Heart Rate Variability: Measurement and Clinical Utility
`Robert E. Kleiger, M.D.,∗ Phyllis K. Stein, Ph.D.,∗ and J. Thomas Bigger, Jr., M.D.†
`From the ∗Washington University School of Medicine, St. Louis, MO and †Columbia University, New York, NY
`
`Electrocardiographic RR intervals fluctuate cyclically, modulated by ventilation, baroreflexes, and
`other genetic and environmental factors that are mediated through the autonomic nervous system.
`Short term electrocardiographic recordings (5 to 15 minutes), made under controlled conditions, e.g.,
`lying supine or standing or tilted upright can elucidate physiologic, pharmacologic, or pathologic
`changes in autonomic nervous system function. Long-term, usually 24-hour recordings, can be used to
`assess autonomic nervous responses during normal daily activities in health, disease, and in response
`to therapeutic interventions, e.g., exercise or drugs. RR interval variability is useful for assessing risk
`of cardiovascular death or arrhythmic events, especially when combined with other tests, e.g., left
`A.N.E. 2005;10(1):88–101
`ventricular ejection fraction or ventricular arrhythmias.
`
`autonomic nervous system
`
`Heart rate responds dynamically to physiologic
`perturbations mediated by the autonomic nervous
`system via efferent vagal and sympathetic nerve
`impulses.1,2 Even at rest heart rate fluctuates cycli-
`cally. High frequency (HF) cyclic fluctuations are
`modulated by ventilation, mediated entirely by
`changes in vagal outflow.3–7 Slower fluctuations
`occur due to baroreflexes or due to thermoreg-
`ulation.3–7 The greatest variation of heart rate
`occurs with circadian changes, particularly the dif-
`ference between night and day heart rate, mediated
`by complex and poorly understood neurohormonal
`rhythms.6,8 Exercise and emotion also have pro-
`found effects on heart rate. Fluctuations in heart
`rate reflect autonomic modulation and have prog-
`nostic significance in pathological states.9–45
`There are two common settings in which heart
`rate variability (HRV)
`is measured. First, HRV
`is assessed under controlled laboratory conditions
`with short-term measurements before and after
`tilt, drugs, controlled ventilation, or other maneu-
`vers selected to challenge the autonomic system.
`Secondly, HRV can be determined from 24-hour
`electrocardiographic (ECG) recordings made while
`subjects perform their usual daily activities.
`Twenty-four-hour ECG recordings are particularly
`useful for risk stratification in a variety of patholog-
`ical entities, but can also be useful for quantifying
`autonomic dysfunction.5,12,16,46–52
`
`Methods for quantifying HRV are categorized as:
`time domain, spectral or frequency domain, geo-
`metric, and nonlinear. Baroreflex sensitivity (BRS)
`and heart rate turbulence can also be considered
`measures of HRV. A short discussion of each will
`follow.
`
`TIME DOMAIN MEASURES
`OF HEART RATE VARIABILITY
`
`In time domain analysis, the intervals between
`adjacent normal R waves (NN intervals) are mea-
`sured over the period of recording.53 A variety of
`statistical variables can be calculated from the in-
`tervals directly and others can be derived from the
`differences between intervals (Table 1).53–55
`SDNN, the standard deviation of all normal RR
`(NN) intervals during a 24-hour period, is the most
`commonly used time domain measure of HRV. A
`major component of SDNN magnitude (approxi-
`mately 30–40%) is attributable to day:night differ-
`ence in NN intervals. Accurate calculation of SDNN
`requires careful editing to exclude ectopic beats,
`artifact, and missed beats. Artificially short or long
`intervals occurring as a result of these events can ar-
`tificially increase SDNN. Most laboratories require
`at least 18 hours of usable data to calculate SDNN
`in a 24-hour recording.
`
`Address for reprints: J. Thomas Bigger, Jr., M.D., Columbia University, 630 West 168th Street, PH 9-103, New York, NY 10032. Fax:
`212-305-7141; E-mail: jtb2@columbia.edu
`
`88
`
`APPLE 1033
`
`1
`
`

