Recovery from exercise can be divided into an early, rapid period and a late, slower period. Although early heart rate (HR) recovery 1 minute after treadmill exercise independently predicts survival, the prognostic value of late HR recovery has not been well studied. The aim of this study was to evaluate the independent prognostic value of late HR recovery for all-cause mortality. A total of 2,082 patients referred to the nuclear cardiology laboratory of an urban academic medical center for treadmill exercise with imaging from August 1998 to December 2003 were followed for all-cause mortality. During 9.9 ± 1.5 years of follow-up, 196 deaths (9%) occurred. To avoid overlap with early HR recovery or the baseline HR, late HR recovery was defined as the percentage of the cycle length change between rest and peak exercise that had been recovered after 5 minutes. Lower values represent impaired recovery, by analogy with 1-minute HR recovery. Impaired late HR recovery was a significant univariate predictor of all-cause mortality (hazard ratio 0.28 per percentage, 95% confidence interval 0.17 to 0.46, p <0.001). It significantly improved a nested, multivariate model (change in chi-square 8.66, p = 0.003), including 1-minute HR recovery, with independent prognostic value (adjusted hazard ratio 0.58, 95% confidence interval 0.41 to 0.84, p = 0.004). In conclusion, late HR recovery after treadmill exercise stress adds prognostic value for all-cause mortality to a multivariate model including early, 1-minute HR recovery.
Heart rate (HR) response at the start of exercise, and its recovery in the early post-exercise recovery period, have been shown to be independent prognostic factors for survival. However, complete HR recovery after exercise extends well beyond the 1-minute mark and is a complex interplay among intrinsic, sympathetic, and parasympathetic components. The prognostic impact of the later component of HR recovery after exercise has received limited study. Therefore, we hypothesized that incorporating late HR response would improve prediction of all-cause survival when added to early (1-minute) recovery. We specifically focused on overall mortality as an unbiased, objective end point.
Methods
The study was approved by the institutional review board of Northwestern University. Our cohort consisted of patients referred to our nuclear cardiology laboratory for a treadmill exercise stress study with single-photon emission computed tomographic (SPECT) perfusion imaging. Inclusion criteria were age ≥18 years, a study performed from August 1998 to December 2003 (to ensure the availability of archived and decipherable digitized electrocardiographic (ECG) waveforms), and available ECG tracings at peak exercise and at 1 and 5 minutes into recovery. Exclusion criteria were missing Social Security number, a previous included study (only the first of multiple studies was included), or an increase in heart rate of >10% between serial tracings during recovery. All-cause mortality was determined from the Social Security Administration Death Master File. Length of survival was computed as the number of days between the treadmill exercise stress test and the date of death or the end of December 2010, whichever came first. All patients had ≥7 years of follow-up.
Available demographic and clinical information included age, sex, previous mechanical revascularization, classic risk factors, body mass index, and current cardiac medications. Bruce protocol treadmill exercise duration, treadmill-induced angina, and peak blood pressure were recorded. In contrast to previous work, our exercise laboratory does not use a formal 2-minute cool-down period, although patients continue on the treadmill for approximately 30 seconds after peak exercise before the belt stops. Rest, stress, and recovery 12-lead ECG tracings were interpreted using standard criteria.
Most SPECT studies used a single-day dual-isotope protocol. Rest images were acquired using intravenous thallium-201 and stress images with technetium-99m sestamibi. A minority of studies used a 2-day technetium-99m sestamibi protocol in patients weighing >350 lb to improve image quality. Images were interpreted using a 20-segment scoring system in which 0 = normal tracer activity and 4 = no tracer activity. Total scores for the stress and rest images determined the summed stress score and summed rest score, respectively, the difference of which represents the combined size and severity of stress-induced perfusion defects. Most stress SPECT images were electrocardiographically gated to yield the left ventricular ejection fraction, although arrhythmia limited gating in a minority of patients. Transient cavity dilatation was reported on the basis of a combination of visual assessment and software-reported ratio.
ECG waveforms in digital format were extracted from our Marquette MUSE database (GE Healthcare, Milwaukee, Wisconsin) using custom software. These tracings provide 10 seconds of data in standard 12-lead arrangement. The Marquette software recorded the HR in beats per minute for each tracing, from which the RR interval in milliseconds was calculated. The HRs at baseline, peak exercise, and 1 and 5 minutes into recovery were recorded in beats per minute and converted to RR intervals in milliseconds (60,000/HR [beats/min]).
