Abstract
Background
Early repolarization (ER) is defined as an elevation of the QRS‐ST junction in at least two inferior or lateral leads of the standard 12‐lead electrocardiogram (ECG). Our purpose was to create an algorithm for the automated detection and classification of ER.
Methods
A total of 6,047 electrocardiograms were manually graded for ER by two experienced readers. The automated detection of ER was based on quantification of the characteristic slurring or notching in ER‐positive leads. The ER detection algorithm was tested and its results were compared with manual grading, which served as the reference.
Results
Readers graded 183 ECGs (3.0%) as ER positive, of which the algorithm detected 176 recordings, resulting in sensitivity of 96.2%. Of the 5,864 ER‐negative recordings, the algorithm classified 5,281 as negative, resulting in 90.1% specificity. Positive and negative predictive values for the algorithm were 23.2% and 99.9%, respectively, and its accuracy was 90.2%. Inferior ER was correctly detected in 84.6% and lateral ER in 98.6% of the cases.
Conclusions
As the automatic algorithm has high sensitivity, it could be used as a prescreening tool for ER; only the electrocardiograms graded positive by the algorithm would be reviewed manually. This would reduce the need for manual labor by 90%.
Keywords: early repolarization, J point elevation, cardiac, sensitivity, specificity
Early repolarization (ER) is characterized by an elevation of the junction between the QRS complex and the ST segment (J point) in at least two inferior (II, III, aVF) or lateral (I, aVL, V4–V6) leads of the standard 12‐lead ECG. For decades, it was considered a benign finding.1 Recent studies, however, have shown that the presence of ER is associated with vulnerability to ventricular fibrillation2 as well as to cardiac and arrhythmic death, especially when present in the inferior leads3 and accompanied by a horizontal or descending ST segment.4
The detection of ER is based on visual assessment of the earliest QRS offset in any of the 12 leads. The earliest QRS offset serves as a reference line for the presence of ER in other leads, and elevation of the J point at this line or thereafter is defined as ER if its amplitude is ≥0.1 mV. In some cases, the QRS offset may vary significantly among leads, and the reference line measured from the earliest QRS offset may occur inside or near the middle of the QRS complex in other leads. Therefore, in many cases the earliest QRS offset must be visually interpreted, making the decision process more subjective and time‐consuming. In general, visual measurements have been shown to vary considerably, and average differences of 15 ms among readers have been reported for QRS offset.5, 6 Due to these challenges in defining the earliest QRS offset, we designed the algorithm to detect ER based on signal morphology rather than on the earliest QRS offset.
The goals of this study were to create an algorithm for the automated detection, localization and classification of ER, and to compare its performance with determinations by experienced readers.
MATERIALS AND METHODS
Patient Population
The patients for this study were drawn from the Health 2000 Study, which is a cross‐sectional, general population‐based epidemiological survey conducted in Finland between 2000 and 2001. The study enrolled a sample of Finnish adults aged ≥30 and <80 years and was representative of the entire Finnish adult population. Patient interviews and health examinations were performed at the survey baseline. Detailed descriptions of the survey protocol and disease definitions used in this study have been published elsewhere.7 Institutional Ethics Committees of the Helsinki and Uusimaa Hospital districts approved the study, and it was performed according to the Declaration of Helsinki. Written informed consent was obtained from all patients.
Electrocardiograms and Manual Grading of Early Repolarization
Digital 12‐lead ECGs were recorded with Marquette MAC5000 (GE Marquette Medical Systems, Milwaukee, WI, USA). Averaged ECGs from the full 10‐second recordings were automatically produced for each lead by the QT Guard software v.1.3 (GE Marquette Medical Systems). These average complexes were used in both the manual and automated analysis of ER. Manual J‐point measurements were performed on the digital ECGs with custom‐made software 8 by two experienced readers (J.T. and K.P.) who were blinded to the results of the automated method. In manual grading, an ECG was considered ER positive if J‐point elevation of ≥0.1 mV was observed in at least two leads in the inferior (II, III, aVF) or lateral (I, aVL, V4–V6) territory, which is in line with the previous studies.3, 4 Each ER‐positive lead was further classified as notched, discrete, or slurred. An example of an ER positive recording is shown in Figure 1.
Figure 1.
An example of infero‐lateral ER. Notched ER patterns are visible in the inferior leads II and aVF and slurred ER patterns in the lateral leads I, V5, and V6. Paper speed is 50 mm/s.
Automated Detection of Early Repolarization
Data Processing
ECG recordings were resampled from 256 to 500 Hz. Leads were then filtered with a 50‐Hz subtraction filter,9 after which the Savitzky–Golay smoothing filter with a 22‐ms window and a third‐order polynomial fit was applied. After R‐wave peak detection, QRS onset and offset were located for each lead with an algorithm that identifies the smallest angle between the isoelectric line (PQ‐interval/ST‐segment) and the ascending/descending part of the QRS complex.6 Median values of QRS onsets and offsets in leads other than V1–V3 were employed to determine the global QRS onset and offset (QRSon and QRSoff, respectively). The global QRS onset in each lead was used to identify the baseline as an average value of the PQ interval, from QRSon‐30 ms to QRSon−10 ms, which was then subtracted from the rest of the signal.
Six different templates (R, RS [three templates with R/S ratios of 2, 1, and 0.5], QS, RSr’), which were time‐ and amplitude‐scaled according to the QRS duration (QRSoff−QRSon), were devised and applied to determine the main QRS morphology based on the QRS complexes within each lead. This information was used to categorize ER morphology types and to minimize the number of false positives.
The steepest slope (Dfmax) after the last R or S wave exceeding 50% of the lead's maximum absolute amplitude was located. Then a line was fitted to the samples between Dfmax−2 ms and Dfmax+2 ms with the least squares method. The first point to deviate from this line was considered the yield point, after which slurring or notching of the QRS complex was determined. In order to locate the yield point, the distance to the line starting from the steepest slope was calculated with the following Equation (1):
where A is the slope, t is the timeline (in seconds), y is the signal amplitude (in millivolts), B = 1, C is the intercept, and n is the sample number. A threshold of 0.004 was used in this implementation. If the yield point was >0.1 mV, then the largest slope between the yield point and the last sample point prior to the QRS offset exceeding 0.09 mV was located. This slope was searched from the first‐order derivative of the signal filtered with a 7‐point, 14‐ms moving average. If a sample point that fulfills the above‐mentioned criteria was located, a line was fitted through the time (t) and amplitude (y) value pairs of seven consecutive samples, with the maximum slope as the central sample. The slope of the fitted line is referred to as terminal slope, and it describes the slurring of the J point (Fig. 2).
Figure 2.
Determining the yield point and terminal slope for a lead. A line is fitted through the point of the steepest slope (Dfmax), and the first sample to deviate from this line is marked as the yield point. Then, the flattest slope after the yield point is located. This is the terminal slope which is used in the detection of QRS slurring.
Detection and Classification of Early Repolarization
The algorithm used for the detection of ER is described in Figure 3. The algorithm analyzes inferior and lateral leads independently and classifies each lead as notched, discrete, slurred, or negative. In addition to the four categories, a fifth category, indeterminate, is used if the morphology of a lead cannot be definitely classified by the software.
Figure 3.
Schematics for the automated detection and categorization of early repolarization.
Notch/Discrete Detection. First, a 10th order polynomial function is fitted to the values of y from Dfmax to QRSoff +20 ms (Fig. 4). However, if the signal goes below baseline prior to the QRS offset, the polynomial function is fitted starting from the ascending part of the S wave in order to avoid jitter in the fitted signal. Subsequently, all local peaks are identified from the signal, and their amplitudes from the baseline (Apeak) and timing (TPeak) are determined. These parameters are used in the notch/discrete notch detection. A notch is detected if the yield point amplitude is >0 mV, notch amplitude is ≥0.09 mV, and the peak of the notch occurs within the timeframe spanned by the yield point (TYield) and QRS offset (TQRSoff) + 20 ms. In addition, the QRS morphology must fit either the QR or R template, and the signal cannot go below baseline prior to the examined peak. A discrete notch is detected if the notch amplitude is ≥ 0.09 mV and TYield ≤ TPeak ≤ TQRSoff + 20 ms. In addition, the signal must fall to or below baseline prior to the examined peak, and the QRS morphology cannot fit the RSr’ template.
Figure 4.
Notch detection. Signal from the steepest slope to QRS offset + 20 ms is extracted and a 10th‐order polynomial function is fitted to its values. Subsequently local peaks (▵) are located within the fit and their amplitude and timing are measured.
Slurred. If none of the local peaks fulfills the notched or discrete criteria, then the signal can be slurred, indeterminate, or negative. A lead is considered slurred if no S wave is present, yield point is ≥0.1 mV and terminal slope is >‒7.5, which corresponds to an angle of ‒71° at 25 mm/s or ‒56° at 50 mm/s paper speeds.
Indeterminate. If the lead does not meet the criteria for the notched, discrete, or slurred morphologies, it is considered either indeterminate or negative. A lead is categorized as indeterminate if it is not notched or slurred but the amplitude at the global QRS offset still exceeds the 0.1 mV threshold.
Negative. If none of the criteria are met, the signal is considered negative.
Statistics
Comparisons between patients with and without ER were made with nonparametric U‐test for unequally distributed data and with t‐test for normally distributed data. Categorical variables were compared with Fisher's exact test. Interobserver reliability (κ‐value) in the manual grading of ER was assessed using the entire patient population. All statistical analyses were performed using the SPSS version 21 (IBM SPSS Statistics, Armonk, NY, USA). All tests were two‐sided and P < 0.05 was considered statistically significant.
RESULTS
Patients with preexcitation syndrome, paced rhythm, atrial fibrillation or flutter, low‐quality ECG or missing data were excluded from this analysis. After exclusions, 6,047 eligible patients remained in the cohort. Table 1 shows the clinical characteristics of the study population and Figure 5 illustrates the flow diagram of this study.
Table 1.
Clinical Characteristics of the Study Population
| Variable | ER‐Negative | ER‐Positive | P Value |
|---|---|---|---|
| Age (years) | 51.4 ± 12.6 | 44.6 ± 11.6 | <0.001 |
| Body mass index (kg/m2) | 27.0 ± 4.7 | 25.4 ± 3.5 | <0.001 |
| Systolic blood pressure (mmHg) | 134.6 ± 21.2 | 125.2 ± 14.6 | <0.001 |
| Total cholesterol/HDL ratio | 4.8 ± 1.7 | 4.7 ± 1.6 | 0.301 |
| Heart rate (bpm) | 63.4 ± 10.7 | 58.4 ± 9.7 | <0.001 |
| QRS duration (ms) | 93.2 ± 9.3 | 86.7 ± 9.4 | <0.001 |
| Male | 2589 (44%) | 140 (77%) | <0.001 |
| Smoking | 1257 (22%) | 58 (32%) | 0.001 |
| LVH (ECG) | 854 (15%) | 72 (39%) | <0.001 |
| Hypertension | 2791 (48%) | 55 (30%) | <0.001 |
| Diabetes | 351 (6%) | 6 (3%) | 0.151 |
| CAD | 420 (7%) | 7 (4%) | 0.080 |
| Previous MI | 151 (3%) | 3 (2%) | 0.632 |
Values are shown as mean ± standard deviation (SD) and number of cases with percentages.
Abbreviations as in text. LVH (ECG) = electrocardiographic signs of left ventricular hypertrophy; CAD = coronary artery disease; MI = myocardial infarction.
Figure 5.
Flow diagram of this study. Out of the 6,305 patients 258 patients were excluded due to missing leads (N = 12), paced rhythm (N = 8), noisy ECG (N = 11), atrial fibrillation (N = 93), and flutter (N = 1), right (N = 71) and left (N = 62) bundle branch block. The automated detection of ER was tested on the remaining 6,047 patients and the results were compared with those of the manual grading, which served as the reference.
Manual Grading
Out of the 6,047 ECGs, readers graded 183 ECGs (3.0%) as ER positive. Of these, 144 (2.4%) were lateral ERs and 65 (1.1%) were inferior ERs. Both inferior and lateral ERs were exhibited by 26 patients. κ‐Value for manual grading, using the entire cohort, was 0.63 (95% CI: 0.58–0.69), indicating substantial agreement among readers.
Automated Detection of Early Repolarization
The algorithm graded a total of 759 (12.6%) cases as ER positive. It correctly detected 176 of the 183 ER‐positive cases, resulting in sensitivity of 96.2%. Of the 5,864 ER‐negative recordings, the algorithm classified 5,281 correctly, resulting in 90.1% specificity. Positive and negative predictive values for the algorithm were 23.2% and 99.9%, respectively, and its accuracy was 90.2%.
Results from the comparison between manual and automated grading are summarized in Table 2. Inferior ER was correctly detected in 84.6% and lateral ER in 98.6% of the manually graded ER cases. ER‐positive leads in the inferior territory were correctly detected in 80.8% of the cases and lateral ER‐positive leads in 93.0% of the cases. Furthermore, the morphology of the ER‐positive leads was correctly categorized in 70.9% and 73.7% of the inferior and lateral ER cases, respectively. However, when indeterminate ERs were transformed into slurred or discrete, based on QRS morphology (QRS complexes without S waves were categorized as slurred and otherwise as discrete), the percentages of correctly categorized ER‐positive leads rose to 72.4% and 82.2% for inferior and lateral leads, respectively. If only those leads that were correctly detected by the algorithm were taken into account, the categorization was accurate in 89.6% of the inferior and 88.4% of the lateral leads.
Table 2.
Accuracy of the Computer Algorithm for ECGs Manually Graded as ER‐Positive (N = 183)
| Algorithm | Inferior | Lateral |
|---|---|---|
| Detected ER‐positive ECGs | 84.6% | 98.6% |
| Detected ER‐positive leads | 80.8% | 93.0% |
| Categorized ER morphology correctly | ||
| In all cases | 70.9% | 73.7% |
| In correctly detected ER cases | 89.6% | 88.4% |
A total of 583 false‐positive cases were detected by the algorithm. Two‐thirds of these false‐positive detections were due to false slurred detections (67.0%), whereas false notched (23.1%) and discrete (9.9%) detections were responsible for the remaining one‐third of the false positive detections.
DISCUSSION
This is the first study to compare automated detection of ER against manual grading in a large‐scale general population sample. The results of this study show that the accuracy of automated detection and classification of ER based on quantification of signal morphology is high when compared against manual grading. The prevalence of ER was low in the studied population (3%). Consequently, the positive predictive value of the algorithm was hindered by the relatively high number of false positives compared to true positives. The vast majority of the false‐positive detections were due to false slurred detections.
Due to challenges in the reliable detection of the earliest QRS offset, we implemented an algorithm that relies on the analysis of signal morphology alone rather than on exact time measurements. Notched and discrete ER morphologies are rather easy to define as the characteristic peak is often easily detectable. However, conduction defects and corrupted signals, for example, due to muscle noise and power‐line interference, can induce morphologies that resemble notched or discrete ER patterns, making the identification of true ER morphologies more difficult. Slurring of the terminal part of QRS complex is more challenging to characterize than notched or discrete morphologies as slurred patterns have a wide range of variations. Especially, in inferior leads slurred morphology seems to vary more than in precordial leads, which might be related to the differences in signal amplitudes. In inferior leads, the slurring may start immediately after the R‐wave peak and continue until the QRS offset, without the typical terminal slurring of the QRS complex.
Recently, Clark et al.10 presented a method for the detection of QRS notching and slurring in the descending part of the QRS complex. Their algorithm yielded high sensitivity (92.1%) and specificity (96.6%) for the detection of notching or slurring in 300 leads drawn from a set of 50 patients (leads I and aVL were left out of the analysis). For comparison, in the 183 manually graded ER‐positive cases (183 cases × 8 leads/case = 1,464 leads), our algorithm detected ER‐positive leads with sensitivity of 88.4% and specificity of 87.2%. Both studies suggest that QRS notching and slurring can be detected accurately with automated methods. However, as Clark et al. stated, the need for internationally agreed definitions exists to ensure uniform assessment of ER. Especially in the case of slurred morphology, threshold for the degree of slurring should be defined more clearly. In this study, most of the false detections (67%) were due to false slurred detections.
CONCLUSIONS
The automated detection of ER patterns from standard 12‐lead ECGs presented in this study shows promising results. Since the algorithm has a high sensitivity, it could be used as a prescreening tool for positive ER to identify ECGs for manual review, reducing effort by 90%.
Limitations
The algorithm detected 96.2% of the ER‐positive cases. However, the detection of inferior ER was not as accurate as that of the lateral ER. This is a weakness, as inferior ER is the most significant ER pattern because it indicates the greatest risk for arrhythmic events.3, 4 Despite this shortcoming, only seven of the 183 ER‐positive cases were not detected by the algorithm. The difference in the sensitivity for the detection of ER between inferior and lateral leads (84.6% vs. 98.6%, respectively) might be related to different characteristics in the slurred morphology between the two territories.
Acknowledgments
Sandra Verrier is gratefully acknowledged for her careful revision of the language of the manuscript. Finally, the authors would like to disclose that T.K., J.T.T., and H.V.H. hold a patent for the detection of early repolarization.
This study was in part supported by Finnish Foundation for Cardiovascular Research.
REFERENCES
- 1. Klatsky AL, Oehm R, Cooper RA, et al. The early repolarization normal variant electrocardiogram: Correlates and consequences. Am J Med 2003;115(3):171–177. [DOI] [PubMed] [Google Scholar]
- 2. Haissaguerre M, Derval N, Sacher F, et al. Sudden cardiac arrest associated with early repolarization. N Engl J Med 2008;358(19):2016–2023. [DOI] [PubMed] [Google Scholar]
- 3. Tikkanen JT, Anttonen O, Junttila MJ, Aro AL, Kerola T, Rissanen HA, et al. Long‐term outcome associated with early repolarization on electrocardiography. N Engl J Med 2009;361(26):2529–2537. [DOI] [PubMed] [Google Scholar]
- 4. Tikkanen JT, Junttila MJ, Anttonen O, et al. Early repolarization: Electrocardiographic phenotypes associated with favorable long‐term outcome. Circulation 2011;123(23):2666–2673. [DOI] [PubMed] [Google Scholar]
- 5. Recommendations for measurement standards in quantitative electrocardiography . The CSE working party. Eur Heart J 1985;6(10):815–825. [PubMed] [Google Scholar]
- 6. Daskalov IK, Christov II. Electrocardiogram signal preprocessing for automatic detection of QRS boundaries. Med Eng Phys 1999;21(1):37–44. [DOI] [PubMed] [Google Scholar]
- 7. Heistaro S. Methodology Report: Health 2000 Survey. Publications of the National Health Institute, Helsinki, 2008. [Google Scholar]
- 8. Noseworthy PA, Tikkanen JT, Porthan K, et al. The early repolarization pattern in the general population: Clinical correlates and heritability. J Am Coll Cardiol 2011;57(22):2284–2289. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Levkov C, Michov G, Ivanov R, et al. Subtraction of 50 Hz interference from the electrocardiogram. Med Biol Eng Comput 1984;22(4):371–373. [DOI] [PubMed] [Google Scholar]
- 10. Clark EN, Katibi I, Macfarlane PW. Automatic detection of end QRS notching or slurring. J Electrocardiol 2014;47(2):151–154. [DOI] [PubMed] [Google Scholar]