Prevalent Low-Frequency Oscillation of Heart Rate
Novel Predictor of Mortality After Myocardial Infarction
Background— This study evaluates a novel method for postinfarction risk stratification based on frequency-domain characteristics of heart rate variability (HRV) in 24-hour Holter recordings.
Methods and Results— A new risk predictor, prevalent low-frequency oscillation (PLF), was determined in the placebo population of the European Myocardial Infarction Amiodarone Trial (EMIAT). Frequencies of peaks detected in 5-minute low-frequency HRV spectra were averaged to obtain the PLF index. PLF ≥0.1 Hz was the strongest univariate predictor of all-cause mortality associated with relative risk of 6.4 (95% CI, 3.9 to 10.6; P<10−12). In a multivariate Cox’s regression model including clinical risk factors, mean RR interval, HRV index, low- and high-frequency HRV spectral power, and heart rate turbulence, PLF was the most powerful mortality predictor, with a relative risk of 4.6 (95% CI, 2.2 to 9.3; P=0.00003). Predictive power of PLF was blindly validated in the population of the Autonomic Tone and Reflexes After Myocardial Infarction (ATRAMI) trial. PLF ≥0.1 Hz was associated with univariate relative risk of 6.1 (95% CI, 2.9 to 12.9; P<10−5) for cardiac mortality or resuscitated cardiac arrest. In multivariate Cox’s regression model including age, left ventricular ejection fraction, baroreflex sensitivity, mean RR interval, standard deviation of normal RR intervals, low- and high-frequency HRV spectral power, and heart rate turbulence, only left ventricular ejection fraction and PLF were significant predictors, with relative risks of 4.2 (95% CI, 1.5 to 11.7; P=0.007) and 3.6 (95% CI, 1.3 to 10.5; P=0.02), respectively.
Conclusions— An innovative analysis of frequency-domain HRV, which characterizes the distribution of spectral power within the low-frequency band, is a potent and independent risk stratifier in postinfarction patients.
Received January 16, 2004; de novo received April 1, 2004; revision received May 12, 2004; accepted May 18, 2004.
The mortality of high-risk patients surviving myocardial infarction (MI) can be effectively reduced by prophylactic implantation of an automatic cardioverter-defibrillator. Results of the Multicenter Automatic Defibrillator Implantation Trial II1 have a major impact on health economics. Therefore, more specific selection of patients at risk of death, based on more than simply ejection fraction, is crucial for the future development of cost-effective prophylactic treatment.
Numerous risk factors have been proposed in postinfarction patients. In addition to clinical predictors, a number of noninvasive electrophysiology risk factors have been investigated. Whereas some have not fulfilled initial expectations (late potentials,2 QT dispersion3), the predictive value of others [mean heart rate, heart rate variability (HRV)4–6] is relatively modest. Recently, heart rate turbulence (HRT) was shown to be a stronger risk predictor compared with a number of other stratifiers, including age, left ventricular ejection fraction (LVEF), mean heart rate, and HRV.7
This study explored a novel method for postinfarction risk stratification based on modified processing of the frequency-domain HRV characteristics. Whereas the conventional approach uses integrated power spectrum density in predefined frequency bands, this method was developed to characterize the distribution of spectral power within the low-frequency (LF) band. A single descriptor of heart rate oscillatory pattern in the LF band was called prevalent LF oscillation (PLF). Its predictive power was first investigated in the placebo arm of the European Myocardial Infarction Amiodarone Trial (EMIAT)8 and subsequently validated in the population of the Autonomic Tone and Reflexes After Myocardial Infarction (ATRAMI)9 trial.
A total of 743 patients (112 women) 60.7±9.2 years of age at 5 to 21 days after index MI with LVEF ≤40% (30.2±7.5%) were randomized into the placebo group; 438 (60.0%) had thrombolysis for the treatment of the index MI, and 333 (44.8%) were receiving β-blockers at hospital discharge. During the follow-up period of 665±104 days, 102 patients in the placebo arm died.
A total of 1284 patients (165 women) 57.3±10.0 years of age at <28 days after MI with LVEF of 49.0±11.8% were enrolled; 784 (61.1%) had thrombolysis for the treatment of the index MI, and 254 (19.8%) were treated with β-blockers at baseline. During the follow-up period of 664±236 days, 49 patients reached an end point defined as cardiac mortality or resuscitated cardiac arrest with documented ventricular fibrillation.
Baseline 24-hour Holter recordings were centrally analyzed with the Laser Holter 8000 System (Marquette Medical Systems). The original RR interval series were filtered by use of a multipass approach. At each step, the RR interval was identified that differed most from the interpolated mean of immediately preceding and following RR intervals. This interval was eliminated if the difference from the interpolated mean exceeded 20% for time-domain HRV and HRT analyses and 10% for the frequency-domain HRV analysis. This filtering was iterated until no RR interval required elimination.
Time-Domain Analysis of HRV
A Holter recording was considered suitable for time-domain analysis of HRV if lasting ≥18 hours and containing ≥80% of analyzable data (the sum of all RR intervals after filtering). Mean RR and HRV index10 were assessed for each recording that fulfilled these criteria.
Heart Rate Turbulence
Indexes of HRT [turbulence onset (TO) and turbulence slope (TS)] were calculated as previously described.7 Ventricular premature complexes (VPCs) were included in the analysis if preceded and followed by ≥15 sinus RR intervals, if their prematurity was ≤80%, and if the compensatory pause was ≥120%. At least 5 VPCs (EMIAT) and 1 VPC (ATRAMI) were required for this analysis.
Frequency-Domain Analysis of HRV
Each Holter recording was divided into 5-minute nonoverlapping segments. In each segment containing ≥95% analyzable data, the discrete series of sinus RR intervals was linearly interpolated at a 2-Hz frequency, any linear trend was subtracted, and power spectral analysis was performed with the method of averaging periodograms.11 Specifically, discrete Fourier transform was applied to Hanning-filtered and zero-padded 60-sample segments with 50% overlap. This approach provided a frequency resolution of 1/60 Hz, and power spectra were obtained over the frequency range of 0.017 to 0.5 Hz.
Spectra were integrated in LF (0.04 to 0.15 Hz) and high-frequency (HF; 0.167 to 0.400 Hz) bands. Median LF oscillation (MLF) was calculated as the frequency that divided the LF band into 2 regions of equal power. These standard frequency-domain indexes were averaged for all (≥10) available 5-minute segments over 24 hours.
PLF index was calculated from individual power spectra of all 5-minute segments containing ≥95% analyzable data. With the 1/60-Hz frequency resolution, the LF band contained 7 power spectrum values at frequencies of 0.033, 0.050, 0.067, 0.083, 0.100, 0.117, and 0.133 Hz. In each spectrum, all local peaks defined as spectral position with power spectrum density more than both adjacent power spectrum densities were detected within the LF band, and the maximum peak (the most powered) was included in the PLF computation. Frequencies of all maximum peaks (≤1 per each 5-minute segment) were averaged over the whole recording to obtain the single value of PLF. Detectable peaks in ≥10 segments per Holter recording were required for a valid PLF calculation.
Risk factors were compared between survivors and nonsurvivors through the use of a 2-tailed t test for independent samples. LF and HF indexes with clearly nonnormal distributions were log-transformed before the analysis. Pearson’s correlation analysis was performed to characterize the interrelationship between individual variables.
In the EMIAT population, conventional predictors were dichotomized at previously established cutoffs (age ≥65 years, NYHA class II or higher, LVEF ≤30%, mean RR ≤800 ms, HRV index ≤20 U, TO ≥0%, and TS ≤2.5 ms/RR). For other continuous variables (QRS duration, LF, HF, and PLF), the dichotomies were set at 40% sensitivity for all-cause mortality. The analysis also included other available nominal clinical stratifiers: gender, presence of previous MI, presence of diabetes mellitus, thrombolysis, and treatment with β-blockers at hospital discharge. The association of all dichotomized predictors with all-cause mortality was examined with univariate and multivariate Cox’s regression analyses.
The validation of the predictive power of the PLF index in the ATRAMI population was performed by use of the dichotomies established in the EMIAT population, except the dichotomy for LVEF (<35%), which was chosen to comply with previous studies. For the same reason, standard deviation of normal RR intervals <70 ms was used instead of HRV index. The association of all available risk predictors, including baroreflex sensitivity (BRS) <3 ms/mm Hg, with the combined end point was investigated through the use of univariate and multivariate Cox’s regression analyses.
Holter recordings and RR interval data files were available in 633 cases. In this subgroup, 87 patients died during the follow-up period of 662±105 days. The median Holter duration was 24.0 hours [interquartile range (IQR), 23.5 to 24.3 hours]. Proportions of sinus RR intervals were 99.2% (IQR, 97.2% to 99.7%), 99.1% (IQR, 96.9% to 99.6%), and 98.7% (IQR, 95.8% to 99.5%) after exclusion of ectopic complexes and after filtering for time- and frequency-domain analysis, respectively. There were 265 (IQR, 214 to 284) analyzable 5-minute segments per Holter recording, and 20 (IQR, 12 to 34) had a detectable LF peak. The number of analyzable Holter recordings (percent of all available Holter recordings) was 592 (93.5%), 607 (95.9%), 431 (68.1%), and 520 (82.1%) for time-domain and frequency-domain HRV, HRT, and PLF index, respectively.
Although MLF did not significantly differ between survivors and nonsurvivors, PLF was significantly shifted to higher frequencies in nonsurvivors and was the statistically strongest predictor among all other risk factors (Table 1). All previously described risk predictors (HRV index, LF, HF, and TS) were moderately interrelated (r=0.57 to 0.84, P<0.0001), and all correlated with mean RR (r=0.33 to 0.58, P<0.0001). On the contrary, PLF correlated with neither mean RR (r=−0.02) nor other indexes (Figure 1), apart from a weak negative correlation with LF (r=−0.33, P<0.0001).
Selected dichotomies for individual predictors and corresponding relative risks are shown in Table 2. The PLF index had the strongest association with all-cause mortality with a relative risk of 6.4 (P<10−12).
When all significant univariate predictors (except gender) were entered into the Cox multivariate regression model, PLF remained the most powerful predictor (relative risk, 4.6; P=0.00003), followed by HF (relative risk, 2.6; P=0.03) and previous MI (relative risk, 2.1; P=0.01). None of the other risk predictors was an independent predictor of all-cause mortality (Table 2). Kaplan-Meier survival curves for mean RR, HF, TS, and PLF and their composites are shown in Figures 2 and 3⇓. The combination of abnormal PLF and abnormal HF was the most powerful mortality predictor, which defined a small population of 20 patients (3.8% of the population with analyzable Holter data) with a 70% mortality during the follow-up and relative risk of 11.6 (P<10−15). In EMIAT, indeterminate PLF alone was associated with an increased risk (mortality of 19.5% compared with overall mortality of 13.7%). However, this risk is considerably lower than that associated with abnormal PLF; consequently, it is not useful for risk stratification.
Holter recordings and RR interval data files were available in 1139 cases. Of these, 45 patients reached the combined end point during a follow-up period of 674±234 days. A comparison of individual risk predictors in both groups is shown in Table 3.
Table 4 shows selected dichotomies for individual predictors and corresponding relative risks. LF had the strongest (relative risk, 6.5; P<10−8) and PLF had the second-strongest (relative risk, 6.1; P<10−5) association with the combined end point. When all significant univariate predictors (except gender and TO) were entered into the Cox’s multivariate regression model, only LVEF (relative risk, 4.2; P=0.007) and PLF (relative risk, 3.6; P=0.02) remained independent predictors of the combined end point. Figure 4 shows Kaplan-Meier survival curves for LVEF, BRS, TS, and PLF.
PLF was established and validated to be a very potent risk predictor in postinfarction patients. Being fully independent of mean RR, PLF does not correlate with most previously established risk stratifiers. Rather than overall HRV, PLF appears to quantify a dynamic and transitory pattern of RR interval modulation, which cannot be assessed with conventional indexes.
Several studies reported that the specific distribution of spectral power of RR interval and/or blood pressure fluctuations within the LF band might provide additional information compared with the total spectral power in this band. Different descriptive methods have been used, including spectral power in the midfrequency band,12 central frequency derived from autoregressive models,13,14 median frequency from Fourier analysis in the LF band,15 and maximum coherence between the RR interval and blood pressure oscillations detecting the dominant oscillatory component in the LF band.16 Some of these indexes have been shown to be superior to the traditional descriptors of HRV in the discrimination of healthy control subjects and patients, eg, diabetics,17 elderly,12,14 borderline hypertensives,18 and patients with coronary artery disease16. All these studies supported the fact that RR interval or blood pressure oscillations in the LF band are shifted toward lower frequencies in patients compared with healthy control subjects and that this shift progressively increases with severity of organic heart impairment.
Because of the previous observations, we expected MLF and PLF to be shifted toward lower frequencies in high-risk postinfarction patients. Surprisingly, MLF failed to distinguish high-risk patients, probably because low- and high-risk postinfarction patients differ less than healthy subjects and patients investigated in previous studies.17,18 More surprisingly, a higher PLF was strongly and independently associated with clinical end points, indicating that PLF differs qualitatively from indexes describing central or dominant frequency within the LF band.