`

`A.N.E.
`
`r January 2005 r Vol. 10, No. 1 r Kleiger, et al.
`
`r Measurement and Clinical Utility r 89
`
`Table 1. Time Domain Measures of HRV Calculated over 24 Hours
`
`Standard deviation of all normal to normal R-R (NN) intervals
`SDNN
`Standard deviation of 5-minute average NN intervals
`SDANN
`Mean of the standard deviations of all NN intervals for all 5-minute segments in 24 hours
`ASDNN (index)
`Square root of the mean of the squares of successive NN interval differences
`rMSSD
`The number of NN intervals differing by >50 ms from the preceding interval
`NN50
`The percentage of intervals >50 ms different from preceding interval
`pNN50
`Night-day difference Mean night R-R interval minus mean day R-R interval
`
`SDANN, the standard deviation of the 5-minute
`average NN intervals, provides a “smoothed out”
`version of SDNN, i.e., measures long-term fluc-
`tuations.12 SDANN is less subject to editing er-
`ror than SDNN because averaging several hundred
`NN intervals minimizes the effects of unedited ar-
`tifacts, missed beats, and ectopic complexity. As
`such, SDANN is also much less affected by abnor-
`mal rhythms and may even permit risk stratifica-
`tion in atrial fibrillation.
`ASDNN (or SDNN index) is the average of the
`5-minute standard deviations of NN intervals.53
`It reflects the average of changes in NN inter-
`vals that occur within 5-minute periods. ASDNN is
`significantly correlated with both SDNN and
`SDANN, because low and high HRV tend to be
`global phenomena, decreasing or increasing all
`measures.
`The most common variables calculated as differ-
`ences between normal R-R intervals are rMSSD,
`NN50, and pNN50.56,57 rMSSD is the square root
`of the squares of the successive differences be-
`tween NN intervals, essentially the average change
`in interval between beats.58 NN50 is the absolute
`count of differences between successive intervals
`>50 ms,17 and pNN50 is the proportion of dif-
`ferences >50 ms.12 In the presence of normal si-
`nus rhythm and normal AV-nodal function, each of
`these measures quantifies parasympathetic modu-
`lation of normal R-R intervals driven by ventilation.
`All other time domain measures are variants of
`those discussed above and correlate highly with one
`or more of the previously discussed measures.
`
`SPECTRAL ANALYSIS OF R-R
`INTERVALS
`
`Either fast Fourier transformation or autoregres-
`sion techniques can be used to quantify cyclic fluc-
`tuations of R-R intervals.59 Traditionally, spectral
`analysis has been done in short-term laboratory
`studies; often standard 5-minute ECG segments are
`analyzed. Two peaks are seen in 5-minute R-R in-
`
`terval power spectra, a HF peak between 0.15 and
`0.40 Hz and a low frequency (LF) peak between
`0.04 and 0.15 Hz (Fig. 1, upper panel).
`High frequency power reflects ventilatory mod-
`ulation of R-R intervals (respiratory sinus arrhyth-
`mia) with the efferent impulses on the cardiac va-
`gus nerves, and is abolished by atropine. When the
`frequency of ventilation is changed, the center fre-
`quency of the HF peak moves with the ventilatory
`rate.60,61 The amplitude of the peak, reflecting the
`degree to which R-R intervals are affected by ven-
`tilation, is similar over normal ventilatory frequen-
`cies60,61
`Low frequency power is modulated by barore-
`flexes with a combination of sympathetic and
`parasympathetic efferent nerve traffic to the sinoa-
`trial node.1,3,6,37,63,64 Standing or head up tilt typ-
`ically causes a modest
`increase in LF power
`and a substantial decrease in HF power.63 At-
`ropine almost abolishes the LF peak, and beta
`blockade prevents the increase caused by stand-
`ing up. Various manipulations of high and LF
`power, e.g., normalization or
`the LF/HF ra-
`tio has been applied in an attempt
`to bet-
`ter estimate sympathetic activity. These manip-
`ulations are based on a somewhat simplistic
`“ying-yang” model of cardiac autonomic function.
`Results have been illuminating under some circum-
`stances (e.g., tilt table testing) and readily misinter-
`preted under others (numerous papers in which in-
`creases in the LF/HF ratio due to reductions in HF
`power have been interpreted as increased sympa-
`thetic activity).
`R-R interval power spectra also have been com-
`puted using data from 24-hour ECG recordings and
`categorized into total power and four mutually ex-
`clusive power bands, ultra low, very low, low,
`and HF power (Fig. 1, lower panel).9,10 Total and
`ultra-low frequency power are best calculated from
`a R-R interval periodogram of the entire 24-hour
`recording. Instead of computing the 24-hour power
`spectrum, spectral analysis often is performed on
`
`2
`
`