Assessing late HR recovery requires careful definition of the variable to study. The simple definition of peak HR minus HR at 5 minutes into recovery is unsuitable for several reasons. First, there is a high degree of correlation between this variable and standard 1-minute HR recovery (in our cohort, Pearson’s r 2 = 0.42 for the relation between 1- and 5-minute HR recovery defined in this manner), because a large component of 5-minute HR recovery occurs in the first minute. Furthermore, in late recovery, the HR trends toward the rest values. Thus, the difference between peak and late HR predominantly reflects rest and peak HR values. To provide an independent index of late HR recovery, we therefore assessed the percentage HR recovery. Late HR recovery at 5 minutes was defined as [1 − (RR 5min − RR peak )/(RR rest − RR peak )] × 100. This gives the percentage of the cycle length change between rest and peak exercise that has been recovered after 5 minutes. Lower values represent impaired recovery, by analogy with 1-minute HR recovery. In our cohort, the correlation between 1-minute HR recovery and this definition of late HR recovery at 5 minutes was not significant (Pearson’s r 2 <0.01, p = 0.43), establishing it as a suitable parameter of late HR recovery for evaluation. Late HR recovery was optimally dichotomized into normal and abnormal groups by maximizing the log likelihood ratio from a Cox proportional-hazards model with the limitation that each subgroup had ≥10% of the cohort.
All statistical tests were performed using R version 2.13 (R Project for Statistical Computing, Vienna, Austria) with missing data (6% of 6 affected variables) imputed by the MI (multiple imputation) package version 0.09–14 for multivariate models only. Continuous variables are expressed as mean ± SD or as median (interquartile range) for non-normal distributions. They were compared between groups using Student’s t tests (or Wilcoxon’s rank-sum tests for non-normal distributions) and among groups using analysis of variance (or Kruskal-Wallis tests for non-normal distributions). Continuous variables were visually examined using Q-Q plots to identify significant deviations from a normal distribution. Frequency variables are expressed as number (percentage) and were compared using chi-square or Fisher’s exact test. Kaplan-Meier curves and the log-rank test examined all-cause survival by quartile of HR recovery. All applicable tests were 2 tailed, and p values <0.05 were considered statistically significant.
Univariate Cox proportional-hazards models explored all variables for subsequent multivariate model building (see list in Table 1 ). Schoenfeld residuals tested the assumption of proportional hazards and found no significant deviations. Plots of β coefficients from univariate Cox proportional-hazards models were visually examined by quartile of each continuous variable for departures from linearity. Two variables, body mass index and HR at rest, appeared visually to have a J-shaped effect. Rest HR showed significant improvement in a nested Cox model after adding its squared term (change in chi-square 4.25, p = 0.039), while body mass index did not (change in chi-square 0.59, p = 0.44). The 2 variables were treated with simultaneous linear and second-order terms in all models. Plots of β coefficients fit a linear regression to a log-transformed version of the equation hazard ratio and its confidence limits = A × exp(k × variable), where A and k are constants, and added a linear offset to align the hazard ratio to unity for the reference quartile. Nested multivariate Cox proportional-hazards models included all variables apart from peak HR and late HR recovery for the baseline model. An incremental, nested Cox model added 5-minute HR recovery. The nested Cox proportional-hazards model provides a baseline chi-square value and a new chi-square value after adding a variable. Change in chi-square for the added variable (1 degree of freedom for late HR recovery) was compared using the chi-square distribution, which is 1 tailed.
