrial overload 100%. Only the statements which have probabilities greater than or equal to 10% are printed. Probabilities greater than 70% have been found most reliable for di- agnosis. A suggested diagnosis is indicated when they are between 40 and 70%. Probabilities for abnormal diag- noses below 40% should be ruled out. For example, for the report in Figure 3 the QRS-T probability of 76% for LVH is high enough to make a diagnosis of LVH. The P-wave probabilities of 49% normal and 43% LAO are very close, indicating that LAO should be considered but the diagnosis is less certain. Special diagnostic state- ments are printed if analysis of the ST and T waves have indicated the presence of such abnormalities. The time and amplitude measurements printed in the report are only a subset of approximately 200 measurements that are calculated for each cycle and can be displayed on the terminal. It is interesting to note that one would not make a di- agnosis of LVH on the basis of the measurements in the report of Figure 3 but that is not uncommon. Although the R amplitudes do not exceed the usual limits for normal and the T-wave measurements look normal, the computer arrived at the diagnosis of LVH through the use of a large number of other measurements. The pa- tient on whom the ECG was made had a previous history of hypertension. Validation of the software for this system has been supplied by several studies where diagnostic accuracy has been tested using independent evidence of patients’ diagnoses (Brohet and Richman, 1979; Milliken et a/., 1983; Pipberger et al., 1975). In one clinical trial (Pip- berger et a/., 1975) computer analyses of more than 1192 orthogonal ECG records were compared with the “true” diagnoses of the patients, based exclusively on objective, independent information according to clinical protocols. The interpretations of 12-lead tracings performed by two well-known and experienced electrocardiographers were also compared with the “true” diagnoses. Of the total of 1192 records 86% were classified correctly by computer. In 5% the automated analysis was partially correct and 9% of the records were misclassified. Results of the 12-lead interpretations were 68%, 4%, and 28% for cor- rect, partially correct, and incorrect classifications, re- spectively. Thus, an 18% increase in correct classifica- tions was achieved by the application of the multivariate analysis. It is also significant that the number of mis- classifications was reduced from 28 to 9%. Note that it is at least as important to minimize the misclassifications as it is to maximize the correct classifications in terms of trouble for the patient. The most striking differences between automated and conventional analysis were found in patients with hypertensive cardiovascular dis- ease and chronic obstructive lung disease. In both groups computer results exceeded those obtained by conven- tional analysis by more than 50 percent. In a more recent study Milliken and co-workers (Milliken et a/., 1983) evaluated the VA computer analysis program and its impact on the cardiologist’s interpretation in a cooperative study. Again, the program was evaluated against ECG-independent evidence of 180 patients’ true diagnoses. In addition, the impact of ECG computer reports on interpretation of nine experienced electrocardiographers was evaluated by comparing their interpretations before and after addition of a computer report to the 12-lead and orthogonal 3-lead ECG. Table III shows a comparison of the accuracy of the ECG diagnoses of the readers and the computer. The computer interpretation alone showed an average increase of 22% in accuracy of diagnosis when compared to the readers’ diagnosis without benefit of the computer report. The superiority of the program was most striking for LVH where it showed a 36% improvement. Average accuracy M. C. Kyle et al.; Microcomputer-based ECG analysis TABLE II] of diagnosis? 453 Comparison of accuracy of ECG diagnosis of experienced electrocardiographers and computer against independent evidence Diagnostic No. of Accuracy of diagnosis (%) entity cases Electrocardiographers Computer Difference Normal 26 70 85 15 MI 65 65 77 12 LVH 61 41 77 36 RVH/COPD 25 36 56 20 BVH 3 74 100 26 180 54 “76 22 7 Data taken from report by Milliken et a/., 1983 of ECG diagnosis of the readers increased by 8 percent when the computer report was added to assist them in making their diagnoses. Results of evaluation of arrhythmia detection by the computer program have been reported by Willems and Pipberger (1972). Computer recognition is achieved in 90% of the cases for the most common arrhythmias. This percentage drops considerably when dealing with second or third degree heart block or other more uncommon arrhythmias. In the latter case, a statement “arrhyth- mia” is usually reported which is to alert the physician of a rhythm problem which might require therapeutic intervention. Discussion Our reason for developing this system was to combine the most advanced ECG diagnostic software with ad- vances in hardware based on state-of-the-art micropro- cessors, Until recent advances in microcomputers it was impossible to run a large, sophisticated analysis program such as the Veterans Administration’s on a portable system. It is especially remarkable that this complete, stand-alone system is more economically feasible than systems requiring telephone transmission of the ECG signals and return transmission of the report. The fidelity of the tracings which are transmitted directly from the patient to the computer by a standard ECG cable is far superior to the fidelity obtainable with telephone transmission. This self-contained, microcomputer-based system is easy to use and allows interaction between the technician and the system encouraging better quality recordings. The