`

`90 r A.N.E.
`
`r January 2005 r Vol. 10, No. 1 r Kleiger, et al.
`
`r Measurement and Clinical Utility
`
`5 minutes, such as those due to day:night differ-
`ences, the 5-minute value is much smaller than to-
`tal 24-hour power. The large difference between
`5-minute and 24-hour total power can cause confu-
`sion; it is the 24-hour value that is more useful for
`prognosis (read below).
`Most of the power of HRV in a 24-hour record-
`ing resides in the frequencies below HF and LF
`power which together account for <10% of the to-
`tal power over 24 hour. About 12% of power is
`accounted for by fluctuations in R-R intervals that
`have a period between 20 seconds and 5 minutes
`(0.0033–0.04 Hz).10 This spectral band is called very
`low frequency (VLF) power. The exact physiologic
`mechanism responsible for VLF is a matter of dis-
`pute, but, like most other forms of HRV, VLF power
`is abolished by atropine, suggesting that it uses a
`parasympathetic efferent limb.64,65 Very low fre-
`quency power is also reduced by about 20% by
`ACE inhibition, suggesting that, at least in part, it
`reflects the activity of the renin-aldosterone sys-
`tem.66,67 Others have suggested that VLF power
`reflects thermoregulation or vasomotor activity.68
`Bernardi et al. showed that physical activity can
`exert a large effect on VLF power.69 In addition,
`sleep-disordered breathing can cause exaggerated
`values for VLF power, seen as clear peaks on plots
`of the HRV power spectrum during the night.70
`The lowest frequency band in the 24-hour R-R
`interval power spectrum is ultra low frequency
`(ULF) power, which quantifies fluctuations in R-R
`intervals with periods between every 5 minutes
`and once per 24 hours (ULF <0.003 Hz). Ultra
`low frequency power is strongly associated with
`SDANN.11
`Although the physiologic basis for ULF and VLF
`power are far less clear than HF and LF power,
`they have proven to be more powerful risk predic-
`tors in cardiovascular diseases.10 It is important to
`point out that accurate editing, and attention to the
`uniformity of beat onset detection, is crucial for 24-
`hour spectral analysis. Including nonNN intervals
`in the R-R interval time series will substantially
`degrade spectral analysis, even more so than for
`time domain analysis. Each of the 24-hour spectral
`measures has an equivalent time domain variable,
`which is highly correlated with it (Table 2) because
`both are influenced by the same physiologic inputs
`and because of mathematical relationships.11 For
`example, SDNN is the square root of the total vari-
`ance in normal R-R intervals, whereas total power
`is equivalent to total variance. In practice, the
`
`Figure 1. R-R interval power spectra. The upper panel
`plots log power versus frequency for a 5-minute peri-
`odogram and the lower panel plots log power versus fre-
`quency for a 24-hour periodogram. In the lower panel,
`frequency is plotted on a log scale and the Y axis is
`markedly compressed compared with the upper panel.
`Note the exponential increase in power as frequency de-
`creases below the low frequency band for both graphs.
`The two graphs resemble each other, but with much
`greater amplitude in the 24-hour plot (lower panel). The
`similarity in the graphs is consistent with fractal behavior
`for power below the low frequency band.
`
`5-minute segments from 24-hour recordings. HF
`and LF power are calculated for each suitable seg-
`ment and then averaged. Either method is suit-
`able for estimating the average 24-hour HF and LF
`power. Unfortunately, commercial Holter systems
`sometimes calculate total power in each 5-minute
`segment and report its average value over 24 hours.
`Because the 5-minute value does not measure fluc-
`tuations in R-R intervals with cycles longer than
`
`3
`
`

`

`A.N.E.
`
`r January 2005 r Vol. 10, No. 1 r Kleiger, et al.
`
`r Measurement and Clinical Utility r 91
`
`Table 2. Highly Correlated Time and Spectral
`Measures of HRV
`
`Time Domain
`
`Frequency Domain
`
`Total power
`SDNN
`ULF power
`SDANN
`VLF power
`ASDNN
`HF power
`PNN50, rMSSD
`Highly Correlated Time Domain Measures
`SDNN
`SDANN
`RMSSD
`pNN50
`
`correlations between TP and SDNN, ULF power
`and SDANN, VLF power, and SDNN index exceed
`0.85 and the correlations between ULF power (ap-
`proximately 80% of the total power) and TP, SDNN,
`SDANN also exceed 0.8. Use of time domain vari-
`ables, e.g., SDNN and SDANN rather than the spec-
`tral measures for a particular study is a matter of
`preference and capability. Because all frequency
`domain and some time domain HRV variables have
`skewed distributions, the data are usually log trans-
`formed for parametric statistical analyses.
`
`GEOMETRIC MEASURES OF R-R
`INTERVALS
`
`Heart rate variability triangular index, a geo-
`metric measure of HRV, has been used exten-
`sively by investigators at St. George’s Hospital in
`London.19,37,54 Bedeviled by difficulties in effi-
`ciently dealing with ectopic complexes, missed
`beats, and noise in analyzing recordings, they cre-
`ated histograms of the intervals by sorting them
`into 7.8 ms bins. They then fitted a triangle, using
`a least squares technique, to the height of each in-
`terval. Two measurements were made, the baseline
`width of the triangle in milliseconds and the ratio
`of the total number of beats divided by the num-
`ber of beats in the modal bin. The latter quantity
`is called HRV triangular index or just HRV index,
`and is essentially the area of the triangle divided by
`the area of the modal bin. The calculation of HRV
`index minimizes the influence of outlier R-R inter-
`vals, i.e., those much longer or shorter than the
`usual, thereby substantially reducing the influence
`of missed beats, artifact and ectopic complexes.
`With accurate editing, HRV index and SDNN are
`strongly correlated and both are powerful risk strat-
`ifiers after myocardial infarction.19,37,54
`
`NONLINEAR MEASURES OF R-R
`INTERVAL FLUCTUATIONS
`
`Although time and frequency domain measures
`of HRV quantify HRV on various time scales, non-
`linear HRV measures attempt to quantify the struc-
`ture or complexity of the R-R interval time series.
`For example, a random series of R-R intervals, a
`normal series of R-R intervals and a totally peri-
`odic series of R-R intervals might have the exact
`same SDNN, but their underlying “organization”
`would be completely different. A large number of
`nonlinear measures of HRV have been studied, but
`only a few have shown clear utility in risk stratifi-
`cation (Fig. 2). These include the power law slope,
`the short- and long-term fractal-scaling exponent,
`and SD12, a measure derived from Poincare plots.
`
`Power Law Slope
`
`In normal sinus rhythm, spectral power, mea-
`sured over 24 hours, shows a progressive, expo-
`nential increase in amplitude with decreasing fre-
`quency.71 (Fig. 1b) This relationship can also be
`plotted as the log of power (Y axis) versus the log of
`frequency (X axis), which transforms the exponen-
`tial curve to a line whose slope can be estimated
`(Fig. 2, bottom panel). In a log-log plot, the power
`−2 and 10
`−4 Hz is linear with
`law slope between 10
`a negative slope, and reflects the degree to which
`the structure of the R-R interval time series is self-
`similar over a scale of minutes to hours. Decreased
`power law slope has been shown to be a marker for
`increased risk of mortality after myocardial infarc-
`tion.72
`
`Detrended Fractal Scaling Exponent
`
`This measure, also referred to as α1, iscomputed
`from detrended fluctuation analysis (DFA) and is
`a measure of the degree to which the R-R interval
`pattern is random at one extreme, or correlated at
`the other on a scale of 3–11 beats (Fig. 2, middle
`panel).73 A totally random R-R interval pattern has
`a value for α1 of 0.5, whereas a totally correlated
`pattern of R-R intervals, i.e., one that is totally pe-
`riodic, has a value of 1.5. α1 is usually repeatedly
`measured within a period of 1000 R-R intervals and
`then averaged. Normal values are about 1.05. De-
`creased values for α1 are strong predictors of out-
`come after MI.73,74 Another measure, α2 (or DFA2)
`can be computed in a similar way on a scale of
`
`4
`
`

