Development and validation of a prediction model for strokes after coronary artery bypass grafting

CARDIOVASCULAR

Development and Validation of a Prediction Model for Strokes After Coronary Artery Bypass Grafting David C. Charlesworth, MD, Donald S. Likosky, PhD, Charles A. S. Marrin, MB, BS, Christopher T. Maloney, MD, Hebe B. Quinton, MS, Jeremy R. Morton, MD, Bruce J. Leavitt, MD, Robert A. Clough, MD, and Gerald T. O’Connor, DSc, PhD, for the Northern New England Cardiovascular Disease Study Group Department of Surgery, Catholic Medical Center, Manchester, New Hampshire; Departments of Medicine, Surgery, and Community and Family Medicine, and The Center For Evaluative Clinical Sciences, Dartmouth-Hitchcock Medical Center, Hanover, New Hampshire; Department of Surgery, Maine Medical Center, Portland, Maine; Department of Surgery, Fletcher Allen Health Care, Burlington, Vermont; Department of Surgery, Eastern Maine Medical Center, Bangor, Maine

Background. A prospective study of patients undergoing coronary artery bypass graft surgery (CABG) was conducted to identify patient and disease factors related to the development of a perioperative stroke. A preoperative risk prediction model was developed and validated based on regionally collected data. Methods. We performed a regional observational study of 33,062 consecutive patients undergoing isolated CABG surgery in northern New England between 1992 and 2001. The regional stroke rate was 1.61% (532 strokes). We developed a preoperative stroke risk prediction model using logistic regression analysis, and validated the model using bootstrap resampling techniques. We assessed the model’s fit, discrimination, and stability. Results. The final regression model included the following variables: age, gender, presence of diabetes, presence of vascular disease, renal failure or creatinine

greater than or equal to 2 mg/dL, ejection fraction less than 40%, and urgent or emergency. The model significantly predicted (␹2 [14 d.f.] ⴝ 258.72, p < 0.0001) the occurrence of stroke. The correlation between the observed and expected strokes was 0.99. The risk prediction model discriminated well, with an area under the relative operating characteristic curve of 0.70 (95% CI, 0.67 to 0.72). In addition, the model had acceptable internal validity and stability as seen by bootstrap techniques. Conclusions. We developed a robust risk prediction model for stroke using seven readily obtainable preoperative variables. The risk prediction model performs well, and enables a clinician to estimate rapidly and accurately a CABG patient’s preoperative risk of stroke.

S

operative atrial fibrillation [6 – 8]. In addition, research has associated strokes with processes of care, such as prolonged cardiopulmonary bypass [9, 10]. Inferences from these reports are traditionally limited by small sample sizes, which provide unstable estimates of risk of stroke after CABG surgery. In addition, the identification of preoperative risk factors is of limited value without a mechanism to implement it in daily clinical practice. The Northern New England Cardiovascular Disease Study Group (NNECDSG) is a voluntary research consortium, composed of clinicians, research scientists, and hospital administrators, representing all medical centers in Maine, Vermont, and New Hampshire, and one medical center in Massachusetts where CABG surgery is performed. Since 1987, the NNECDSG has maintained a prospective registry of all patients undergoing cardiac surgery in the region. The group fosters continuous improvement in the quality of care of patients with cardiovascular disease in the region through the pooling of process and outcome data and the timely feedback of data to clinicians. Outcomes data on 33,062 consecutive isolated CABG procedures performed in this region, from 1992 to 2001, provided an opportunity to identify patient and disease factors associated with stroke, and to develop and vali-

troke is a devastating complication of coronary artery bypass graft (CABG) surgery. As such, the development of a user-friendly and robust risk prediction model will likely benefit both patients and physicians. Clinical researchers have documented neurologic complications, including global encephalopathy and focal neurologic syndromes, after cardiac surgery [1, 2]. These deficits, most likely caused by hypoxia, embolism, hemorrhage, or metabolic factors, have varied widely in severity and permanence [3]. The reported incidence of perioperative stroke, excluding subclinical strokes, ranges from 1.3% to 4.3% [4, 5]. Whereas advances have been made in many aspects of cardiac surgery, such as in diminishing the occurrence of low output syndrome and mortality, perioperative stroke has remained as a substantial problem [2, 5]. Case reports and series have documented the association of numerous patient risk factors with the development of perioperative stroke, including: increased patient age, vascular disease, arteriosclerosis, congestive heart failure, diabetes mellitus, prior CABG surgery, and postAccepted for publication March 10, 2003. Address reprint requests to Dr Charlesworth, Catholic Medical Center, 100 McGregor St, Manchester, NH 03102; e-mail: charlesworth@nhheart. com.