software used in this system has been under de- velopment since the VA cooperative study, comprising eight hospitals, was organized in 1960 for data collection (Pipberger, 1966). The orthogonal Frank-lead and 12-lead ECGs were recorded on each patient who sat- isfied strict protocol criteria. The patients were classified into diagnostic categories, based exclusively on inde- pendent, that is, non-ECG, information such as cardiac catheterization data, autopsy reports, etc. Approximately 30,000 records were collected, providing the data for development and testing. Over 300 measurements were made on each cycle of the ECGs in the Veterans Ad- ministration’s large well-documented data base. Virtu- ally every ECG variable that has been suggested in the electrocardiographic literature was included. Statistics were compiled on the measurements from these well- documented records to determine the parameters which best characterized a disease group and which best dis- criminated between disease groups (Eddleman and Pipberger, 1971; Gamboa et al., 1969; Kerr ef al., 1970). These large amounts of data were essential for devel- opment and rigorous testing of the wave recognition and diagnostic classification techniques. The importance of the wave recognition was empha- sized because any computer diagnosis is dependent on accurate wave recognition. For a computer program to be clinically useful the wave recognition must not only be consistent but correct. The wave recognition program was developed by using statistics calculated from a large number of records in which the important features had been visually identified by trained readers. Thus, the program was made to identify the features as a trained electrocardiographer would. Continued development and testing on new records insured that the algorithms used are sophisticated enough to accurately identify almost any waveform configuration. An important capability with the new system is the option to review the computer’s identifications on the video terminal (Fig. 2). When the waveform identifi- cations are made accurately the measurements calcu- lated from these identifications can be relied upon with confidence. To reduce the effects of noise and physio- logical factors the system uses records of sufficient length to calculate the measurements for many cycles. Mea- surement variation is significantly reduced by elimi- nating ectopic or aberrant beats and averaging those remaining. This technique and the use of simultaneous leads eliminate the problems which occur in systems which record only one lead at a time or lead sets that have insufficient numbers of cycles for statistical anal- ysis. 454 Clin, Cardiol. Vol. 6, September 1983 The multivariate classification program for diagnosis which applies Bayes’ theorem explains the improvement in diagnostic classification. Applications of Bayes’ the- orem for diagnostic classification are becoming in- creasingly widespread (Wagner, 1982). Use of Bayes’ theorem in the ECG program is particularly successful because of the availability of the Veterans Administra- tion’s very large data base consisting of records from patients with firmly documented diagnoses. Many measurements from each patient’s ECG are simulta- neously used with the statistics from this large data base to obtain the patient’s classification. The use of appro- priate prior probabilities in the Bayesian classification procedure proved to be a highly significant factor in di- agnostic accuracy. Ina formalized statistical sense this procedure may appear new or even strange to the phy- sician. However, it represents a procedure that is quite familiar to every clinician because he uses this approach whenever he is faced with a diagnostic problem, i.e., using prior information he has about the patient to help in making his diagnosis. Physicians’ acceptance of computer results in the form of probabilities has been gratifying. This is probably because in all problem cases, various possible diagnoses are considered with a certain likelihood. In addition, in cases with left or right ven- tricular hypertrophy, the probability levels have been found a useful indicator for the degree of hypertrophy that may increase or decrease in time as shown in fol- low-up tracings. The most frequently used measures of performance for computer analysis of ECGs are reproducibility and diagnostic accuracy. In experiments on reproducibility by Bailey and co-workers (Bailey et a/., 1976) the Vet- erans Administration program showed the least vari- ability of all programs tested. The program with the next best reproducibility had changes in diagnostic statements in 30% of the cases tested (Bailey and Horton, 1979). If a user is to have any confidence in a given program, then it must give essentially the same answer for a patient whose ECG is essentially unchanged. Evaluation of diagnostic accuracy can be performed in a variety of ways. Most designers of programs for computer diagnosis of ECGs had to make an individual and arbitrary selection of diagnostic decision rules which were considered optimal by them but not necessarily by others. Aimost without exception, performance tests were done by program designers themselves using the same decision rules. As expected, “diagnostic accuracy” of such evaluations varied between 80 and 95%. That there was any failure rate had to be attributed to tech- nical errors in wave recognition or measurement pro- grams. Since the same diagnostic rules were followed both by the computer and the human reader, it is clear that “diagnostic accuracy” is not tested at all in such studies. Comparisons between results obtained by a given computer program and results of human ECG inter- preters who may use their own diagnostic rules have been less frequently applied. As could be