`

`92 r A.N.E.
`
`r January 2005 r Vol. 10, No. 1 r Kleiger, et al.
`
`r Measurement and Clinical Utility
`
`12–20 R-R intervals. α2, however, has not proved
`to be especially useful in risk stratification.
`
`The Poincar´e Plot
`
`The Poincar´e graph plots each R-R interval as
`a function of the next R-R interval (Fig. 2, top
`panel) and provides an excellent way to visualize
`patterns of R-R intervals.73 Usually, the R-R inter-
`val time series is plotted for an entire 24 hours, but
`plots of shorter periods, e.g., hourly, can reveal de-
`tails obscured in a 24-hour plot that involves about
`100,000 points. Poincar´e plots that reveal abnormal
`R-R interval patterns have been characterized as
`“complex.” In addition, Poincar´e plots that reflect
`extremely low HRV have also been classified as ab-
`normal. SD12 is determined by fitting an ellipse to
`the Poincar´e plot. SD1 is the short axis of this el-
`lipse and SD2 is the long axis. SD12 is their ratio. As
`the plot becomes more complex, the relative mag-
`nitude of SD1 compared to SD2 increases and SD12
`becomes larger (Fig. 2, top panel). In addition, if the
`plot is small and ball-shaped because of relatively
`constant R-R intervals, SD12 also will be large. This
`measure has not been used much for risk stratifi-
`cation, but has proved useful for detecting editing
`problems that significantly influence the calcula-
`tion of HRV variables.
`
`Heart Rate Turbulence
`
`Heart rate turbulence is a novel analytic method,
`which evaluates the perturbation (shortening then
`lengthening) in R-R intervals following premature
`ventricular complexes (VPC).75 Two parameters
`quantify the response to VPC: turbulence onset
`(TO) and turbulence slope (TS). Turbulence onset,
`a decrease in the first two normal R-R intervals fol-
`lowing a VPC compared with the two normal R-R
`intervals just before the VPC, presumably reflects
`baroreceptor reflex activity induced by a decreased
`stroke volume and blood pressure during the com-
`pensatory pause. Normally, the two R-R intervals
`after a VPC are shorter than the two normal R-R
`intervals immediately preceding the VPC. Turbu-
`lence slope quantifies the degree of lengthening of
`R-R intervals following the shortening of R-R in-
`tervals immediately after a VPC, again reflecting
`baroreflex activity.75 It is calculated by determin-
`ing the maximum slope of any 5-beat sequence
`of normal R-R intervals during the 15–20 R-R
`intervals after the VPC. Turbulence onset and
`
`Figure 2. Nonlinear Measures of R-R Interval fluctu-
`ations. The top panel shows a two-dimensional vector
`analysis of a Poincar ´e plot; the middle panel shows cal-
`culation of detrended fluctuation analysis (DFA); and the
`bottom panel shows calculation of the power law slope.
`The Poincar ´e plots and DFA analyses are derived from
`a 1-hour recording at night in a healthy subject. The
`power law slope is derived from a 24-hour recording. Ab-
`breviations: SD1, short-term beat-to-beat R-R variability
`from the Poincar ´e plot (width); SD2, long-term beat-to-
`beat variability from the Poincar ´e plot (length); α1, the
`short-term fractal scaling exponent for 4–11 beats; α2,
`the intermediate-term fractal scaling exponent (11–20
`beats), β, power law slope (adapted from Ref.73)
`
`5
`
`

`

`A.N.E.
`
`r January 2005 r Vol. 10, No. 1 r Kleiger, et al.
`
`r Measurement and Clinical Utility r 93
`
`turbulence slope are calculated from all single VPC
`in a 24-hour recording. Schmidt recommends that
`at least 5 VPC be present in a Holter recording, in
`order to estimate heart rate turbulence.74 Reduced
`heart rate turbulence is strongly associated with in-
`creased death rates after MI.75–77 Heart rate turbu-
`lence will be discussed in detail elsewhere in this
`journal.
`
`DIAGNOSTIC USES FOR HEART
`RATE VARIABILITY
`
`Analysis of HRV has been used to assess
`autonomic function and/or to quantify risk in
`a wide variety of both cardiac and noncardiac
`disorders. These include, among others, stroke,
`multiple sclerosis, end stage renal disease, neonatal
`distress, diabetes mellitus, ischemic heart disease,
`particularly myocardial infarction, cardiomyopa-
`thy, patients awaiting cardiac transplantation,
`valvular heart disease, and congestive heart
`failure.3,11,12,14,15,17–24,27–29,32,33,35–37,39,43,47,49,50,52
`Several authors have reported that HRV analysis
`is a more sensitive indicator of autonomic dys-
`function in alcoholics and in diabetic subjects than
`conventional autonomic tests.78–81 Heart rate vari-
`ability analysis has also been used to assess the au-
`tonomic effects of drugs, including beta-blockers,
`calcium blockers, antiarrhythmics, psychotropic
`agents, and cardiac glycosides.65–67,82–91 Drug
`effects on HRV can be established with relatively
`small numbers of study participants because HRV
`measurements are quite stable over the short- and
`long-term.92,93
`Heart rate variability analysis has had its great-
`est cardiologic use in post MI risk stratification
`and in assessing risk for arrhythmic events. Wolff
`et al. in 1978 first observed that HR variability mea-
`sured on admission to the coronary care unit was a
`predictor of mortality.94 They calculated the vari-
`ance of 30 consecutive R-R intervals taken from
`a 1minute ECG recording in 176 patients with
`acute myocardial infarction. The group of patients
`(n = 73) with R-R interval variance <32 ms had sig-
`nificantly higher hospital mortality than the group
`with preserved sinus arrhythmia (n = 103). Clin-
`ically, patients with low HRV were older, more
`likely to have an anterior infarct, and more likely to
`have heart failure. It was not clear from this study
`whether decreased HRV was an independent pre-
`dictor of adverse outcome or if it predicted long-
`term risk after myocardial infarction.
`
`The first study that clearly documented the in-
`dependent and long-term predictive value of HRV
`analysis after myocardial infarction was reported in
`1987 by the Multi-Center Post-Infarction Program
`(MPIP).28 Eight hundred and eight patients who had
`survived acute myocardial infarction had 24-hour
`ambulatory electrocardiograms prior to discharge.
`Besides Holter variables, which included mean
`heart rate, ventricular arrhythmias, and SDNN, pa-
`tients were evaluated clinically, had a radionuclide
`ejection fraction determined and were evaluated by
`a low level exercise test. During a mean follow up
`of 31 months, there were 127 deaths (Fig. 3). Forty-
`three of these deaths occurred in the group of pa-
`tients with SDNN <50 ms (125 patients), approxi-
`mately 16% of the total group. Thus, over a third
`of these patients died during follow-up and a third
`of the deaths occurred in the group with SDNN
`<50 ms, establishing a sensitivity and positive pre-
`dictive accuracy of about one third (Table 3). The
`relative risk of mortality in patients with SDNN
`<50 ms versus those with SDNN ≥50 ms was 2.8.
`Reduced SDNN was significantly associated with
`low ejection fraction, poor exercise performance,
`high New York Heart Association functional class,
`and short R-R intervals (Table 4), but the correla-
`tions were weak (0.15–0.52). Multivariate analysis
`clearly demonstrated that SDNN was an indepen-
`dent risk factor for mortality. SDNN also was the
`Holter variable with the strongest association with
`
`Figure 3. Kaplan-Meier survival curves from the Multi-
`Center Post-Infarction Study demonstrating decreased
`survival among patients with SDNN <50 ms (from
`Ref.28)
`
`6
`
`

