Customized probability model for perioperative all-cause mortality.
Customize your JAMA Network experience by selecting one or more topics from the list below.
Kertai MD, Boersma E, Klein J, van Urk H, Poldermans D. Optimizing the Prediction of Perioperative Mortality in Vascular Surgery by Using a Customized Probability Model. Arch Intern Med. 2005;165(8):898–904. doi:10.1001/archinte.165.8.898
Accurately assessing the probability of perioperative mortality can be useful in preoperative risk assessment and management. This study aimed to revise and customize the revised cardiac risk (Lee) index to estimate the probability of perioperative all-cause mortality in patients undergoing noncardiac vascular surgery.
We studied 2310 patients (mean age, 67.8 ± 11.3 years; 1747 males) who underwent acute or elective major noncardiac vascular surgery between January 1, 1991, and December 31, 2000, at the Erasmus Medical Center, Rotterdam, the Netherlands. A total of 1537 patients were assigned for model development, in which the associations between predictor variables and mortality occurring within 30 days after surgery were identified to modify the Lee index, which was then evaluated in a validation cohort of 773 patients.
The perioperative mortality rates were similar in the development (n = 103 [6.7%]) and validation (n = 50 [6.5%]) populations. The customized risk-prediction model for perioperative mortality identified type of vascular surgery, ischemic heart disease, congestive heart failure, previous stroke, hypertension, renal dysfunction, and chronic pulmonary disease as being associated with increased risk, whereas β-blocker and statin use were associated with a lower risk of mortality. The performance of the customized index had excellent discriminative ability in both derivation and validation populations (concordance statistic, 0.88 and 0.85, respectively).
The customized index provides more detailed information than the Lee index about the type of vascular procedure, clinical risk factors, and concomitant medication use. The customized probability model can be a useful tool to estimate the risk of perioperative all-cause mortality and facilitate subsequent treatment strategies.
The purpose of preoperative risk assessment of patients undergoing major noncardiac surgery is to identify patients at increased risk for perioperative mortality and morbidity. Since perioperative cardiac mortality and morbidity are the most frequently occurring adverse events in noncardiac surgery, especially in vascular surgery, most previously developed risk indexes have focused only on the evaluation of perioperative cardiac risk.1,2 However, there is a strong relationship between perioperative cardiac and noncardiac complications and subsequent mortality; almost half of the patients who experience cardiac morbidity develop other types of noncardiac complications and mortality.3 Nevertheless, the currently available risk indexes estimate the risk of cardiac death rather than all-cause mortality, which may lead to inappropriate interpretations and underestimation of the risk of all-cause mortality.4
Furthermore, once patients are identified to be at increased risk for perioperative mortality, the most challenging issue is to modify the risk that may involve medical treatment or further interventions. β-Blockers and, to a lesser extent, statins have been shown to be beneficial in reducing noncardiac and cardiac mortality in patients undergoing noncardiac procedures including major vascular surgery.5-9 Again, however, the currently available risk indexes do not account for current medication use as such, and thus clinicians may not be able to accurately estimate patients’ likelihood of perioperative mortality. Therefore, the aim of this study was to develop a simple risk index that accounts for significant clinical risk factors and current medication use for the prediction of perioperative mortality in patients undergoing major vascular surgery.
As part of an ongoing research effort to improve perioperative risk assessment and treatment of patients undergoing major vascular surgery, a database was set up at the Erasmus Medical Center, Rotterdam, using the computerized hospital information system that contains medical files, surgical reports, anesthesiologic and postoperative charts, discharge letters, and records of outpatient clinic visits. With the help of this system, 2310 patients who underwent 2758 noncardiac vascular procedures performed between January 1, 1991, and December 31, 2000, were identified. Among these patients, 342 underwent multiple vascular procedures during the 10-year observation period. For the present study for these patients the initial vascular procedure was used as the time when they had surgery.
Data on the following risk factors of all-cause mortality were selected1,2,10,11: advanced age at operation (age >70 years); history of or current stable angina pectoris; history of myocardial infarction; congestive heart failure; stroke or transient ischemic attack; diabetes mellitus; hypertension; chronic pulmonary disease; and renal dysfunction. Data on noninvasive exercise testing such as dobutamine hydrochloride stress echocardiography or dobutamine technetium Tc 99m tetrofosmin myocardial perfusion scintigraphy, if performed, were also retrieved. The result of dobutamine stress echocardiography was considered positive if new wall-motion abnormalities occurred,11 and the result of perfusion scintigraphy was positive if a reversible perfusion defect was detected.12 Information about previous coronary revascularization was also noted.
