Heart rate recovery: validation and methodologic issues

Exercise testing

Patients underwent symptom-limited treadmill testing using the U.S. Air Force School of Aerospace Medicine (5)or an individualized ramp treadmill protocol (6). The physiologic distinction of these protocols is that the patient is subjected to small, frequent increments in workload rather than uneven increases every 3 min. Information gathered from a questionnaire enabled maximal exercise to be reached at approximately 10 min (7). Patients did not perform a cool-down walk but were placed supine as soon as possible after exercise. The reasons for termination were angina, >2 mm of abnormal ST depression, drop in systolic blood pressure or ominous arrhythmias. Visual ST-segment depression was measured at the J junction and corrected for pre-exercise ST-segment depression. An abnormal response was defined as 1 mm or more of horizontal or downsloping ST-segment depression. Blood pressure was taken manually, and metabolic equivalents (METS) were estimated from treadmill speed and grade. Heart rate was measured supine, standing, during each minute of exercise, at maximum exercise and in recovery at 1, 2, 3 and 5 min. Heart rate recovery was defined as (maximum HR—HR at specified time period after exercise) and represented the drop in HR during that time interval. No test was classified as indeterminate (8); medications were not withheld, and maximal HR target was not used as an end point. The exercise tests were performed, analyzed and reported with a standard protocol utilizing a computerized database.

Coronary angiography

Coronary artery narrowing was visually estimated and expressed as percent lumen diameter stenosis. Patients with a 50% narrowing of the left main, left anterior descending, left circumflex or right coronary arteries or their major branches were considered to have significant angiographic coronary artery disease (CAD). Severe disease was considered to be two-vessel disease if the proximal left anterior descending was involved; otherwise, three-vessel or left main disease were considered severe. The 50% criterion was chosen to be consistent with definitions used by the Coronary Artery Bypass Graft Surgery Trialists’ Collaboration (9). In addition, the Duke coronary artery jeopardy score was calculated (10). Ejection fraction was estimated from biplane left ventricular angiograms. Decisions for cardiac catheterization were consistent with clinical practice.

Follow-up

The social security death index was used to match all of the patients’ names to their social security numbers. The index was updated weekly, and the most current records were used. Death status was determined as of July 2000 and was 100% complete. No other information regarding hospitalizations, cardiac interventions or cause of death during the follow-up was known.

Statistical methods

All-cause mortality was used as the end point for follow-up, and coronary angiography was utilized for diagnostic gold standard.

Survival analysis was performed using Kaplan-Meier curves to compare variables and cut-points, and the Cox hazard function was used to demonstrate which variables were independently and significantly associated with death. Automatic selection of variables was performed with a Z value cutoff of 2 and 20 iterations. Hazard ratios were calculated along with their 95% confidence intervals.

Logistic regression was used to separate subjects into those with and without significant angiographic disease, based on clinical and measured exercise variables in the diagnostic subpopulation. Forward selection was used with entry at a significance level <0.05. The general linear logistic regression model used took the following form:

$\text{Probability}(0\text{to}1)=1/(1+{\text{e}}^{-[\text{a}+\text{bx}+\text{cy}...]})$

where a is the intercept, b and c are coefficients and x and y are variable values.

How well the models separated patients with and without a given outcome (abnormal angiogram or death) was assessed by means of the area under a receiver operating characteristic (ROC) curve, which ranged from 0 to 1, with 0.5 corresponding to no discrimination (i.e., random performance) and 1.0 to perfect discrimination.

Number Crunching System Software (Salt Lake City, Utah) was used for all statistical analyses.

Exercise test responses

No complications were encountered during testing. Results for the entire population, and specifically for those who survived and the 413 patients who died, along with significance levels for differences are presented in Table 2.

Validation of prior criteria

Our first examination of HR recovery was to apply the previously published threshold values for abnormal at the specified time points. Figure 1illustrates the Kaplan-Meier curves for three criteria (12 beats/min drop at 1 min after exercise, 18 beats/min drop at 2 min after exercise and 42 beats/min drop at 2 min after exercise), showing a significant ability of all criteria to predict mortality (1–3). Significant hazard ratios were associated with each specified criterion.