| Variable | Cohort (n = 2,082) | Alive (n = 1,886) | Dead (n = 196) | p Value ⁎ | Hazard Ratio (95% CI) | p Value † |
|---|---|---|---|---|---|---|
| Age (years) | 55 ± 11 | 54 ± 11 | 63 ± 12 | <0.001 | 1.96 (1.72–2.23) | <0.001 |
| Men | 1,187 (57%) | 1,061 (56%) | 126 (64%) | 0.034 | 1.37 (1.03–1.84) | 0.033 |
| Hypertension | 764 (37%) | 659 (35%) | 105 (54%) | <0.001 | 2.09 (1.58–2.77) | <0.001 |
| Dyslipidemia (by history or prescribed medications) | 902 (43%) | 812 (43%) | 90 (46%) | 0.45 | 1.15 (0.87–1.52) | 0.34 |
| Diabetes mellitus | 238 (11%) | 189 (10%) | 49 (25%) | <0.001 | 2.82 (2.03–3.90) | <0.001 |
| Previous or current tobacco use | 279 (13%) | 246 (13%) | 33 (17%) | 0.15 | 1.39 (0.96–2.02) | 0.08 |
| Family history of coronary disease | 554 (27%) | 514 (27%) | 40 (20%) | 0.041 | 0.72 (0.51–1.02) | 0.07 |
| Body mass index (kg/m 2 ) | 27 (24–31) | 27 (24–31) | 27 (24–32) | 0.84 | 0.95 (0.83–1.09) ‡ | 0.49 |
| Previous mechanical revascularization | 262 (13%) | 208 (11%) | 54 (28%) | <0.001 | 2.88 (2.11–3.95) | <0.001 |
| Antiplatelet medications | 635 (30%) | 550 (29%) | 85 (43%) | <0.001 | 1.82 (1.37–2.42) | <0.001 |
| β blockers | 458 (22%) | 387 (21%) | 71 (36%) | <0.001 | 2.12 (1.58–2.84) | <0.001 |
| Angiotensin-converting enzyme inhibitors | 283 (14%) | 238 (13%) | 45 (23%) | <0.001 | 2.10 (1.51–2.94) | <0.001 |
| Anticholesterol medications | 625 (30%) | 553 (29%) | 72 (37%) | 0.033 | 1.43 (1.07–1.91) | 0.017 |
| Calcium channel blockers | 224 (11%) | 187 (10%) | 37 (19%) | <0.001 | 1.96 (1.37–2.81) | <0.001 |
| Diuretics | 233 (11%) | 190 (10%) | 43 (22%) | <0.001 | 2.42 (1.72–3.39) | <0.001 |
| Antihypertensive medications | 220 (11%) | 196 (10%) | 24 (12%) | 0.40 | 1.25 (0.82–1.92) | 0.30 |
| Exercise time (minutes) | 9.9 ± 3.2 | 10.1 ± 3.2 | 8.2 ± 3.0 | <0.001 | 0.83 (0.79–0.87) | <0.001 |
| Double product (mm Hg/min) | 28,392 ± 5,242 | 28,676 ± 5,128 | 25,668 ± 5,548 | <0.001 | 0.90 (0.88–0.92) | <0.001 |
| Treadmill angina | 118 (6%) | 106 (6%) | 12 (6%) | 0.75 | 1.00 (0.54–1.84) | 1.00 |
| Treadmill ECG changes | 461 (22%) | 407 (22%) | 54 (28%) | 0.06 | 1.39 (1.01–1.90) | 0.041 |
| Summed difference score | 0 (0–2) | 0 (0–2) | 0 (0–4) | <0.001 | 1.06 (1.04–1.09) | <0.001 |
| Transient cavity dilation | 101 (5%) | 86 (5%) | 15 (8%) | 0.08 | 1.86 (1.10–3.15) | 0.022 |
| Ejection fraction (%) | 59 ± 10 | 59 ± 10 | 56 ± 12 | 0.001 | 0.97 (0.95–0.98) | <0.001 |
| Rest HR (beats/min) | 72 ± 13 | 72 ± 12 | 73 ± 14 | 0.45 | 0.43 (0.20–0.95) § | 0.038 |
| Peak HR (beats/min) | 156 ± 18 | 158 ± 18 | 143 ± 18 | <0.001 | 0.68 (0.63–0.73) | <0.001 |
| 1-minute HR recovery (beats/min) | 23 ± 9 | 23 ± 9 | 18 ± 9 | <0.001 | 0.57 (0.48–0.66) | <0.001 |
| Late HR recovery (%) | 42% (31%–51%) | 42% (32%–51%) | 35% (24%–44%) | <0.001 | 0.28 (0.17–0.46) | <0.001 |
⁎ Comparison between alive and dead.
† Hazard ratio from univariate Cox model.
‡ Linear term hazard ratio listed, squared term 1.00 (95% confidence interval 1.00 to 1.00, p = 0.43).
§ Linear term hazard ratio listed, squared term 1.06 (95% confidence interval 1.01 to 1.12, p = 0.03).
Results
From August 1998 to December 2003, our nuclear cardiology laboratory performed 12,780 stress SPECT studies. Of these, 6,768 (53%) used treadmill exercise stress. The application of inclusion and exclusion criteria limited the final cohort to 2,082 patients, which represented 31% of all treadmill exercise studies during this period. Table 1 lists cohort characteristics and the univariate predictors of survival. During the mean follow-up period of 9.9 ± 1.5 years, typical risk factors were significant univariate predictors of survival. A cutoff of ≤13% produced an optimal binary variable of late HR recovery for predicting survival (univariate hazard ratio 2.26, 95% confidence interval 1.49 to 3.44, p <0.001). Table 2 lists cohort characteristics by quartile of late HR recovery. Patients with lower late HR recovery were older, more likely to be women, and had a higher incidence of diabetes. Interestingly, they also had higher rest HRs, lower peak exercise HRs, and higher 1-minute HR recovery.