Long-term medication use in patients with acute vascular conditions was noted on arrival at the hospital and for electively treated patients routinely at the outpatient clinic visit 1 month before vascular surgery. During the outpatient clinic visit, additional medications, including β-blockers or statins, were prescribed at the discretion of the attending physicians. Long-term medication use was ascertained if medication use was documented at least 1 to 3 months before hospital admission for vascular surgery and included angiotensin-converting enzyme inhibitors, aspirin, β-blockers, calcium channel blockers, diuretics, insulin, nitrate, and statins. According to a local practice, patients continued taking their medication on the evening before surgery and, if there was no contraindication, oral medication use was resumed on day 1 after surgery. Aspirin use was discontinued at least 10 days before major vascular surgery because of the perioperative use of low-molecular-weight heparin or planned anesthesia with neuraxial blockade, except for patients who underwent carotid endarterectomy.
The outcome chosen for the present study was all-cause mortality occurring before discharge or within 30 days after surgery. For patients who died at the Erasmus Medical Center, hospital records and autopsy results were retrieved and reviewed. For patients who died outside of the center, general practitioners were approached to ascertain the cause of death.
To derive a customized risk index for perioperative mortality after major vascular surgery, we used the revised cardiac risk index developed by Lee et al.2 The Lee index contains 6 predictors (high-risk surgery, ischemic heart disease, congestive heart failure, stroke, type 1 [insulin-dependent] diabetes mellitus, and renal dysfunction) of major cardiac complications. Since the Lee index was developed for predicting cardiac complications but not all-cause mortality, and the study population had a small proportion (3.8%) of patients undergoing major vascular surgery, certain modifications were made: the high-risk (abdominal aortic aneurysm surgery) vs low-risk (all other types of vascular surgery) categories were redefined as low risk (carotid endarterectomy), low-intermediate risk (infrainguinal bypass surgery), high-intermediate risk (abdominal and thoracoabdominal aortic surgery), and high risk (acute abdominal aortic aneurysm surgery) according to the published mortality rates of these procedures13-15; clinical characteristics such as advanced age, type 2 (non–insulin-dependent) diabetes mellitus, chronic pulmonary disease, and hypertension were considered as potential predictors; and information on β-blocker and statin use was taken into account. Other summary variables of the Lee index were also used, including ischemic heart disease (defined as a history of myocardial infarction or a positive exercise test result, and history of or current angina pectoris) and cerebrovascular disease (defined as a history of transient ischemic attack or stroke).
Sixty-seven percent (n = 1537) of the total cohort was randomly assigned to the derivation cohort, which we used to customize the Lee index. Clinical characteristics were compared between the derivation and validation cohorts by χ2 test for categorical variables and independent t test for continuous variables. Univariate logistic regression analyses were used to examine associations of the summary variables of the Lee index and each of the individual clinical risk factors, with perioperative mortality as the dependent variable. In subsequent multivariate regression analyses, the predictive performance of the Lee index with the original 6 predictors of perioperative mortality was evaluated. Thereafter, in a multivariate model, the risk of surgical procedures was changed to low, low-intermediate, high-intermediate, and high. Finally, along with the modified predictor of risk of surgical procedures and predictors of the Lee index, all univariate variables significant at a nominal 2-tailed P<.25 were then evaluated by means of a backward stepwise technique to create the final model containing variables with P<.05. Odds ratios and corresponding 95% confidence intervals are reported. The performance of multivariate logistic regression models was evaluated with the Hosmer-Lemeshow goodness-of-fit test16 and with the C (for concordance) statistic, which is identical to the area under the receiver operating characteristic curve.17
To develop a simple risk score for perioperative mortality, coefficients of the significant (P<.05) predictors of the final logistic regression model were multiplied by 10 and rounded to the nearest integer. The weighted scores were then assigned to each categorical predictor, which allowed a total risk score for each patient to be calculated. This then was applied to a probability plot, which shows the corresponding probability of perioperative mortality.18
The mean age of the entire cohort was 67.8 ± 11.3 years, and 1747 (76%) of the patients were men. A history of ischemic heart disease and hypertension was present in more than 36% of patients. Twenty-nine percent of patients had a history of cerebrovascular disease; almost 12%, diabetes mellitus; and less than 6%, renal dysfunction. The greatest percentages of patients underwent infrainguinal bypass surgery (36%), elective abdominal aortic surgery (29%), and carotid endarterectomy (22%). In total, 874 patients (38%) had a history of noninvasive testing: 720 patients underwent dobutamine stress echocardiography and 154 patients underwent myocardial perfusion scintigraphy. A history of a positive noninvasive stress test result was documented in 213 patients (24%). Perioperative mortality occurred in 153 patients (6.6%). Among these, there were 76 cardiac deaths (49%), 18 deaths (12%) due to septic complications, 14 cases of stroke syndrome (9%), 14 cases of hemorrhage (9%), 12 cases of renal failure (8%), 12 cases of intestinal necrosis (8%), and 7 cases of respiratory insufficiency (5%). When the 10-year study period was divided into 2 periods, from 1991 to 1995 and 1996 to 2000, the prevalence of type of surgery and clinical risk factors and the incidence of perioperative mortality did not show significant differences between the 2 periods.