expected, percent- ages of disagreement rose sharply and reached almost 50% in one report (Caceres and Hochberg, 1970). Fail- ure of agreement was attributed toa variety of sources, such as technically poor recordings, physician error, imprecise measurements, differences in opinion, and imprecise criteria. Considering the accuracy rate of slightly more than 50% obtained by expert electrocar- diographers in the cooperative study by Simonson et al. (1966), one may wonder about the real meaning of such a comparison. Both the American Heart Association and the American College of Cardiology recommended, many years ago, evaluation of ECG computer programs only against independent evidence of diagnosis. The Tenth Bethesda Conference on Optimal Electrocardiography (Rautaharju ef al., 1978) calls for accuracy to be “evaluated by using a library of cases for which the di- agnosis is determined or excluded by nonelectrocardio- graphic means.” Evaluations where these recommen- dations have not been used are of limited value because of bias introduced by observers or evaluators (Pipberger and Cornfield, 1973). While this recommendation has not been followed for evaluation of other systems, the diagnostic performance of the Veterans Administration program has been extensively tested against independent evidence for the diagnoses and found to be quite satis- factory and reliable (Brohet and Richman, 1979; Mil- liken et al., 1983; Pipberger ef a/., 1975). It is fortunate that these evaluations against independent evidence do exist. The results of these studies all showed that the diagnostic accuracy rate of the computer program was significantly higher than the rate achieved by expert electrocardiographers. Thus they strongly suggest that diagnostic ECG classification can be significantly en- hanced through the use of multivariate analysis. The ECG diagnostic software used in this new system has undergone extensive development and testing and is a well-established computerized ECG program. The application of this software to a small, mobile micro- computer is new and now makes it possible for clinicians to utilize the ECG diagnostic software in a complete, stand-alone system that is ready and practical for use in a doctor’s office, clinic, or small hospital. References Bailey JJ, Horton MR: Reproducibility of version 2 of the IBM program with and without the serial comparison option. Fifth International Congress on Electrocardiography. Progress in Electrocar- M. C. Kyle et al.: Microcomputer-based ECG analysis 455 diology. (PW McFarland, Ed.) Pittmann Medical, (1979) pp. 259-264 Bailey JJ, Horton MR, Itscoitz SB: The importance of reproducibility testing of computer programs for electrocardiographic interpretation: Application to the Automatic Vectorcardiographic Analysis Program (AVA 3.4). Comput Biomed Res 9, 307 (1976) Brohet CR, Richman HG: Clinical evaluation of auto- mated processing of electrocardiograms by the Veterans Administration program (AVA 3.4). Am J Cardiol 43, 1167 (1979) Caceres CA, Hochberg HM: Performance of the com- puter and physician in the analysis of the electro- cardiogram. Am Heart J 79, 439 (1970) Cornfield J, Dunn RA, Batchlor CD, Pipberger HV: Multigroup diagnosis of electrocardiograms. Comput Biomed Res 6, 97 (1973) Eddleman EE, Jr., Pipberger HV: Computer analysis of the orthogonal electrocardiogram and vector- cardiogram in 1002 patients with myocardial in- farction. Am Heart J 81, 608 (1971) Gamboa R, Klingeman JD, Pipberger HV: Computer diagnosis of biventricular hypertrophy from the orthogonal electrocardiogram. Circulation 39, 72 (1969) Ishikawa K, Batchlor CD, Pipberger HV: Reduction of electrocardiographic beat-to-beat variation through computer wave recognition. Am Heart J 81, 236 (1971) Kerr A, Jr, Adicoff A, Klingeman JD, Pipberger HV: Computer analysis of the orthogonal electrocardi- ogram in pulmonary emphysema. Am J Cardiol 25, 34 (1970) Klingeman JD, Pipberger HV: Computer classifications of electrocardiograms. Comput Biomed Res |, 1 (1967) Klingeman JD, Pipberger HV: Time-shared computer analysis of orthogonal electrocardiograms. In Clinical Electrocardiography and Computers. (Caceres CA and Dreifus LS, Eds.). Academic Press, Inc., New York (1970) pp. 121-131 Kyle MC, Klingeman JD, Conrad JD, Freis ED: Mi- crocomputer-based systolic time interval acquisi- tion and analysis system. Proceedings of the Fourth Annual Symposium on Computer Applications in Medical Care, 1120-1124 (November, 1980). Milliken JA, Pipberger HA, Pipberger HV: The impact of an ECG computer analysis program on the car- diologist’s interpretation. A cooperative study. J Electrocardiol 16(2), 141 (1983) Pipberger HV: Vectorcardiography. (Hoffman I, Ed.). Lippincott, Philadelphia (1966) pp. 418-421 Pipberger HV, Cornfield J: What ECG computer pro- gram to choose for clinical application: The need for consumer protection. Circulation 47, 918 (1973) Pipberger HV, McCaughan D, Littmann D, Pipberger HA, Cornfield J, Dunn RA, Batchlor CD, Berson AS: Clinical application of a second generation electrocardiographic computer program. Am J Cardiol 35, 597 (1975) Rautaharju PM, Ariet M, Pryor TA, Arzbaecher RC, Bailey JJ, Bonner R, Goetowski CR, Hooper JK, Klein V, Millar CK, Milliken JA, Mortara DW, Pipberger HV, Pordy L, Sandberg RL, Simmons RL, Wolf HK: Tenth Bethesda conference on op- timal electrocardiography. Task force I]]: Com- puters in diagnostic electrocardiography. Am J Cardiol 41, 158 (1978) Simonson E, Tuna N, Okamoto N, Toshima H: Diag- nostic accuracy of the vectorcardiogram and elec- trocardiogram. A cooperative study. 4m J Cardiol 17, 829 (1966) Wagner HN, Jr: Bayes’ theorem: An idea whose time has come? Am J Cardiol 49, 875 (1982) Willems JL, Pipberger HV: Arrhythmia detection by digital computer. Comput Biomed Res 5, 263 (1972)