Best time/cut-point

In order to compare the candidate cut-points with each other and with the published cut-points in this population, they were entered into a proportional hazards regression model with automatic selection. Heart rate recovery measurements at 2 min recovery were superior to all other time periods, and a value of 22 beats/min was optimal at that time period as judged by the Z values and significance levels. Figure 2shows a Kaplan-Meier plot of this criterion for comparison with previously published cut-points presented in Figure 1.

The effect of beta-blockers and maximal HR

Previous results have been divided over the effects of beta-blockers. We found no difference in the Kaplan-Meier survival curves of those receiving beta-blockers when compared with those not receiving beta-blockers when both exhibited abnormal HR in recovery. The Kaplan-Meier curves in Figure 3demonstrate no difference in the prognostic value of our optimal criteria of a drop of 22 beats/min at 2 min whether or not patients were receiving beta-blockers. When beta-blocker administration was confirmed by not reaching target HR, the same result was obtained. In addition, the same result was obtained in those who did and those did not obtain target HR.

Comparison with other variables

We then determined how HR recovery ranks with traditional treadmill responses. To do this, we entered all of the traditional treadmill responses and several clinical variables that were univariately significant (resting HR and blood pressure along with HR recovery) into a proportional hazards regression model. The top variables chosen as significant predictors were METS, age, history of typical angina and HR drop at 2-min recovery. Notably, the treadmill angina index and exercise-induced ST depression—both ischemic variables—were not chosen.

Prognostic score

Because the ischemic components of the Duke treadmill score were not predictive of all-cause mortality, we elected to consider the combined use of METS and HR recovery. Figure 4presents a Kaplan-Meier plot using a METS criterion of 5 and HR recovery criterion of 22 beats/min as well as for abnormal by both criteria.

We also created a prognostic score using the variables chosen in the Cox hazard model and their coefficients as follows:

$\begin{array}{l}\text{Age}(\text{years})+10\cdot \text{definite}\text{angina}\text{pectoris}(0=\text{no},1=\text{yes})\\ -\text{HR}\text{drop}\text{at}2\text{min}\text{recovery}-5\cdot \text{METs}\end{array}$

We chose cut-points in this prognostic score by which to stratify patients into high-, intermediate- and low-risk as illustrated in Figure 5. The high-risk group had a hazards ratio of greater than five times. The score was entered along with data from cardiac catheterization into a proportional hazards regression model with automatic selection. The above score was chosen first with a Z value of 13.3 (p < 0.001) followed by ejection fraction with a Z value of 3 (p < 0.002) and, finally, the Duke jeopardy score with a Z value of 2.6 (p = 0.01). When cardiac catheterization variables were added to the Cox hazard model, they were superceded by the score variables.

Diagnostic characteristics

The diagnostic value of HR recovery (i.e., prediction of the presence of significant angiographic disease) was addressed in the subpopulation without MI. There was a 59% prevalence of significant angiographic CAD. When HR recovery was entered into the logistic regression model with variables previously noted to predict angiographic coronary disease (including maximal HR, ST depression, age and chest pain characteristics) (11,12), it was not chosen. The area under the curve of an ROC curve plotted with any of the continuous HR recovery measurements was never >0.58 (p < 0.001) as compared with the area under the ROC curve obtained for the ST response alone (0.67) or that from a previously published diagnostic score (0.79) (11).

This study

In our study, we attempted again to validate the use of HR recovery for prognosis in a male veteran population. The mortality rate in our study was higher than it was in previous studies of HR recovery. Using similar statistical analyses, we found that a decrease of HR in recovery of <22 beats/min at 2 min after exercise identified a high-risk group of patients. While these data confirm the utility of HR recovery, it is important to note that several studies have employed different cut-points for defining an abnormal test, even when the testing protocol (e.g., presence or absence of cool-down, minute of recovery when measurements are made) are the same. Therefore, it is difficult to know which value is most applicable to the general population.

Beta-blockade

We also found that the clinical administration of beta-blockers had no significant impact on the prognostic value of HR recovery. Given the impact of beta-blockers on the ability to achieve maximum HR, it would not be unreasonable to imagine that the medication might have an impact on HR recovery. However, the same results were found in those who reached target HR and those who did not. Further analysis of this issue would be warranted to better understand the true impact of this medication and maximal HR.