`

`94 r A.N.E.
`
`r January 2005 r Vol. 10, No. 1 r Kleiger, et al.
`
`r Measurement and Clinical Utility
`
`Table 3. SDNN Prediction of Mortality in MPIP
`
`SDNN
`<50 ms
`
`≥50 ms
`
`683 (84.5%)
`84 (12.3%)
`
`Number of patients
`Deaths, number (%)
`Sensitivity
`Specificity
`Positive predictive
`accuracy
`False negative rate
`12.3% (84/683)
`Relative risk = 2.8 (34.4%/12.3%)
`
`125 (15.5%)
`43 (34.4%)
`33.9% (43/127)
`88.0% (599/681)
`34.4% (43/125)
`
`all-cause mortality, exceeding that of any ventric-
`ular arrhythmia measure. Using combinations of
`risk variables such as SDNN and ejection fraction
`or SDNN and repetitive VPC subgroups of MI pa-
`tients could be determined with either very high
`(50%) mortality or very low (<2%) 31 months mor-
`tality.28,29
`Multi-Center Post-Infarction Program data have
`been analyzed using other HRV measures. Bigger
`et al. evaluated the predictive value of 24-hour
`spectral measures.9–11 Because of the previously
`cited physiologic associations of various frequency
`bands, it was thought that spectral analysis might
`provide mechanistic insight into death and arrhyth-
`mias after myocardial infarction. The anticipated
`selectivity was not found. All four frequency bands
`predicted all-cause and arrhythmic mortality, but
`ultra-low frequency power had the strongest associ-
`ation with these fatal outcomes. Frequency domain
`measures of HRV had similar predictive value for
`death of all causes, cardiac death, and arrhythmic
`death. The MPIP data also have been analyzed us-
`ing heart rate turbulence, which in the MPIP data
`
`Table 4. Correlations of SDNN with Other Variables
`in MPIP
`
`Age
`Rales in the CCU
`Peak BUN
`Ejection fraction
`Duration of exercise test
`Twenty-four-hour average
`RR interval
`Ln VPC frequency
`Ln ventricular paired VPC
`Ln ventricular runs per hour
`
`r
`−0.19
`−0.25
`−0.15
`0.24
`0.15
`0.52
`−0.12
`−0.07
`−0.02
`
`P
`
`0.0001
`0.0001
`0.0007
`0.0001
`0.0007
`0.0001
`
`0.0004
`0.04
`0.57
`
`set is even a stronger risk predictor than conven-
`tional time or frequency domain variables.75
`Multi-Center Post-Infarction Program was done
`in the late 1970s, prior to the institution of much of
`what is standard therapy today. Few of the patients
`received aspirin or B-blockers and none had reper-
`fusion therapy, thrombolytics, angioplasty, or coro-
`nary artery bypass graft surgery. Thus, the ques-
`tion arose as to whether the MPIP results apply
`in the era of reperfusion. Multiple studies since
`MPIP have confirmed the power of HRV analysis
`in risk stratification post infarction. Some of these
`are summarized in Table 5.
`Some of the most important of these studies were
`performed at St. George’s hospital in London by
`Camm, Malik and co-investigators.19,33,34,37,54 Far-
`rell reported 68 patients with acute myocardial
`infarction who had both baroreceptor sensitivity
`and HRV determined before discharge from hos-
`pital.95 The latter was measured using HRV trian-
`gular index. Both BRS and HRV index were deter-
`mined to be good risk stratifiers for survival; BRS
`was superior. Subsequently, these investigators ex-
`tended their studies to over 400 survivors of my-
`ocardial infarction. In all these studies, they uti-
`lized HRV index.19,37,54 Approximately 60% of their
`patients received thrombolytic therapy or angio-
`plasty. Besides HRV index, late potentials, ejection
`fraction, clinical variables, and ventricular arrhyth-
`mias were measured. In addition, the mechanisms
`of death, arrhythmic or nonarrhythmic, and ma-
`lignant ventricular rhythms were adjudicated. De-
`creased HRV index best predicted both total cardiac
`mortality and malignant arrhythmias better than
`decreased ejection fraction, abnormal late poten-
`tials, or increased frequency of ventricular ectopy
`in 24-hour Holter ECG recordings. Furthermore,
`combining decreased HRV index with another risk
`variable, such as decreased ejection fraction or ab-
`normal late potentials, created subgroups of post
`MI patients with high risk as well as subgroups
`with very low risk of death or malignant ventric-
`ular arrhythmias.19
`The GISSI study of thrombolytic therapy in acute
`myocardial infarction evaluated HRV.52 In GISSI,
`all 12,490 patients were treated with streptokinase.
`A subset of 567 patients had a valid 24-hour ambu-
`latory ECG recording and 52 of them died during
`a 1000-day follow-up. Time domain analysis utiliz-
`ing SDNN, NN50+, and rMSSD identified high risk
`groups comprising 16–18% of the subset with mor-
`talities ranging from 20.8 to 24.2% in the high risk
`
`7
`
`

