QRS micro-fragmentation as a mortality predictor

Abstract Aims Fragmented QRS complex with visible notching on standard 12-lead electrocardiogram (ECG) is understood to represent depolarization abnormalities and to signify risk of cardiac events. Depolarization abnormalities with similar prognostic implications likely exist beyond visual recognition but no technology is presently suitable for quantification of such invisible ECG abnormalities. We present such a technology. Methods and results A signal processing method projects all ECG leads of the QRS complex into optimized three perpendicular dimensions, reconstructs the ECG back from this three-dimensional projection, and quantifies the difference (QRS ‘micro’-fragmentation, QRS-μf) between the original and reconstructed signals. QRS ‘micro’-fragmentation was assessed in three different populations: cardiac patients with automatic implantable cardioverter-defibrillators, cardiac patients with severe abnormalities, and general public. The predictive value of QRS-μf for mortality was investigated both univariably and in multivariable comparisons with other risk factors including visible QRS ‘macro’-fragmentation, QRS-Mf. The analysis was made in a total of 7779 subjects of whom 504 have not survived the first 5 years of follow-up. In all three populations, QRS-μf was strongly predictive of survival (P < 0.001 univariably, and P < 0.001 to P = 0.024 in multivariable regression analyses). A similar strong association with outcome was found when dichotomizing QRS-μf prospectively at 3.5%. When QRS-μf was used in multivariable analyses, QRS-Mf and QRS duration lost their predictive value. Conclusion In three populations with different clinical characteristics, QRS-μf was a powerful mortality risk factor independent of several previously established risk indices. Electrophysiologic abnormalities that contribute to increased QRS-μf values are likely responsible for the predictive power of visible QRS-Mf.


Recording characteristics of analysed short-term ECGs
Standard settings of the equipment were used with removal of alternating current frequencies. Where the exported sampling frequency differed from 1000 Hz, cubic spline re-sampling to this frequency was used. Although low-pass filtering with 100 Hz cut-off was applied (see subsequent Supplementary Figure  1), the 1000 Hz frequency was used for the purposes of obtaining interval measurements (in representative beats) with 1 millisecond precision. LSB -least significant bit, s -seconds.
Example of ECG pre-processing shown in a case of an atrial fibrillation patient. The left panel shows the original ECG signal in blue superimposed by filtered signals in red. The filtering was performed in two steps: (a) A low pass infinite-impulse-response Butterworth filter with 100 Hz cut-off frequency was used to eliminate high-frequency noise (it also harmonised the frequency contents of all the study ECGs). (b) Subsequently, for each detected QRS complex (combination of maximum absolute amplitudes in the native signal and its derivative) a window of preceding 100 ms was used to identify the point with minimum standard deviation across all leads. These points identified baseline wander nodes and a cubic spline interpolation across these nodes was subtracted from the filtered signal to remove baseline wander.
The right panel shows representative beatforms derived, for each ECG lead, by obtaining sample by sample medians across all superimposed QRS complexes. These representative beatforms of all 12 leads were superimposed on the same isoelectric axis and the resulting image was used to detect the global QRS onset and offset as well as the T wave offset (red vertical lines).  Table 2 of the article was computed without (Model of Score 1) and with (Model of Score 2) QRS micro-fragmentation. The table shows the resulting beta coefficients (log hazard ratios) assigned to each of the variables that were retained during the backwards stepwise elimination for Model Score 1. The Model score 2 shows the Cox regression beta coefficients after QRS microfragmentation was added to the variables of Model Score 1. The beta coefficients were used as weights of the variables to obtain weighted average risk scores. The blue lines of the table show the overall χ 2 statistics of the Cox regression models that provided the beta coefficients, areas under the receiver operator characteristic of the derived risk scores (for events across the complete follow-up), and the Harrell's C-index values of the derived risk scores. Note that in all three populations, the inclusion of QRS micro-fragmentation increased the χ 2 statistics, the area under the receiver operator characteristic, and the C-index statistics.
For each of the investigated populations, the left panels show multifactorial receiver operator characteristic (ROC) for events during the complete follow-up. The right panels show the areas under these ROC curves. Two groups of ROC curves are shown: Those labelled ART combined age, heart rate, and total cosine R to T; those labelled ARTµF included also QRS micro-fragmentation. Within each group, the ROC curves differed by the definition of true/false positive/negative: as the dichotomies of the risk factors involved were varied, for each of their combinations, positive cases were defined as those subjects for whom the values of 1, 2, 3, or 4 risk factors were above the given dichotomy. To ease the comparison of the ROC curves, their values are shown above the 50% identity line, i.e., the panels on the left show the dependency of (specificity+sensitivity-1) on sensitivity. The colours of these curves correspond to the bar graphs. Multivariable analysis used backwards stepwise elimination. In addition to hazard ratios, Wald statistics are shown. QRS micro-fragmentation was used after logarithmic transformation with base 2 -hazard ratios correspond to value increases by a factor of 2.