© 2003 by The Society of Thoracic Surgeons Published by Elsevier Inc

(Ann Thorac Surg 2003;76:436 – 43) © 2003 by The Society of Thoracic Surgeons

0003-4975/03/$30.00 PII S0003-4975(03)00528-9

date internally a risk prediction model to estimate the risk of perioperative stroke.

Patients and Methods Data Collection We conducted a prospective cohort study of stroke associated with CABG surgery at all NNECDSG centers, using data collected on consecutive patients undergoing isolated CABG surgery with or without extracorporeal circulation. For this analysis, we did not include patients having CABG surgery incidental to heart valve repair or replacement, resection of a ventricular aneurysm, or other surgical procedures. Previous publications by the NNECDSG have discussed in detail our data collection methodology and definitions [11]. In short, we prospectively collected the following preoperative variables: age, gender, diabetes, vascular disease (cerebrovascular disease: prior stroke, prior transient ischemic attack, prior carotid surgery, carotid stenosis or bruit; lower extremity disease: claudication, amputation, prior lower extremity bypass, absent pedal pulses, or lower extremity ulcers), congestive heart failure (this or prior admission), renal failure requiring dialysis, preoperative serum creatinine, treated chronic obstructive pulmonary disease, left main stenosis, preoperative ejection fraction (EF), left ventricular end diastolic pressure (LVEDP), number of diseased coronary arteries, history of cardiac surgery, recent myocardial infarction (⬍ 1 week), and current unstable angina [12]. The cardiothoracic surgeons, using definitions previously described, assessed priority of surgery (elective, urgent, emergency) [11]. We defined a stroke as a new focal neurologic deficit that appears and is still at least partially evident more than 24 hours after its onset, occurring during or after the CABG procedure and established before discharge.

Statistical Analysis To compare the characteristics of patients having a stroke to those free from stroke, we used standard statistical methods (␹2 for categorical data and t tests for continuous variables). Multivariate logistic regression was used to explore the association between strokes and potential risk factors and to adjust for potentially confounding variables; it was performed using the STATA 7.0 program (Stata Coroporation, College Station, TX) [13]. The logistic regression model predicts the probability of an event (stroke), conditional on patient and disease characteristics specified by the analyst. Using this logistic regression model, we calculated the predicted risk of stroke for an individual with any combination of demographic or disease characteristics. We tested our model’s discrimination via the relative operating characteristic (ROC) curve. The ROC curve graphically represents the relationship between the truepositive rate (ie, the sensitivity of the test) and the false-positive rate (ie, the specificity of the test). A suitable single-valued measure of test accuracy is the area of

CHARLESWORTH ET AL PREDICTING STROKE RISK AFTER CABG SURGERY

437

the graph lying beneath the curve. In a review by Swets of ROC curves, a test with a ROC score of 0.50 “can achieve that performance by chance alone.” Such a test has no discrimination, and “the true- and false-positive proportions are equal.” A test with a ROC score of 1.0 has perfect discrimination [14]. After development of the multivariate risk prediction model, we calculated each patient’s predicted risk of stroke, and internally validated the model using bootstrap techniques [15]. We developed the multivariate model on the entire data set, and randomly drew with replacement 200 samples of 100%. We then calculated the ROC curve area for each sample. A clinical risk assessment tool was developed from the multivariate risk prediction model and was scored by rounding the adjusted odds ratio for each variable (see Fig 5). These weights were then summed. The relationship between this clinical risk score and the probability calculated from the risk prediction model was read from a graph. This risk assessment tool therefore approximates the risk that would have been calculated from the risk prediction model. We assessed the calibration of the risk prediction model by applying it to the individual patients in the data set and comparing the observed and expected numbers of strokes by decile of predicted risk. We calculated the Lemeshow-Hosmer goodness of fit statistic. We used the relative contribution of each of the patient and disease characteristics to the prediction of risk of peri-operative stroke as a measure of the total ␹2 uniquely associated with each factor [16].