Although the exact physiological background of PLF currently is not known, several suggestions can be made. Spectra obtained from high-risk subjects are flat and frequently with less apparent 0.1-Hz periodicity (Figure 5). A reduced power in the LF band is an independent risk predictor for sudden death in heart failure patients.19 Likewise, in the present study in post-MI patients, a reduced power in the LF band was a significant mortality predictor. Respiratory arrhythmia is also generally less prominent in high-risk patients. All these factors provide optimal conditions for the spurious and low-powered peaks of unknown origin to emerge in the high-frequency area of the LF band.
The extensive filtering was needed because, although the original files were edited to a “high clinical standard,” the possibility still existed that some ectopic complexes or artifacts were wrongly annotated. The multipass filtering approach that, compared with single-pass, more effectively manages misrecognized arrhythmias and/or artifacts was defined prospectively. Spectral HRV analysis is particularly sensitive to QRS recognition errors; thus, more stringent filtering (<10%) for spectral analysis was set prospectively on the basis of experience in independent Holter databases. Although we tried to ensure that no erroneous sinus RR intervals were included, even at the cost of omitting some genuine sinus RR intervals, the filter eliminated only <1% of the data. When the data were reanalyzed without filtering, the relative risks of all HRV/HRT indexes were lower, but the dominance of PLF in the multifactorial analysis was fully preserved (data not presented).
PLF index is calculated from a relatively low number of highly selected 5-minute segments in which ≥1 LF peak was found. Spectral smoothing resulting from the Welch method (and consequently, low prevalence of LF peaks) appeared to be crucial for risk stratification power of PLF. When we tried to reduce the number of spectra entering the averaging process, the prevalence of LF peaks increased, but at the same time, the predictive power of PLF was reduced. Nevertheless, LF spectral power alone, which modestly correlated with PLF, had considerable predictive power. However, LF predictive power was lost in multivariate analysis even if cutoff values of both PLF and LF were set a priori at the same level of sensitivity. Therefore, PLF must convey some additional information relevant to risk stratification.
It has not been investigated whether different approaches to spectral analysis (parametric autoregressive models) might provide an even more powerful PLF index. The 10% threshold of the initial filtering procedure, 5-minute duration of RR interval segments processed by spectral analysis, and the analysis threshold of 95% of noise- and/or ectopy-free RR interval series were all selected prospectively but empirically and may not be optimal.
In multivariate analysis, some risk factors were obviously handicapped by their previously established cutoff values with sensitivities largely >40%. However, risk stratification with PLF at a relatively low level of sensitivity provided the highest positive predictive accuracy of all risk predictors in both the study populations; this may be a valuable approach if costly treatment (such as implantable cardioverter-defibrillators) results from the risk stratification.
In both studies, PLF was analyzable only in ≈80% to 85% Holter recordings because of an inadequate number of 5-minute spectra with LF peak. However, TS (the strongest electrophysiological competitor to PLF) encounters a similar problem because the presence of VPCs is necessary for its assessment. Consequently, PLF and TS may serve as complementary stratifiers.
In conclusion, a novel, qualitatively unique approach based on frequency-domain analysis of HRV provides a very potent and independent postinfarction risk stratifier. However, before direct clinical applicability is proposed, the pathophysiological background to this observation should be elucidated and the predictive power of PLF should be confirmed prospectively in a contemporary population of postinfarction patients.
This work was supported in part by the Wellcome Trust (grant 060683) and the Ministry of Health of the Czech Republic (research grant 6685-3).
Zabel M, Klingenheben T, Franz MR, et al. Assessment of QT dispersion for prediction of mortality or arrhythmic events after myocardial infarction: results of a prospective, long-term follow up study. Circulation. 1998; 97: 2543–2550.
Malik M, Farrell T, Cripps T, et al. Heart rate variability in relation to prognosis after myocardial infarction: selection of optimal processing techniques. Eur Heart J. 1989; 10: 1060–1074.
Bigger JT, Fleiss JL, Steinman RC, et al. Frequency domain measures of heart period variability and mortality after myocardial infarction. Circulation. 1992; 85: 164–171.
Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology. Heart rate variability: standards of measurement, physiological interpretation and clinical use. Circulation. 1996; 93: 1043–1065.
Veerman DP, Imholz BPM, Wieling W, et al. Effects of ageing on blood pressure variability in resting conditions. Hypertension. 1994; 24: 120–130.
Molgaard H, Hermansen K, Bjerregaard P. Spectral components of short-term RR interval variability in healthy subjects and effects of risk factors. Eur Heart J. 1994; 15: 1174–1183.
La Rovere MT, Pinna GD, Maestri R, et al. Short-term heart rate variability strongly predicts sudden cardiac death in chronic heart failure patients. Circulation. 2003; 107: 565–570.