Association between mortality and continuous values of risk factors in aetiology sub-groups of EU-CERT-ICD
CI -confidence interval, bpm-beats per minute, deg -degrees, HR -hazard ratio, LVEF -left ventricular ejection fraction, ms -milliseconds, TCRT -total cosine R to T, µ-fragmentation -micro-fragmentation.
Characteristics of EU-CERT-ICD population per contributing centre Interestingly, when non-parametric Kruskal-Wallis one-way analysis of variance was used to test that the distribution of the risk factors shown is the same across centres, the distributions of all variables with the exception of QRS micro-fragmentation were found highly significantly different between centres (p < 0.0001 for age, heart rate, LVEF, QRS duration, and QTc interval; p = 0.0007 for TCRT). However, no differences were found between the distributions of QRS micro-fragmentation (p = 0.4173).
bpm -beats per minute, deg -degrees, ms -milliseconds, TCRT -total cosine R to T, µ-fragmentationmicro-fragmentation. Multivariable analysis used backwards stepwise elimination. In addition to hazard ratios, Wald statistics are shown. QRS micro-fragmentation was used after logarithmic transformation with base 2 -hazard ratios correspond to value increases by a factor of 2.

Association between mortality and continuous values of risk factors in contributing centres of EU-CERT-ICD
CI -confidence interval, bpm-beats per minute, deg -degrees, HR -hazard ratio, LVEF -left ventricular ejection fraction, ms -milliseconds, µ-fragmentation -micro-fragmentation.
Individual panels of the figure show survival differences stratified by QRS micro-fragmentation ≤ 3.5% (green lines) and > 3.5% (red lines) in sub-populations of the EU-CERT-ICD data defined by sex (top row); by age dichotomised at 65 years (middle row), and by heart rate dichotomised at 75 beats per minute (bottom row). In each panel, the χ 2 statistics is shown together with the corresponding p-value (log-rank test). Numbers of patients at risk in the different strata are shown below the panels in colours corresponding to the Kaplan-Meier survival curves.
Individual panels of the figure show survival differences stratified by QRS micro-fragmentation ≤ 3.5% (green lines) and > 3.5% (red lines) in sub-populations of the EU-CERT-ICD data defined by QRS duration dichotomised at 120 ms (top row); by QTc interval dichotomised at 450 ms (middle row), and by the total cosine R to T (TCRT) dichotomised at 100°(bottom row). In each panel, the χ 2 statistics is shown together with the corresponding p-value (log-rank test). Numbers of patients at risk in the different strata are shown below the panels in colours corresponding to the Kaplan-Meier survival curves (small number of cases with missing data excluded).
Individual panels of the figure show survival differences stratified by QRS micro-fragmentation ≤ 3.5% (green lines) and > 3.5% (red lines) in sub-populations of the EU-CERT-ICD data defined by creatinine plasma levels dichotomised at 1.35 mg/dL (top row); by the rhythm of the analysed electrocardiogram (middle row -see Supplementary Table 6 for further details); and by the distinction on whether the patients were, for clinical reasons, implanted with a cardiac resynchronisation defibrillator or with a device without the resynchronisation function (bottom row). In each panel, the χ 2 statistics is shown together with the corresponding p-value (log-rank test). Numbers of patients at risk in the different strata are shown below the panels in colours corresponding to the Kaplan-Meier survival curves (small number of cases with missing data excluded).
Individual panels of the figure show survival differences stratified by QRS micro-fragmentation ≤ 3.5% (green lines) and > 3.5% (red lines) in sub-populations of the EU-CERT-ICD data defined by intention to treat by beta-blockers (top row); amiodarone (middle row), and statins (bottom row). In each panel, the χ 2 statistics is shown together with the corresponding p-value (log-rank test). Numbers of patients at risk in the different strata are shown below the panels in colours corresponding to the Kaplan-Meier survival curves (small number of cases with missing data excluded). Of the 1948 patients of the EU-CERT-ICD data collection, 1558 had the ECG classified as sinus rhythm, 214 were in atrial fibrillation, 123 had the rhythm classified as "other" (trigeminy, frequent ectopic beats, atrial flutter, paced rhythm, etc.) and 53 patients had the rhythm unclassified. The analyses shown in this table show only patients with confirmed sinus rhythm and confirmed atrial fibrillation.

Association between mortality and continuous values of risk factors in EU-CERT-ICD in sinus rhythm and in atrial fibrillation
Multivariable analysis used backwards stepwise elimination. In addition to hazard ratios, Wald statistics are shown. QRS micro-fragmentation was used after logarithmic transformation with base 2 -hazard ratios correspond to value increases by a factor of 2.
For each of the investigated populations, the panels show survival differences stratified by the presence (blue lines) and absence (green lines) of visible QRS macro-fragmentations in subpopulations with QRS micro-fragmentation < 3.5% (left panels) and ≥ 3.5% (right panels). In each panel, the χ 2 statistics is shown together with the corresponding p-value (log-rank test). Numbers of patients at risk in the different strata are shown below the panels in colours corresponding to the Kaplan-Meier survival curves.
For each of the populations, the graphs show outcome probabilities in sub-groups stratified by a combination of QRS micro-fragmentation dichotomised at 3.5% and of total cosine R to T (TCRT) dichotomised at 100°. The green, blue, and red lines correspond to both factors normal, only one factor normal, and both factor abnormal, respectively. In each panel, the χ 2 statistics is shown together with the corresponding p-value (log-rank test). Numbers of patients at risk in the different strata are shown below the panels in colours corresponding to the Kaplan-Meier survival curves.