Results Univariate Trends in Rates of Stroke There were a total of 532 strokes among 33,062 patients. The crude rates of stroke were relatively stable from 1992 to 2001 (1.59 vs 1.72; Fig 1). The crude rate of stroke increased with increasing age (0.49 vs 3.11; Fig 2).

Patient Characteristics The odds of stroke increased nearly sevenfold for patients more than 80 years versus those younger than 55 years (OR ⫽ 6.55, ptrend, ⬍ 0.001) (Table 1). Female patients had a 1.3-fold higher odds of stroke than male patients (p ⫽ 0.004). Diabetics had a 1.6-fold increased risk of stroke versus nondiabetics (p ⬍ 0.001). Patients with renal failure or creatinine greater than or equal to 2 mg/dL,or vascular disease had higher odds of stroke than those without either of these comorbidities (p ⬍ 0.001).

Trends in Treatment Practices, Angiography, and Hemodynamics Patients with lower ejection fraction (EF) had increased odds of stroke (ptrend, ⬍ 0.001) (Table 2). Those with left main stenosis or left ventricular end-diastolic pressure (ptrend, ⬍ 0.001) and greater number of diseased vessels (ptrend, ⬍ 0.001) had increased odds of having a stroke. Those with emergency surgery (ptrend, ⬍ 0.001), unstable

CARDIOVASCULAR

Ann Thorac Surg 2003;76:436 – 43

438

CHARLESWORTH ET AL PREDICTING STROKE RISK AFTER CABG SURGERY

Ann Thorac Surg 2003;76:436 – 43

Fig 1. Rates of stroke after isolated coronary artery bypass surgery in northern New England between 1992 and 2001. CARDIOVASCULAR

angina (p ⫽ 0.05), prior CABG surgery (p ⫽ 0.32), and recent myocardial infarction (MI) (p ⫽ 0.26) had an increased odds of developing a stroke.

Development and Validation of the Multivariate Risk Prediction Model All variables associated with the risk of stroke in univariate analyses (p ⬍ 0.10) were entered into a multivariate analysis using logistic regression. Nonsignificant variables were dropped from the model. The final model is summarized in Table 3. Variables in the final model included age, presence of diabetes, EF less than 40%, gender, priority of surgery, renal failure or creatinine greater than or equal to 2 mg/dL, and vascular disease. The regression model significantly (␹2 [14 d.f.] ⫽ 258.72, p ⬍ 0.0001) predicted the occurrence of stroke in this data set. This final model was then compared with the full model, which included all variables found significantly associated with stroke in univariate analyses) through the Fig 2. Age-stratified rates of stroke after isolated coronary artery bypass surgery in northern New England between 1992 and 2001.

respective ROC curves (full model ⫽ 0.71, final model ⫽ 0.70). The discrimination did not change appreciably between the models. Coefficients from the logistic regression model may be used to calculate predicted probability of perioperative stroke for an individual patient. The calculation as well as the parameterization of the variables in the model are described in Table 3. The correlation between the observed and expected number of strokes was high (r ⫽ 0.99), and the Lemeshow-Hosmer goodness-of-fit statistic was not statistically significant (␹2 [8 d.f.] ⫽ 6.31, p ⫽ 0.6129), indicating that there was no statistically signifi cant departure from a perfect fit (Fig 3). To study discrimination of the prediction model, we calculated the area under the ROC curve. This risk prediction model had an ROC area of 0.70, indicating moderate ability to distinguish between individuals who will experience a stroke and those who will not. We used resampling techniques to internally validate the risk

Ann Thorac Surg 2003;76:436 – 43

CHARLESWORTH ET AL PREDICTING STROKE RISK AFTER CABG SURGERY

439

Table 1. Univariate Associations Between Patient Characteristic and Disease Variables, and Risk of Stroke

Age (years) ⬍55 55–59 60 – 64 65– 69 70 –74 75–79 ⱖ80 Body mass index (kg/m2) ⬍25.0 25.0 –27.7 27.8 –31.1 ⱖ31.2 Body surface area (m2) ⬍1.60 1.60 –1.79 1.80 –1.99 ⱖ2.00 Gender Male Female Chronic obstructive pulmonary disease No Yes Diabetes No Yes Congestive heart failure No Yes Renal failure or creatinine ⱖ 2 mg/dL No Yes Vascular disease No Yes