Multivariable analysis

Through multivariate analysis, we evaluated the power of several other clinical and treadmill variables to see how they compared with HR recovery in their ability to predict outcome. Similar to Cole et al. (2), we found that a low METS capacity was the most powerful variable associated with outcome. This finding was not shared by Nishime et al. (1), in which the strongest predictor of death was resting tachycardia. It is important to note that in our study and those of Nishime et al. (1)and Cole et al. (3), ischemic ST responses on the treadmill did not predict death. Exercise-induced angina was also not a significant variable in the study by Nishime et al. (1)nor in our study. The Duke treadmill score (18)includes both of these variables, and they have been predictors of cardiovascular outcome in several other studies.

This raises the question as to why ischemic variables included in the Duke score that clearly have diagnostic power do not predict all-cause mortality. While all-cause mortality has advantages over cardiovascular mortality as an end point (19), the Duke score was generated using the end points of infarction and cardiovascular death (18). Furthermore, cardiology interventions such as bypass surgery or catheter interventions were censored in the Duke study (i.e., subjects were removed from the survival analysis when these events occurred). Such censoring should increase the association of ischemic variables with outcome by removing patients whose disease has been alleviated and, thereby, would not be as likely to experience the outcome. We did not censor patients in this study on the basis of whether or not they had a cardiovascular procedure during follow-up because we do not have that information. From a previous study using a similar patient population, our group found that 75% of deaths were cardiovascular deaths, and 20% of patients were censored in follow-up due to bypass surgery (20). If the proportions of these statistics are similar in our current population, it would not be unreasonable to expect a bias against the predictive power of these variables. However, this hypothesis would need to be proven since censoring had no impact on HR recovery findings in the study that did censor (2)nor in the Duke score. In fact, exercise-induced ST depression predicted all-cause mortality with and without censoring in an earlier study (21). These contradictory results could potentially be due to the more effective methods of treatment currently available for coronary disease.

Diagnostic characteristics

A distinct advantage over previous studies is that we selected a group who underwent coronary angiography. This made it possible to evaluate the diagnostic ability of HR recovery. Surprisingly, HR recovery was not selected among the standard variables to be included in a logistic model, and the ROC curve did not indicate any discriminatory value. Thus, while HR recovery has been validated as an important prognostic variable, it did not help the diagnosis of coronary disease in this study.

Study limitations

While our study was unique in its strict adherence to guidelines in the performance of exercise testing and its inclusion of angiographic data, there were some limitations. The use of a ramp protocol differed from previous studies using the Bruce protocol, but, rather than being a limitation, this extends earlier findings to another protocol. The lack of inclusion of women is quite important, especially given the different characteristics of treadmill testing between women and men that have previously been shown. Additionally, we were unable to provide information on cause of death and did not have information about cardiovascular procedures during the follow-up period, so we could not censor on them. As mentioned earlier, the inability to censor may skew results. Without censoring, we are predicting prognosis in spite of, or in addition to, modern therapy rather than predicting who should have interventions. The pathophysiology of an inadequate HR decrease in recovery has not been explained by our study. Finally, as mentioned previously, our angiographic subset was subject to work-up bias.

Conclusions

We found that, among male patients with or without prior MI but without prior bypass surgery referred for clinical exercise testing done without cool-down walk but with prompt supine placement, an HR drop of 22 beats/min at 2 min recovery had prognostic, but not diagnostic, value. Heart rate at 1 or 2 min of recovery has been validated as a prognostic treadmill measurement and should be recorded as part of all treadmill tests. The prognostic power of this measurement does not appear to be affected by beta-blockade, but its cut-point is most likely affected by population selection and protocol. A score including HR recovery, METS, age and history of typical angina pectoris was superior to cardiac catheterization data for predicting prognosis. To date, this measurement has not been evaluated as a diagnostic indicator, and we found that it was not diagnostic of angiographic CAD. The drop of HR in recovery or its score should supplement, but not replace, the Duke treadmill score that has been validated as a predictor of infarct-free survival and diagnostic of angiographic CAD. Though our analysis has not explained the pathophysiology of an inadequate decline of HR after exercise, it may well represent a marker of habitual physical activity level (16).