`

`A.N.E.
`
`r January 2005 r Vol. 10, No. 1 r Kleiger, et al.
`
`r Measurement and Clinical Utility r 95
`
`Continued
`
`betterspecificityforsensitivity<60%
`forall-causemortality.HRV+LVEF
`≤40%whichhadspecificityof40%
`specificity52%comparedwithLVEF
`
`mortalityforfirst6monthsonly
`independentpredictoroftotalcardiac
`mortalityforwholefollow-upbut
`
`adjustment
`Remainedsignificantafter
`1.96–5.15),PPA65%andNPA86%.
`mortality(RR3.17,95%CI
`
`α1<0.85bestunivariatepredictorof
`
`HRVindex<20univariatepredictorof
`
`151–1618daysHRVindex<39sensitivity75%,
`
`24-hour,before
`
`discharge
`
`4weeksto5
`
`years
`
`24-hour,before
`
`discharge
`
`4years
`
`24-hour,traditionaland
`
`nonlinear
`
`death
`byarrhythmicandnonarrhythmic
`predictorafteradjustment.Predicted
`causemortality,independent
`685±360daysα1<0.75RR3.0,95%CI2.5–4.2forall
`
`SDNN<70msvsSDNN≥70ms
`
`21±8months
`
`betterthanLVEF
`all-cause,cardiacandsuddendeath
`specificity,20%PPA)andpredicted
`death(45%sensitivity,85%
`interval<700mspredictedcardiac
`andHRVindexequal.MeanR-R
`sensitivity≥40%meanR-Rinterval
`RRintervalforsensitivity<40%.For
`HRVindexbetterpredictorthanmean
`
`predictor
`covariates,VLFwasthestrongest
`mortality.Afteradjustmentfor
`univariatepredictorsofall-cause
`
`ULF,VLF,LF,HFallsignificant,
`
`SDNNfor5-minuteshadlowerPPA
`
`ECGforriskstratification
`thosewhorequire24-hourHolter
`thanHRVindex,butcouldpreselect
`
`1year
`
`>2years
`
`3years
`
`HRVPredictors/Endpoints
`
`Follow-Up
`
`24-hour,<28daysafter
`
`MI
`
`24-hour,predischarge,
`
`nonlinearHRV
`traditionaland
`
`beforedischarge
`period,5–8days
`24-hour,5-minute
`
`daysafterMI)
`discharge(median7
`
`24-hour,before
`
`meds
`1weekafterstopping
`enrollinginCAPSand
`
`24-hour,1yearafter
`
`WhenObtained
`HRVMeasure
`
`N=385(44deaths,14
`
`sudden)
`
`deaths)
`totaldeaths,15sudden
`
`N=433(firstMI),(46
`
`Odemuyiwaetal.37
`
`Odemuyiwaetal.101
`
`N=159withLVEF≤35,
`
`sudden)
`deaths,5nonfatal
`N=1284(44cardiac
`
`72deaths
`
`M¨akikallioetal.(TRACE)100
`
`LaRovereetal.(ATRAMI)99
`
`nonarrhythmic
`arrhythmic,28
`0.35,114deaths,75
`
`Huikurietal.(DIAMOND-MI)74N=446withLVEF≤
`
`deaths,24sudden)
`N=700(45cardiac
`
`Feietal.98
`
`N=579,(54deaths,42
`
`cardiac,26sudden)
`
`Copieetal.97
`
`N=331(30deaths)
`Patients(Events)
`
`Numberof
`
`Biggeretal.(CAPS)96
`
`(StudyName)
`
`Source
`
`Table5.RepresentativePost-MPIPConfirmatoryStudiesofHRVasaPredictorofAll-CauseorCardiacMortalityAfterMI.NumberinParenthesis
`
`ReferstoReferenceList.OthersReferencedBelowTable
`
`8
`
`