The incidence of perioperative mortality was also similar in the development (6.7%) and validation (6.5%) cohorts. The patients who died in the derivation and validation cohorts had similar baseline characteristics except for statin use, which was more common in the validation cohort (Table 1). The incidence of perioperative mortality was also comparable in the derivation and validation cohorts during the periods from 1991 to 1995 and 1996 to 2000: 6.4% vs 6.9% (P = .70) in the derivation population and 6.6% vs 6.3% (P = .90) in the validation population.
Most of the predictors and summary variables of the Lee index were also significant univariate predictors of perioperative mortality in the derivation cohort (Table 2). There was a systematic increase in the risk of perioperative mortality in patients who underwent abdominal, thoracoabdominal, and acute abdominal aortic surgery compared with carotid endarterectomy (Table 2). Conversely, patients with a history of transient ischemic attack had a reduced incidence of perioperative mortality. Similarly, β-blocker and statin use were associated with a reduced incidence of perioperative mortality.
In a univariate logistic regression analysis, perioperative mortality significantly increased as the number of risk factors for the Lee index increased (Table 3). The C statistic (0.78) for the prediction of perioperative mortality using the Lee index showed good discriminative ability. In a subsequent multivariate analysis, the contribution of the original 6 predictors of the Lee index showed that high-risk type of surgery; ischemic heart disease; and histories of congestive heart failure, cerebrovascular disease, and renal dysfunction were significant predictors of perioperative mortality (Table 3). In fact, high-risk type of surgery was the strongest predictor of perioperative outcome, followed by congestive heart failure and renal dysfunction. On the contrary, although not significant, insulin therapy for diabetes indicated a risk reduction for perioperative mortality. The C statistic of the model was 0.82, and the Hosmer-Lemeshow goodness of fit was not significant for lack of fit (P = .20). When type of surgery was redefined, the results showed that the risk of mortality was substantially increasing from low- to high-risk type of vascular surgery, and the prognostic value of the other predictors remained the same (Table 3). The Hosmer-Lemeshow goodness of fit for this model was adequate (P = .30) and the discriminative ability of the model significantly improved (Table 3).
Finally, the multivariate logistic regression analysis with backward deletion identified risk of surgical procedures, cardiovascular morbidity, renal dysfunction, and chronic pulmonary disease as significant predictors of perioperative mortality. In contrast, the use of β-blockers and statins was associated with a reduced incidence of perioperative mortality (Table 3). The final multivariate model had excellent discriminative ability, and the Hosmer-Lemeshow goodness of fit for this model was not significant (P = .40).
The final logistic regression model with the 9 independent predictors of perioperative mortality was used to create a variable-weight index where scores were assigned on the basis of parameter estimates of the individual predictors. By summing the individual scores from the given predictors and using the total risk score, the patient’s probability of perioperative mortality can be estimated from the Figure.
Eight of the 9 predictors of the customized probability model remained significantly associated with increased risk of perioperative mortality (Table 4). The predictor variable of hypertension also showed a positive association with perioperative mortality, but this was statistically not significant. However, the odds ratios for perioperative mortality for hypertension were not significantly different in the derivation and validation cohorts. The performance of the customized probability model with the 9 predictors was slightly decreased when applied to the validation population (C statistic, 0.88 vs 0.85).
We have customized and validated a simple prediction tool that can be used to estimate the risk of perioperative mortality in patients undergoing acute or elective major vascular surgery. We have demonstrated that previously identified risk factors of the Lee index for cardiac complications were also predictors of perioperative all-cause mortality and had similar prognostic impact in a contemporary group of patients undergoing vascular surgery. By modifying the Lee index, the customized prediction model of perioperative mortality showed excellent discrimination and had a good ability to identify patients at increased risk.
The present study showed that 5 of the 6 original predictors of the Lee index (high-risk type of surgery, ischemic heart disease, congestive heart failure, cerebrovascular disease, and renal dysfunction) were associated with increased risk of perioperative all-cause mortality. This observation is in agreement with previous studies that showed that patients who require surgical treatment of peripheral vascular disease are at increased risk of perioperative mortality related to the surgical procedure, to concomitant coronary artery disease, and to renal and pulmonary comorbidities.10,19 However, the risk associated with type of vascular surgery as defined by the Lee index only differentiates between high-risk procedures (abdominal aortic surgery) and low-risk procedures (infrainguinal bypass, carotid endarterectomy, aortobifemoral bypass). This definition is probably inadequate, since there are substantial differences in complexity of vascular procedures categorized as low risk, with subsequently reported differences in mortality rates.20 In addition, the Lee index was developed in patients who underwent only elective vascular surgery. With this in mind, we redefined type of surgery on the basis of timing (acute or elective), complexity, and reported mortality rates, which substantially improved the predictive performance of the Lee index and confirms that type of vascular surgery was an important predictor of mortality.
