OUP user menu

Different spectral components of 24 h heart rate variability are related to different modes of death in chronic heart failure

Stefano Guzzetti, Maria Teresa La Rovere, Gian Domenico Pinna, Roberto Maestri, Ester Borroni, Alberto Porta, Andrea Mortara, Alberto Malliani
DOI: http://dx.doi.org/10.1093/eurheartj/ehi067 357-362 First published online: 9 December 2004


Aims To assess whether analysis of heart rate variability (HRV) from 24 h Holter recordings provides information about the mode of death (pump failure vs. sudden death) in chronic heart failure (CHF).

Methods and results We analysed 24 h HRV in 330 consecutive CHF patients in sinus rhythm. Indices derived from time domain, spectral domain, and fractal analyses of 24 h automatic HRV were evaluated. Data from clinical assessment, echocardiography, right heart catheterization, exercise test, blood biochemical examination, and arrhythmia pattern were analysed. Patients were followed up for 3 years. Two simple multivariable models, both including 24 h spectral indices, were able to identify patients at higher risk of progressive pump failure and sudden death, respectively. Depressed power of night-time HRV (≤509  ms2) below 0.04 Hz [very low frequency (VLF)], high pulmonary wedge pressure (PWP≥18 mm Hg) and low left ventricular ejection fraction (LVEF≤24%) were independently related to death for progressive pump failure, while the reduction of power between 0.04 and 0.15 Hz at night (LF≤20 ms2) and increased left ventricular end-systolic diameter (LVESD≥61 mm) were linked to sudden mortality.

Conclusion Automatic spectral analysis of 24 h HRV provides independent risk indices related to mode of death in sinus rhythm CHF patients.

  • Heart failure
  • Sudden death
  • Progressive pump failure
  • Heart rate variability
  • Autonomic nervous system


Progressive pump failure is the leading cause of mortality in patients with chronic heart failure (CHF), while sudden death accounts for one-third to one-half of all deaths.1 While a large number of invasive and non-invasive parameters have been shown to identify patients at increased risk for worsening heart failure, arrhythmic and clinical patterns have a low predictive power for sudden events.2 Some measures of heart rate variability (HRV) have been shown to provide independent prognostic information in CHF patients.35 The plausibility of these results is related to the link between these indices and autonomic modulation of the heart.6 However, it has not yet been established whether long-term power spectral analysis might differentiate heart failure from sudden death risk. Moreover, the predictive value of slower oscillations [very low frequency (VLF) band, ≤0.04 Hz] has not been analysed in patients with CHF.

The aim of the present prospective study was to assess whether spectral analysis of HRV, from 24 h Holter recordings, provides information about the mode of death (pump failure vs. sudden death) in CHF.


Study population

The patient population consisted of 352 consecutive CHF patients in sinus rhythm, with stable optimized therapy during the prior 2 weeks, and Holter recordings available for off-line analysis, who were admitted to the Heart Failure Unit of the Scientific Institute of Montescano between July 1991 and January 2001. These patients were part of an already published population,5 of which 25% were in atrial fibrillation or had a pacemaker implanted. Of the 352 patients enrolled, 22 were excluded for poor quality recordings. All the remaining 330 patients had a 24 h recording analysable for at least half of the night-time (00:00–05:00 h) and half of the daytime (09:00–19:00 h). The relevant measures from clinical and blood biochemical examination, 2D echocardiography, right heart catheterization, exercise testing, and the arrhythmia pattern of the enrolled patients are summarized in Table 1. At the time of enrolment, patients were treated with diuretics (91%), angiotenain converting enzyme-(ACE-)inhibitors (94%), digitalis (68%), nitrates (50%), amiodarone (23%), and beta-blockers (15%). According to the changes in guidelines, beta-blocker treatment has been progressively increased during follow-up. At the end of follow-up 42% of patients were on beta-blockers. Seventeen patients had implantable cardioverter defibrillator (ICD), either already present at baseline or implanted during the follow-up.

View this table:
Table 1

Baseline clinical and functional characteristics of study population (330 patients)

 Age, years54 (47–59)
 Male, %87
 BMI, kg/m225.2 (23.1–27.9)
 SBP, mmHg110 (100–120)
 Diabetes, %16
 Smoking history
  Current smoker, %25
  Previous smoker, %49
  Never smoker, %26
 NYHA Class
  I, %11
  II, %53
  III, %36
  Ischaemic, %49
  Idiopathic, %45
  Valvular, %4
  Other, %2
 LVEF, %24 (19–28)
 LVESD, mm61 (54–68)
 LVEDD, mm71 (66–78)
 DT, ms125 (95–160)
 MR grade 3–4, %41
 PWP, mmHg18 (10–27)
 RAP, mmHg4 (2–7)
 CI, L/min/m22.2 (1.9–2.6)
 VPCs/h, N18 (4–59)
 RR 24-h, ms813 (730–902)
 RR night, ms907 (786–1017)
 SDNN 24-h, ms89 (68–121)
 SDNN night, ms71 (54–94)
 VLF 24-h, ms2894 (422–1734)
 VLF night, ms21147 (509–2530)
 LF 24-h, ms263 (25–189)
 LF night, ms270 (20–235)
 HF 24-h, ms2127 (58–269)
 HF night, ms2140 (60–360)
 LF 24-h, nu21 (12–32) %
 LF night, nu23 (12–33) %
 HF 24-h, nu58 (48–66) %
 HF night, nu65 (53–74) %
 LF/HF 24-h0.82 (0.37–1.57)
 1/f slope−1.18 (−1.33 to −1.04)
Cardiopulmonary exercise
 VO2 peak, mL/kg/min14.3 (11.9–17.4)
Blood chemistry
 Cholesterol total, mg/dL211 (176–239)
 Triglycerides, mg/dL137 (99–197)
 BUN, mg/dL47 (39–59)
 Creatinine, mg/dL1.19 (1.03–1.34)
 Na, mEq/L139 (137–141)
 K, mEq/L4.36 (4.1–4.6)
 Bilirubin (mEq/L)0.99 (0.68–1.32)

Continuous variables are reported as median (lower quartile – upper quartile).

LVEDD, left ventricular end diastolic diameter; DT, deceleration time; RAP, right atrial pressure; CI, cardiac index; VPCs, ventricular premature contractions; nu, normalized units; LF/HF ratio.

The local ethics committee approved the study and all patients gave their informed consent.

Holter data analysis

All Holter recordings were performed using a two-channel bipolar recorder. Analogue-to-digital conversion of ECG recordings was performed at 250 Hz. The RR interval series (tachograms) were automatically obtained; when artefacts or arrhythmia were present, corrections were made using appropriate software (ELATEC software version 3.0; Ela Medical, SNIA Milano, Italy). A consecutive series of 300 beats, overlapped with 150 beats, was considered for the analysis (∼600 series/day). A series was automatically discharged if it did not contain at least 95% of sinus beats correctly recognized. In agreement with the published recommendations,6 mean RR (ms) and the standard deviation of normal-to-normal RR intervals (SDNN) (ms) were calculated. The power spectral density was assessed for each RR series using a recursive autoregressive algorithm, as previously described.7 Total power (ms2) and the power in very low frequency (VLF, below 0.04 Hz), low frequency (LF, 0.04–0.15 Hz) and the high frequency (HF, 0.15–0.40 Hz) bands were evaluated. Power of LF and HF components was also considered in normalized units.6 The power in each band was calculated by summing all components with central frequency inside the band. The 24 h and the night-time values of HRV parameters were calculated as the mean of HRV indices across all accepted RR series. The night-time results were separately reported because the night represents a more homogeneous period among patients during which they stayed in bed, whilst during the daytime patients accomplished different unrestricted activities, thus affecting HRV parameters in a different way.8

The fractal characteristics of 24 h HRV were evaluated using a non-linear approach.9 The power spectrum was computed on the 24 h RR series not divided into short series of 300 beats, over the range 0.0001–0.04 Hz, using Fast Fourier Transformation, and plotted on a log-log graph. The slope (referred to as the 1/f slope in the following) of the regression line of the log(power) vs. the log(frequency) was determined. It represents the exponent of the power log relationship P=Kfα where P is the power spectral density, f is the frequency, and K is a constant. The 1/f slope was considered a measure of fractal complexity.10


During follow-up, patients were periodically re-evaluated and re-admitted to hospital if they became clinically unstable. In case of death, the date and modality (progressive heart failure or sudden) were accurately investigated. Sudden (presumably arrhythmic) death was defined as death occurring within 1 h of onset of symptoms in a previously medically stable patient, death during sleep, unwitnessed death occurring within 1 h of the patient last being seen alive or appropriate and documented ICD discharge for fast ventricular tachycardia or ventricular fibrillation. Time–event information for each subject was recorded in a dedicated database together with the clinical, functional, and HRV parameters recorded at baseline.

Statistical analysis

Endpoints of survival analysis were progressive pump failure death + urgent transplantation (defined as ‘patients requiring ventricular support or in-hospital intensive care at the time of transplantation’), and sudden death. When we analysed the endpoint of pump failure, patients who died of sudden death or of non-cardiac causes (cancer or accidental death, n=13), and those who underwent elective heart transplantation (i.e. patients who were not under intensive support at the time of transplantation, n=37), were considered censored observations. The same procedure was carried out when we analysed the endpoint of sudden death. Continuous variables characterized by higher risk at lower values (e.g. peak VO2) were categorized as follows: below the first quartile (high risk), between the first quartile and the median (intermediate risk) and above the median (low risk). Similarly, continuous variables characterized by higher risk at higher values (e.g. pulmonary wedge pressure, PWP) were categorized according to the median and the upper quartile.

Survival functions were estimated using the Kaplan–Meier method. The univariable association between each variable and the event was assessed by the Cox proportional hazards regression model. The risks associated with adjacent risk classes were compared statistically (high vs. intermediate risk and intermediate vs. low risk), and the two classes merged together in case of a non-significant difference, thus leading to a two-level categorization (dichotomous variable). All predictors of the same category of variables (clinical, echocardiographic, etc.) having a P<0.1 were considered as possible candidates for a multivariable Cox model, in order to identify those containing independent prognostic information. In this way we obtained different multivariable models, one for each category of variables. Then, all variables selected in these models were taken together to build the final prediction model of the study. Results are presented as relative risk (RR) and corresponding 95% confidence intervals (CIs). Multivariable modelling was carried out using a stepwise procedure and verified using the best subset selection approach. We used P<0.25 for entering a variable into the model and P<0.15 to remain in the model. At the end of the stepwise process, only variables with P<0.05 were retained in the model. The assumption of proportional hazards was assessed by examination of plotted Schoenfeld residuals against time.

Because of the skewness in the distribution of many variables, descriptive statistics are given as median and interquartile range (IQR). All statistical tests were two-sided. Statistical analyses were performed with the SAS/STAT statistical package, release 8.02 (SAS Institute Inc., Cary, NC, USA).


Follow-up data

During a 3 year follow-up (median 34; range 1–36 months), 79 patients (24%) died because of pump failure (62 patients) or underwent urgent transplantation (17 patients), while sudden death occurred in 29 patients (9%), including one patient with appropriate ICD discharge.

Progressive pump failure death

A large number of clinical and test variables, including the use of beta-blockers, low cholesterol, and systolic blood pressure values and many HRV parameters (excluded LF and HF expressed in normalized units) were significantly related to non-sudden death in univariable models (Table 2). For all continuous variables, dichotomous categorization described the prognostic information better than three-level categorization, as Figure 1 shows for patients with the VLF power at night below and above the cut-off value of 509 ms2. In multivariable analysis, three factors were found to be independently associated with pump failure death: a depressed power of night-time HRV below 0.04 Hz (VLF≤509 ms2), an increased pulmonary wedge pressure (PWP≥18 mm Hg), and a reduced LVEF (≤24%) (Table 3). The cumulative 3 year mortality rates were: 7% for patients without risk factors, 20% for patients who had only one risk factor, 32% for patients with two, and 44% for patients with three risk factors.

Figure 1 Kaplan–Meier survival curves in relation to the endpoint of death from progressive heart failure and urgent transplantation. Patients with reduced power at night (n) of heart rate variability in the very low frequency band (VLF, <0.04 Hz) were at increased risk of mortality (dash lines). Survival curves related to the LF-night power below or above the cut-off value are also shown (solid lines).

View this table:
Table 2

Significant univariable predictors of progressive pump failure death and urgent transplantation

Variable (cut-off value)χ2PRR (95% CI)
 NYHA class (=3)15.8<0.00012.5 (1.6–3.8)
 SBP (≤ 110 mmHg)9.10.00262.2 (1.3–3.6)
 LVEF (≤ 24%)15.3<0.00012.7 (1.6–4.4)
 LVESD (≥68 mm)9.40.00222.1 (1.3–3.2)
 LVEDD (≥78 mm)6.80.0091.8 (1.2–2.9)
 MR (≥3) (1.1–2.7)
 PWP (≥18 mmHg)13.90.00022.8 (1.6–4.7)
 RAP (≥7 mmHg)11.60.00072.4 (1.5–4.1)
 CI (≤1.9 L/min/m2)10.60.00122.3 (1.4–3.8)
 RR 24-h (≤730 ms)7.30.00681.9 (1.2–3.0)
 RR night (≤786 ms)12.70.00042.3 (1.5–3.7)
 SDNN 24-h (≤68 ms)13.60.00022.4 (1.5–3.7)
 SDNN night (≤71 ms)11.80.00062.3 (1.4–3.6)
 VLF 24-h (≤894 ms2)17.0<0.00012.7 (1.7–4.4)
 VLF night (≤509 ms2)22.6<0.00013.0 (1.9–4.7)
 LF 24-h (≤63 ms2)13.80.00022.4 (1.5–3.9)
 LF night (≤70 ms2)11.50.00072.3 (1.4–3.7)
 HF 24-h (≤58 ms2)6.00.01461.8 (1.1–2.8)
 HF night (≤60 ms2)8.00.00471.9 (1.2–3.1)
 LF/HF 24-h (≤0.37)6.50.0111.8 (1.1–2.9)
 1/f slope (≤−1.33)15.2<0.00012.4 (1.6–3.8)
Cardiopulmonary exercise
 VO2 (≤11.9 mL/kg/min)4.90.0261.7 (1.1–2.8)
Blood chemistry
 Cholesterol total (≤176 mg/dL)10.20.00142.1 (1.3–3.4)
 Sodium (≤137 mEq/L)5.30.0211.7 (1.1–2.7)
 Bilirubin (≥0.99 mEq/L)7.80.00531.9 (1.2–3.1)
Baseline therapy
 Beta-blockers4.60.030.11 (0.02–0.82)
View this table:
Table 3

Significant multivariable predictors of progressive pump failure death and urgent transplantation

Variable (cut-off value)χ2PRR (95% CI)
VLF night (≤509 ms2)9.70.00182.3 (1.4–3.8)
PWP (≥18 mmHg)6.00.01452.0 (1.1–3.5)
LVEF (≤24%)5.00.02521.9 (1.1–3.3)

Sudden death

Two 24 h spectral HRV indices (LF-night, HF-night) and one echocardiographic parameter (left ventricular end-systolic dimension LVESD) were the only variables significantly related to sudden death in univariable models (Table 4). In multivariable analysis, only LF-night showed a highly significant relationship with sudden death (RR 2.6, 95% CI 1.2–5.5, P=0.012), while an increased left ventricular dimension (LVESD≥61 mm) had borderline significance (RR 2.3, 95% CI 1.0–5.2, P=0.052). No relationship was found between aetiology (ischaemic vs. non-ischaemic cardiomyopathy) and sudden death, both in univariable (P=0.25) and multivariable analysis (P=0.33 testing for interaction). The Kaplan–Meier survival curves for sudden death according to night-time LF power below or above the cut-off value (≤20 ms2) are reported in Figure 2. The 3 year mortality increased from 8% to 21% in patients below the cut-off value.

Figure 2 Kaplan–Meier survival curves in relation to the endpoint of sudden (presumably arrhythmic) death. Patients with reduced power of heart rate variability in the low frequency (LF, 0.04–0.15 Hz) band at night (n) showed an increased risk of event (solid line). The VLF-night curves drawn according to the cut-off value (509 ms2) are shown (dashed line).

View this table:
Table 4

Significant univariable predictors of sudden (presumably arrhythmic) death

Variable (cut-off value)χ2PRR (95% CI)
LF night (≤20 ms2)7.00.00792.7 (1.3–5.6)
HF night (≤60 ms2)4.30.03852.2 (1.0–4.6)
LVESD (≥61 mm)5.10.02392.6 (1.1–5.9)

In Figure 1, the contribution of LF power≤20 ms2 to pump failure death is also shown; Figure 2 shows the contribution of VLF≤509 ms2 to sudden death.


This study indicates that in CHF patients, 24 h HRV spectral parameters remained independently related to non-sudden and sudden mortality in two multivariable models taking into account clinical assessment, blood biochemical examination, echocardiography, right heart catheterization, cardiopulmonary exercise test, and arrhythmia pattern. Although in previous publications, HRV spectral parameters appeared to describe the autonomic balance better while expressed in normalized units,7,11,12 in this larger study absolute units (ms2) provided the highest significant association with death. We suggest that in CHF the stronger predictive value of the absolute power may derive from its link to the condition of the markedly reduced total variability characteristic of these patients. Indeed, the independent prognostic value of SDNN (the simplest measure of total variability) in CHF was demonstrated by Nolan et al.,3 while Galinier et al.,4 and La Rovere et al.5 showed that a reduction of the LF spectral component of HRV was strongly related to sudden death, performing 24 h and short-term HRV analysis, respectively. In the automatic analysis from routine Holter recordings, a great number of variables were considered and the evaluation of the data according to the mode of death represents the leading novelties of the present study.

Progressive pump failure

Patients who died from progressive pump failure had a reduced variability below 0.04 Hz (VLF), which was the strongest predictor of non-sudden death. Although all other HRV parameters were significant predictors in univariable analysis, they did not enter the multivariable model when VLF power was added. Different physiological mechanisms for VLF oscillations have been proposed: physical activity,8 thermoregulation,13 renin-angiotensin-aldosterone system,14 slow respiratory patterns,15 and parasympathetic mechanisms.16 The present HRV results, obtained from 24 h Holter recordings, made during unrestricted daily activity, could have been influenced by a reduced physical activity in the patients who were more ill.8 Nevertheless, if the relationship between VLF and the risk of progressive pump failure were just expression of a reduced physical activity, the VLF power during the day period, and not the night, would be the best predictor of death.

Our data show that the evolution toward non-sudden death is linked to a severely impaired capacity of autonomic control mechanisms to influence the heart rate at all frequencies, although VLF oscillations were those better related to a worse prognosis.

The 1/f slope, a measure of fractal complexity,10 was related to pump failure in the univariable model (Table 2), yet linear spectral indices provided more powerful prognostic information than fractal long-term properties of HRV.

While several studies have addressed the value of reduced left ventricular ejection fraction (LVEF) in predicting cardiac death,17 it has recently been regarded as a risk stratifier for patients who may benefit from ICD implantation.18 The present study suggests that LVEF is related only to the risk of progressive pump failure death. It is worth saying that in our population, however, the cut-off value for LVEF (<24%) was well below that used in primary prevention trials. Finally, high PWP is a reliable index of pulmonary congestion; thus the strong relationship between PWP and risk of non-sudden death was not unexpected.

Sudden death

The relationship between decreased LF power and sudden death was previously described.4,5 The interpretation of a reduced LF in CHF patients is still an open question including a depressed sinus node responsiveness,19 central abnormality in autonomic modulation,20 limitation in responsiveness to high levels of cardiac sympathetic activation,21 depressed baroreflex, and increased chemoreceptor sensitivity.22 These altered receptor sensitivities would be partly mediated by reflex sympathetic hyperactivity emerging from the activation of cardiovascular sympathetic sensory endings.23 The relationship between reduced LF power and increased sympathetic activity in CHF20 could explain why reduced LF power, as shown in Figure 1, is also a significant univariable predictor of pump failure. However, multivariable modelling showed that VLF ‘contains’ the prognostic information of LF for pump failure death.

Clinical implications

The results of the present study integrate well into the current debate around the turmoil created by the practical implications of the results of the MADIT-II trial.24

Although a reduced LVEF has emerged as the most important predictor of arrhythmia risk, even in patients with idiopathic dilated cardiomyopathy,25 basing the decision of implanting ICD only on LVEF has been shown to be rather unattractive in terms of cost-effectiveness. A ‘more detailed exploration of risk stratification’ has been claimed to improve the decision process.24 Our data show that among a large population of patients at increased arrhythmic risk (all our patients had an LVEF <30%), subjects ‘less likely to benefit from ICD implantation’24 could be identified by the use of the information provided by spectral analysis of the 24 h HRV. Indeed, the 3 year sudden death was 3% in the 37% of patients exhibiting a preserved LF power and a less dilated left ventricle. These data are well in line with a recent analysis on MADIT-II patients including microvolt T-wave alternans.26 Thus, the current criteria for risk stratification and treatment decisions in patients with CHF should be reappraised and the design of future cost-effectiveness studies should take the emerging variables into account.


We express our gratitude to Anna Testori and Elisabetta Achilli for their skilful secretarial support in the management of the study database.


View Abstract