Percent of Subjects

Stroke (%)

Odds Ratio

18.63 11.62 14.54 18.25 17.45 13.18 6.33

0.49 0.83 1.31 1.72 2.21 2.52 3.11

1.00 1.72 2.71 3.59 4.64 5.29 6.55

24.94 24.88 25.00 25.17

5.46 17.70 32.92 43.92

1.81 1.59 1.69 1.35

1.79 1.61 1.95 1.32

1.00 0.87 0.93 0.74

1.36 1.22 1.48 1.00

Confidence Interval (95%)

p Value

Reference 1.04 –2.83 1.76 – 4.20 2.38 –5.39 3.11– 6.91 3.53–7.94 4.24 –10.13 ptrend

⬍0.001

Reference 0.69 –1.11 0.74 –1.17 0.58 – 0.95 ptrend

0.040

0.93–1.99 0.95–1.56 1.22–1.81 Reference ptrend

0.030

CARDIOVASCULAR

Variable

71.72 28.28

1.48 1.92

1.00 1.30

Reference 1.09 –1.56

0.004

89.33 10.67

1.52 2.32

1.00 1.54

Reference 1.21–1.95

⬍0.001

69.37 30.63

1.37 2.14

1.00 1.57

Reference 1.32–1.87

⬍0.001

98.39 1.61

14.18 23.50

1.00 1.86

Reference 1.52–2.28

⬍0.001

96.75 3.25

1.53 3.91

1.00 2.62

Reference 1.90 –3.61

⬍0.001

80.88 19.12

1.30 2.90

1.00 2.27

Reference 1.90 –2.72

⬍0.001

prediction model. We randomly drew 200 samples each containing 100% of the total number of subjects. The risk prediction model was applied to each sample, and the area under the ROC curve was calculated. There was a tight distribution around a mean ROC area of 0.70 (95 % CI, 0.67 to 0.72). The relative importance of each patient and disease characteristic in predicting the risk of perioperative stroke is summarized in Figure 4. We grouped the variables in our model into three categories: patient characteristics, patient presentation, and comorbidities. We calculated the percentage of total explainable preoperative risk for each of these categories. Patient characteristics (age and gender) contributed 52% of the total predictive ability of the model. Comorbidities (diabetes, renal failure, or creatinine greater than or equal to 2

mg/dL, and vascular disease) contributed 32%. Patient presentation (priority of surgery and EF) contributed 16% of the predictive ability of the model.

Comment In this regional prospective study of isolated consecutive CABG patients, we developed and validated a multivariate preoperative risk prediction model for the risk of stroke. The prediction model uses seven readily obtainable preoperative patient and disease characteristics assigning an independent weight to each to provide quantitative information about risk. The prediction model demonstrated relatively strong discriminatory ability (area under ROC curve ⫽ 0.70) and excellent parameterization (ie, high correlation between observed and ex-

440

CHARLESWORTH ET AL PREDICTING STROKE RISK AFTER CABG SURGERY

Ann Thorac Surg 2003;76:436 – 43

Table 2. Univariate Associations Between Angiographic and Hemodynamic Data, Past Medical History, and Risk of Stroke Variable CARDIOVASCULAR

Ejection fraction (%) ⱖ60 50 –59 40 – 49 ⬍40 Left main stenosis (%) ⬍50 50 – 89 ⱖ90 Left ventricular end-diastolic pressure (mm Hg) ⬍14 15–18 19 –21 ⱖ22 Number of diseased vessels 1 2 3 Priority Elective Urgent Emergency Unstable angina No Yes Prior coronary artery bypass grafting No Yes Prior percutaneous coronary angiography No Yes Recent myocardial infarction No Yes

Percent of Subjects (%)

Stroke (%)

Odds Ratio

Confidence Interval (95%)

p Value

41.62 25.22 17.58 15.58

1.27 1.49 1.61 2.36

1.00 1.18 1.28 1.89

Reference 1.46 –2.43 0.97–1.68 1.44 –2.42 ptrend

⬍0.001

Reference 1.15–1.73 1.15–2.24 ptrend

⬍0.001

Reference 0.61–1.11 0.91–1.78 1.22–1.99 ptrend

⬍0.001

Reference 0.67–1.16 1.13–1.85 ptrend

⬍0.001

Reference 1.23–1.88 1.82–3.39 ptrend

⬍0.001

75.27 19.62 5.11

39.30 23.75 11.49 25.46

14.81 37.24 47.95

31.62 61.65 6.74

1.45 2.03 2.30

1.33 1.09 1.69 2.05

1.34 1.22 1.99

1.14 1.72 2.78

1.00 1.41 1.60

1.00 0.82 1.28 1.55

1.00 0.88 1.45

1.00 1.52 2.49

41.07 58.93

1.44 1.72

1.00 1.20

Reference 1.00 –1.43

0.05

93.70 6.30

1.59 1.87

1.00 1.18

Reference 0.85–1.64

0.32

82.71 17.29

1.64 1.42

1.00 0.87

Reference 0.68 –1.10

0.25

87.52 12.48

1.58 1.82

1.00 1.15

Reference 0.90 –1.48

0.26

pected values [r ⬎ 0.95], with no significant deviation from perfect fit [significant Lemeshow-Hosmer goodness of fit test]). The discriminatory ability of the prediction model was moderate when tested among important patient subgroups. This prediction model, which makes use of routinely available preoperative data, can serve the clinician and the patient by providing a simple method to assess accurately the risk of stroke for patients anticipating CABG surgery. Also, by accounting for patient variability, the model provides a basis for tracking outcomes of cardiac surgery programs. This model may also be useful by providing the foundation for understanding the effects of treatment variables on stroke risk. There are a few limitations to this study. Aortic and carotid imaging studies are not performed routinely in our region. This is similar to the experience reported by Mick-

leborough and colleagues [17]. In our registry, information on carotid disease is combined with other vascular disease. Our definition of vascular disease includes lower extremity, aortic, and carotid disease. We believe that if one were to separate the carotid disease from the lower extremity and aortic disease, this variable would remain an independent predictor of stroke. In data reported by Mickleborough and colleagues, 65% of patients with carotid disease had other manifestations of vascular disease [17]. Birkmeyer and colleagues, in an extensive medical record review of 2,871 consecutive patients discharged alive after CABG surgery, noted that vascular disease is a marker for more systemic arteriosclerosis [18]. He linked this registry information with the National Death Index to examine the long-term mortality risk of vascular disease. He found that those with vascular disease were at a

Ann Thorac Surg 2003;76:436 – 43

CHARLESWORTH ET AL PREDICTING STROKE RISK AFTER CABG SURGERY

441

Table 3. Prediction of Risk of Stroke

Age (years) 55–59 60 – 64 65– 69 70 –74 75–79 ⱖ80 Diabetes Ejection fraction ⬍ 40% Female gender Priority Urgent Emergency Renal failure or creatinine ⱖ 2 mg/dL Vascular disease Intercept

Adjusted OR 1.63 2.56 3.25 4.07 4.54 5.51

Coefficients

p Value

1.41 1.39 1.04

0.485528 0.939827 1.177941 1.404483 1.513644 1.706814 ptrend 0.346638 0.329046 0.038858

⬍0.001 0.001 0.005 0.68

1.34 2.26 1.70

0.290969 0.814316 0.528380

0.007 ⬍0.001 0.002

1.79

0.583754 ⫺5.878808

⬍0.001

Model: ␹2[14 d.f.] ⫽ 258.72, p ⬍ 0.0001. Calculation of predicted risk using patient data and the logistic regression coefficients. Calculate the ODDS using the patients values and the coefficients: Stroke outcomes: ODDS ⫽ EXP(⫺5.878808 ⫹ [0.485528* age 55–59] ⫹ [0.939827* age 60 – 64] ⫹ [1.177941* age 65– 69] ⫹ [1.404483* age 70 –74] ⫹ [1.513644* age 75–79] ⫹ [1.706814* age ⱖ 80] ⫹ [0.346638* diabetes] ⫹ [0.329046* EF ⬍ 40%] ⫹ [0.038858* gender] ⫹ [0.290969* priority urgent] ⫹ [0.814316* priority emergency] ⫹ [0.528380* renal failure or creatinine ⱖ 2 mg/dL] ⫹ [0.583754* vascular disease]. Use the ODDS to calculate the PREDICTED PROBABILITY: Predicted Probability ⫽ ODDS/(1 ⫹ ODDS).

higher risk of mortality than those without vascular disease. In addition, those with multilevel vascular disease had an almost 1.5 times higher odds of death than those without multilevel arteriosclerosis. These findings suggest that a diagnosis of one’s vascular disease is sufficient for knowing information regarding a patient’s carotid disease and extent of lower extremity disease. In our current study, we had information concerning the presence or location of vascular disease for 54% of our

Fig 4. The relative contributions of each factor to predicting the risk of perioperative stroke. Patient characteristics: age and gender. Comorbidities: diabetes, renal failure or creatinine greater than or equal to 2 mg/dL, and vascular disease. Patient presentation: priority and ejection fraction.

patients. Of those with vascular disease, 36.6% had cerebrovascular disease, 29.4% had lower extremity disease, 19.4% had both, and 15% had vascular disease with insufficient information concerning its location. This large overlap between types of vascular disease suggests that our strategy of using a broad definition of vascular disease captures much of the same information contained in carotid imaging studies. These findings are in agreement with work done by Long and colleagues, who quantified the association between peripheral arterial disease and carotid occlusive disease [19]. Long and colleagues found that, among patients who underwent noninvasive testing for peripheral arterial and carotid disease and who did not have surgical interventions on either their leg or carotid arteries, there was a significant correlation between peripheral arterial disease and the number of diseased carotid arteries (⬎ 50% stenosis). Our data set, though large, was collected from eight medical centers in a single region of the United States. We are confident that our findings are representative of the experience at other centers around the country, as the NNECDSG definition for stroke is quite similar to the definitions used by other large registries, such as the Fig 3. Observed versus expected strokes by decile of predicted risk.

CARDIOVASCULAR

Variables

442

CHARLESWORTH ET AL PREDICTING STROKE RISK AFTER CABG SURGERY

CARDIOVASCULAR

Society for Thoracic Surgeons (STS) [20]. In addition, our rate of stroke is comparable with rates reported by the STS. It was our intent to address some of the limitations in the peer-reviewed literature concerning unstable estimates of preoperative risk factors, by undertaking a multicenter prospective observational study of more than 30,000 consecutive CABG surgeries. Several other multivariate analyses of risk factors for stroke have been performed. Wareing and colleagues reported on 1,200 consecutive patients more than 50 years of age undergoing CABG surgery [21]. They found an association between perioperative stroke and the following: increasing age, presence of arteriosclerosis of the ascending aorta, and severe carotid artery disease. D’Agostino and colleagues reported on 1,835 consecutive patients undergoing isolated CABG, with 42 patients (2.5%) having a perioperative stroke [22]. Multivariate analysis revealed associations between perioperative stroke and postinfarction angina, vascular disease, history of stroke or transient ischemic attack (TIA), duration of cardiopulmonary bypass, aortic arteriosclerosis, use of amrinone or epinephrine, and postoperative atrial fibrillation. Newman and colleagues presented a report of a multicenter preoperative stroke risk index for patients undergoing CABG [1]. Using a systematic sampling interval unlike our consecutive series, they enrolled 2,417 patients at 24 US academic medical centers. The perioperative neurologic complication rate was 3.2% (including major neurologic events, stroke, TIA, or persistent coma). Using logistic regression techniques, they developed a multivariate risk prediction model that included age, unstable angina, history of neurologic disease, prior CABG, diabetes, history of vascular disease, and history of pulmonary disease. This method showed good discrimination (ROC area ⫽ 0.78) and was validated using bootstrap techniques. Differences in our model versus the one presented by Newman and colleagues reside in the definition of a neurologic deficit. Newman and colleagues had a higher complication rate due in part to the inclusion of comas and TIAs, whereas we have chosen to include only focal deficits. Although Newman and colleagues did not present 95% confidence limits for their ROC area, we presume that our model’s ROC (0.70) would be included within that range. Mickelborough and colleagues reported risk factors for stroke among 1,631 CABG patients at a single medical center [17]. Among these, 170 patients were screened for carotid disease. Multivariate analyses revealed vascular disease, carotid artery occlusion, advanced age and history of prior stroke as significant preoperative predictors of perioperative stroke. Roach and colleagues reported results of data on 2,108 patients at 24 US medical institutions [2]. Of these, 3.1% had focal central nervous system injury, stupor, or coma at discharge. Roach and colleagues identified predictors of both type 1 (death due to stroke or encephalopathy, nonfatal stroke, TIA, stupor, or coma before discharge) and type 2 (nonfocal neurologic decline or deficit). Predictors of type 1 deficits were proximal aortic arteriosclerosis, history of neurologic disease, and age. Predictors of type 2 deficits were age, systolic hypertension noted at admission, pulmonary disease, and extreme

Ann Thorac Surg 2003;76:436 – 43

Fig 5. Risk prediction card for stroke. (CABG ⫽ coronary artery bypass grafting; EF ⫽ ejection fraction.)

alcohol consumption. Stamou and colleagues reported results from 16,528 consecutive patients undergoing CABG surgery over 10 years at a single institution. Multivariate analyses included preoperative (chronic renal insufficiency, recent MI, previous stroke, carotid disease, hypertension, diabetes, age ⬎ 75 years, EF ⬍ 34%) and postoperative (low cardiac output and atrial fibrillation) variables [23]. Because the postoperative variables may have occurred after some of the strokes, the inclusion of these as predictor variables is held in doubt. Discrepancies between our model and Stamou and colleagues’ may reside in differing definitions for stroke, a single institution’s experience, and the use of postoperative variables as predictors of stroke. Female gender has previously been found significantly associated with neurologic injury secondary to CABG sur-

gery. Hogue and colleagues examined this relationship in The Society of Thoracic Surgery National Cardiac Surgery Database [24]. In this data set, females accounted for 32% of the population. Neurologic outcomes were classified as new permanent global or focal deficits, transient ischemic attacks, or coma. Female gender was associated univariately with a 57% increased risk of neurologic injury (p ⬍ 0.001). After adjusting for 22 preoperative risk factors, female gender remained a significant predictor of neurologic injury (OR, 1.22; p ⬍ 0.001). Much of the difference in risk may be attributed to the difference in the identified outcome. We predicted the risk of a stroke (rate of 1.6%), whereas Hogue and colleagues predicted the risk of one of three neurologic events (rate of 3.3%) and did not separate out the association between gender and each type of neurologic event. There is general agreement in the observational literature that increasing age, the presence of arteriosclerosis, and in some studies, the presence of diabetes, are associated with increased risk of perioperative stroke [25]. However, as acute strokes are a contraindication for surgery, processes of clinical care (namely intra- and postoperative factors) in part create perioperative strokes. Possible culprits may include low preoperative hematocrit, duration of cardiopulmonary bypass, the application of the aortic cross clamp, and low cardiac output syndrome. The development and validation of a risk prediction model for stroke provides quantitative knowledge of the associations between patient and disease risk factors and incidence of stroke. This work on preoperative risk factors has paved the way for future work aimed at understanding the effects of processes of clinical care on the incidence of perioperative stroke. Clinicians may use this model, which uses seven easily obtainable preoperative variables, to link clinical outcomes not only to patient consultations through the framing of discussions, but also to operative care through improvement of surgical outcomes using preoperative patient risk predictions. In our region, surgeons use a pocket-sized risk prediction card (developed out of our preoperative stroke model) (Fig 5) in counseling patients through the accurate preoperative prediction of their risk. Accurate prediction of stroke may be used for studying treatment factors associated with the development of perioperative stroke, and may serve as a benchmark for quality improvement processes locally and nationally. This study was supported in part by an Individual NRSA Post-Doctoral Fellowship Award (F32 HL68357-01) to DS Likosky.

CHARLESWORTH ET AL PREDICTING STROKE RISK AFTER CABG SURGERY

4. 5. 6. 7.

8. 9. 10.

11.

12. 13. 14. 15. 16. 17.

18.

19.

20. 21. 22.

References 1. Newman MF, Wolman R, Kanchuger M, et al. Multicenter preoperative stroke risk index for patients undergoing coronary artery bypass graft surgery: Multicenter Study of Perioperative Ischemia (McSPI) Research Group. Circulation 1996;94:II74 –80. 2. Roach GW, Kanchuger M, Mangano CM, et al. Adverse cerebral outcomes after coronary bypass surgery: Multicenter Study of Perioperative Ischemia Research Group and the Ischemia Research and Education Foundation Investigators [see comments]. N Engl J Med 1996;335:1857–63. 3. Harrison MJ. Neurologic complications of coronary artery

23. 24. 25.

443

bypass grafting: diffuse or focal ischemia? Ann Thorac Surg 1995;59:1356 –8. Blossom GB, Fietsam R Jr, Bassett JS, Glover, JL, Bendick PJ. Characteristics of cerebrovascular accidents after coronary artery bypass grafting. Am Surg 1992;58:584 –9. Jones EL, Weintraub WS, Craver JM, Guyton RA, Cohen CL. Coronary bypass surgery: is the operation different today? J Thorac Cardiovasc Surg 1991;101:108 –15. Ricotta JJ, Faggioli GL, Castilone A, Hassett JM. Risk factors for stroke after cardiac surgery: Buffalo Cardiac-Cerebral Study Group. J Vasc Surg 1995;21:359 –4. Frye RL, Kronmal R, Schaff HV, Myers WO, Gersh BJ. Stroke in coronary artery bypass graft surgery: an analysis of the CASS experience: the participants in the Coronary Artery Surgery Study. Int J Cardiol 1992;36:213–21. Lynn GM, Stefanko K, Reed JFR, Gee W, Nicholas G. Risk factors for stroke after coronary artery bypass. J Thorac Cardiovasc Surg 1992;104:1518 –23. Libman RB, Wirkowski E, Neystat M, Barr W, Gelb S, Graver M. Stroke associated with cardiac surgery: determinants, timing, and stroke subtypes. Arch Neurol 1997;54:83–7. Brown WR, Moody DM, Challa VR, Stump DA, Hammon JW. Longer duration of cardiopulmonary bypass is associated with greater numbers of cerebral microemboli. Stroke 2000;31:707–13. O’Connor GT, Plume SK, Olmstead EM, et al. A regional prospective study of in-hospital mortality associated with coronary artery bypass grafting: the Northern New England Cardiovascular Disease Study Group [see comments]. JAMA 1991;266:803–9. Coronary artery surgery study (CASS): a randomized trial of coronary artery bypass surgery: survival data. Circulation 1983;68:939 –50. Stata, Stata Statistical Software. Release 7.0. 2001, College Station, TX: Stata Corporation. Swets JA. Measuring the accuracy of diagnostic systems. Science 1988;240:1285–93. Efron B, Tibshirani R. Bootstrap methods of standard errors, confidence intervals, and other methods of statistical accuracy. Stat Sci 1986;1:54 –77. Knaus WA, Wagner DP, Zimmerman JE, Draper EA. Variations in mortality, and length of stay in intensive care units [see comments]. Ann Intern Med 1993;118:753–61. Mickleborough LL, Walker PM, Takagi Y, Ohashi M, Ivanov J, Tamariz M. Risk factors for stroke in patients undergoing coronary artery bypass grafting. J Thoracic Cardiovasc Surg 1996;112:1250 –9. Birkmeyer JD, Quinton HB, O’Connor NJ, et al. The effect of peripheral vascular disease on long-term mortality after coronary artery bypass surgery: Northern New England Cardiovascular Disease Study Group. Arch Surg 1996;131:316–21. Long TH, Criqui MH, Vasilevskis EE, Denenberg JO, Klauber MR, Fronek A. The correlation between the severity of peripheral arterial disease and carotid occlusive disease. Vasc Med 1999;4:135–42. Data Analyses of the STS National Cardiac Surgery Database. January 1999, Society of Thoracic Surgeons. Wareing TH, Davila-Roman VG, Daily BB, et al. Strategy for the reduction of stroke incidence in cardiac surgical patients [see comments]. Ann Thorac Surg 1993;55:1400 –8. D’Agostino RS, Svensson LG, Neumann DJ, Balkhy HH, Williamson WA, Shahian DM. Screening carotid ultrasonography and risk factors for stroke in coronary artery surgery patients. Ann Thorac Surg 1996;62:1714 –23. Stamou SC, Hill PC, Dangas G, et al. Stroke after coronary artery bypass: incidence, predictors, and clinical outcome. Stroke 2001;32:1508 –13. Hogue CW Jr, Murphy SF, Schechtman KB, VG Da-Ra. Risk factors for early or delayed stroke after cardiac surgery. Circulation 1999;100:642–7. Craven TE, Ryu JE, Espeland MA, et al. Evaluation of the associations between carotid artery atherosclerosis, and coronary artery stenosis: a case-control study. Circulation 1990; 82:1230 –42.

CARDIOVASCULAR

Ann Thorac Surg 2003;76:436 – 43