`

`96 r A.N.E.
`
`r January 2005 r Vol. 10, No. 1 r Kleiger, et al.
`
`r Measurement and Clinical Utility
`
`relativerisk;TRACE=TRAndolaprilCardiacEvaluation;VF=ventricularfibrillation;VT=ventriculartachycardia.
`Myocarde;FU=follow-up;HRV=heartratevariability;LVEF=leftventricularejectionfraction;MI=myocardialinfarction;PPA=positivepredictiveaccuracy;RR=
`MortalityonDofetilide;GISSI=GrupoItalianoperloStudiodellaSopravvivenzanell-InfartoMiocardico;GREPI=Grouped’EtudeduPronosticdel’Infarctusdu
`ATRAMI=AutonomicToneandReflexesafterMyocardialInfarction;CAPS=CardiacArrhythmiaPilotStudy;DIAMOND=DanishInvestigationsofArrhythmiaand
`
`=3.0),rMSSD(RR=2.8)
`mortality:NN50+(RR=3.5),SDNN(RR
`VT,resuscitatedVF,ordeath)
`
`Independentpredictorsofall-cause
`
`6–17%forHRValone
`endpointswas16–29%comparedwith
`specificityat70%sensitivitywherePPAfor
`Forbestcombinationofpredictorsmaximum
`
`pNN50notdifferent
`amongnonsurvivors,butrMSSDand
`SDNN,SDANN,ASDNN,LF,HF,LF/HF
`CoxregressionnotperformedDecreased
`
`Mean32monthsSDNNsignificantlyhigherinevent-free(no
`
`daytimeSDNN<100ms(RR=2.6)
`NighttimeAVGNN<750ms(RR=3.2),
`(RR=3.90,95%CI2.03–7.49)
`independentlyassociatedwithmortality
`Afteradjustment,α(1)remained
`mortalityRR5.05,95%CI2.87–8.89).
`
`α1<0.65mostpowerfulpredictorof
`
`mortality/nonfatalinfarction(RR=2.2)
`all-causemortality(RR=1.9)or
`
`36±15monthsLnVLF<5.99independentpredictorof
`HRVPredictors/Endpoints
`
`Follow-Up
`
`1000days
`
`24-hoursatdischarge
`
`(median13days)
`
`24-hourHRV,stable,
`
`beforedischarge
`
`2years
`
`HRV
`MI,standard,nonlinear
`
`24-hour,5–8daysafter
`
`Mean8months
`
`31.4months)
`term(median
`1yearandlong
`
`24-hour,mean83hours
`
`afterMI
`
`24-hourHRV,10days
`
`afterMI
`
`months
`18.4±6.5
`
`24-hour,2–7daysafter
`
`MI
`
`24-hour,mean4days
`
`afterMI
`
`WhenObtained
`HRVMeasure
`
`Table5.Continued.
`
`deaths)
`totaldeaths,44cardiac
`withthrombolysis(52
`Zuanettietal.(GISSI)52N=567malestreated
`N=250(30endpoints)
`Zabeletal.107
`nonfatalVT/VF)
`sudden,34cardiac,13
`suddenarrhythmic,22
`N=572(43all-cause,14
`
`Vossetal.106
`
`45%hadthrombolysis
`long-termFU,9sudden)
`yearFU,39for
`
`N=226(19cardiac
`
`deaths)
`
`Viashnavetal.105
`
`Toubouletal.(GREPI)104N=471(26dea