We also found that the magnitude and direction of risk estimates of ischemic heart disease and cerebrovascular disease were similar to those of the Lee index. In contrast, the predictive value of congestive heart failure and renal insufficiency was greater, whereas the predictive value of insulin therapy for diabetes was not a significant correlate of perioperative mortality, and its predictive value suggested a decreased rather than increased risk. Indeed, the results of the final multivariate analysis also showed that all the clinical risk factors but diabetes mellitus remained predictors of perioperative mortality. The absence of either form of diabetes as a predictor of perioperative mortality may reflect a changing patient population or improved management of diabetic vascular disease with more intensive insulin management.21 Nevertheless, some other risk factors, such as hypertension and chronic pulmonary disease, became predictors of outcome, in agreement with previous studies.10,15 The American College of Cardiology–American Heart Association guidelines for perioperative evaluation identify severe hypertension as a condition that should be controlled before surgery when possible.22 Patients with poorly controlled hypertension or labile preoperative hypertension are at greater risk for perioperative dysrhythmias and myocardial ischemia.22 Poor pulmonary function is also often described as a significant predictor of perioperative mortality,10,15 and a relationship between cardiac and pulmonary complications is emerging.3 Combination of different perioperative strategies designed to emphasize respiratory therapy in high-risk patients may reduce the risk of pulmonary23 and subsequent cardiac3 complications.
The clinical applicability of our prediction rule in the perioperative treatment of patients undergoing vascular surgery should be considered along with the American College of Cardiology–American Heart Association guidelines for evaluation and management of perioperative cardiovascular risk.22 Accordingly, patients with acute abdominal aortic aneurysm should proceed to surgery without any delay, and not more than a cursory preoperative evaluation can be performed. Of note, 50% of these patients have underlying coronary artery disease, and thus appropriate cardiac monitoring, anesthesia, and medical therapy may significantly improve survival after surgery.24 Any other vascular procedures not thought to be urgent allow for more thorough evaluation of perioperative risk. In that respect, our customized probability model may help to identify patients at increased risk for perioperative all-cause mortality. There is consensus that if patients have any of the risk factors described in our probability model, β-blockers, given their favorable action, should be incorporated into perioperative risk-reduction strategies.11,22 Additional information on β-blocker use in our risk model may help to adjust the probability of mortality and subsequent decision for proceeding with surgery or the need for further risk-reduction strategies. Nevertheless, the group of patients identified by clinical risk factors and noninvasive testing as being at high risk often has considerable perioperative mortality despite β-blocker use.11 For these patients, a combination of β-blockers with statins may offer additional prevention.6,8,9 Therefore, we also adjusted our probability model for statin use, allowing further refinement of risk estimation. In patients in whom, despite the combination of β-blockers and statins, the estimated probability remains high and there is clear indication for coronary revascularization independent of the need for vascular surgery, a coronary intervention should be considered.22
Our study has some limitations. First, this was an observational study that relied on administrative data and medical records based on physician documentation of significant comorbidities. Thus, the effects of some of the risk factors, especially cardiac risk factors, may be biased. Nevertheless, the predictive values of these cardiac comorbidities were similar to those described by the original Lee index. Second, the customized probability model was based on a vascular surgery population operated on at a tertiary center. Therefore, the mortality rates for some of the procedures, such as the mortality rate after elective abdominal aortic aneurysm surgery, may appear higher than those reported from other studies (2.7% to 5.5%).10,15 However, these studies with lower mortality rates usually used strict inclusion criteria and selected patients at low risk for perioperative complications. Finally, discontinuation of aspirin treatment in patients with major vascular surgery could have led to an increased risk of perioperative mortality. However, mortality rates observed in our study were similar to those of some previously reported studies,1,22 and we found no association between discontinuation of aspirin use and increased risk of mortality.
In summary, our study showed that a combination of clinical predictors, type of vascular surgery, and current β-blocker and statin use are major determinants of perioperative mortality. The derived and validated customized probability model is a simple risk assessment tool with excellent discriminative ability that may allow clinicians to estimate the probability of perioperative mortality in patients undergoing acute or elective major vascular surgery.
Correspondence: Don Poldermans, MD, PhD, Department of Vascular Surgery, Room H921, Erasmus Medical Center, Dr Molewaterplein 40, 3015 GD Rotterdam, the Netherlands (firstname.lastname@example.org).
Accepted for Publication: October 28, 2004.
Financial Disclosure: None.
Create a personal account or sign in to: