A Multilevel Analysis of Surgical Category and Individual Patient-Level Risk Factors for Postoperative Stroke

Abstract

Background and Purpose:

Many studies supporting the association between specific surgical procedure categories and postoperative stroke (POS) do not account for differences in patient-level characteristics between and within surgical categories. The risk of POS after high-risk procedure categories remains unknown after adjusting for such differences in patient-level characteristics.

Methods:

Using inpatients in the American College of Surgeons National Surgical Quality Initiative Program database, we conducted a retrospective cohort study between January 1, 2000, and December 31, 2010. Our primary outcome was POS within 30 days of surgery. We characterized the relationship between surgical- and individual patient-level factors and POS by using multivariable, multilevel logistic regression that accounted for clustering of patient-level factors with surgical categories.

Results:

We identified 729 886 patients, 2703 (0.3%) of whom developed POS. Dependent functional status (odds ratio [OR]: 4.11, 95% confidence interval [95% CI]: 3.60-4.69), history of stroke (OR: 2.35, 95%CI: 2.06-2.69) or transient ischemic attack (OR: 2.49 95%CI: 2.19-2.83), active smoking (OR: 1.20, 95%CI: 1.08-1.32), hypertension (OR: 2.11, 95%CI: 2.19-2.82), chronic obstructive pulmonary disease (OR: 1.39 95%CI: 1.21-1.59), and acute renal failure (OR: 2.35, 95%CI: 1.85-2.99) were significantly associated with POS. After adjusting for clustering, patients who underwent cardiac (OR: 11.25, 95%CI: 8.52-14.87), vascular (OR: 4.75, 95%CI: 3.88-5.82), neurological (OR: 4.60, 95%CI: 3.48-6.08), and general surgery (OR: 1.40, 95%CI: 1.15-1.70) had significantly greater odds of POS compared to patients undergoing other types of surgical procedures.

Conclusions:

Vascular, cardiac, and neurological surgery remained strongly associated with POS in an analysis accounting for the association between patient-level factors and surgical categories.

Introduction

Postoperative stroke (POS) is associated with high mortality, prolonged hospital stays, and significant postoperative disability.^1–3 The risk of this important surgical outcome is related to factors that may be broadly categorized as either surgery- or patient-specific. Reported patient-specific factors for POS include age,^4–6 hypertension,⁷ ischemic cardiac disease,⁸ peripheral vascular disease,⁹ diabetes mellitus,⁴ aortic atherosclerosis,¹⁰ and prior stroke.^5,11–13 Surgery-related risk factors for POS include the indication for surgery,^14,15 as well as the category of surgery, such as vascular,^11,16 cardiac,^17,18 or general surgery.^8,19,20 While multiple studies have shown that rates of POS are higher in vascular and cardiac procedures than in general surgery, evidence is largely limited to single-center studies and cohorts undergoing a single category, or limited variety of procedures. Moreover, these studies do not account for the degree of confounding that may arise from clustering of patient-level characteristics within and between certain surgical categories.

The lack of investigations adjusting for this specific type of confounding raises the possibility that previously reported risks of POS associated with particular surgical categories may be due in part to unmeasured “clustering” of patient-level comorbidities. In order to characterize the association between both surgical and patient factors and POS while adjusting for clustering between individual-level and surgical risk factors, we conducted an analysis using an inpatient surgical cohort from the National Surgical Quality Initiative Program (NSQIP) encompassing a broad range of surgical procedures. We hypothesized that high-risk procedure categories would retain a significant association with POS, despite adjustment for patient-level factor clustering in a large cohort of patients undergoing multiple categories of surgery.

Methods

Study Design and Population

We conducted a retrospective study of all patients in the NSQIP cohort who underwent inpatient surgery between the dates of January 1, 2000, and December 31, 2010. National Surgical Quality Initiative Program is sponsored by the American College of Surgeons (ACS) and prospectively collects 135 demographic and clinical data points on randomly selected patients undergoing major surgery at over 350 nationwide participating clinical sites for the purpose of improving the quality of private sector, nontrauma surgical care. Cases are identified through a systematic sampling strategy whereby all surgical cases performed in the previous 8 days at an institution are chosen for data abstraction. Data points are abstracted and recorded in the NSQIP database for up to 30 days postoperatively by ACS-trained clinical reviewers. Clinical data are deidentified and include presurgical conditions, as well as postsurgical adverse events, including POS, mortality, and reoperation; collection methods include medical chart review and telephone interviews. To ensure quality standards, NSQIP data are regularly audited to assess the level of disagreement between all collected variables,^21,22 which was recently shown to be approximately 1.6%.²³

As per NSQIP guidelines, patients under the age of 18 and patients classified as brain-dead organ donors by American Society of Anesthesiologists score of 6 or more were excluded, as were those who underwent minor, trauma, or transplant surgery, surgery related to a postoperative complication from a previous operation, or additional surgery performed by a different operating team under the same anesthetic.²¹ Additionally, we excluded patients who underwent outpatient procedures, as these patients had an exceedingly low rate of POS within the 30-day postoperative period (data not shown). The institutional review board of Columbia University Medical Center approved this analysis of deidentified NSQIP data and waived the requirement for informed consent.

Measurements

We were primarily interested in modeling the relationship between patient-level and surgical category characteristics and the primary outcome, which was defined as the development of an embolic, thrombotic, or hemorrhagic cerebrovascular event occurring within 30 days of surgery involving motor, sensory, or cognitive dysfunction persisting for greater than 24 hours.²¹ We additionally sought to adjust the model for the effect of nonrandom clustering of patient-level characteristics with surgical categories.

Demographic data included age, sex, and race. Clinical characteristics included cigarette smoking within the year preceding surgery, alcohol use (defined as consumption of greater than 2 drinks per day), hypertension, diabetes mellitus, chronic obstructive pulmonary disease (COPD), myocardial infarction, percutaneous coronary intervention (PCI), revascularization or amputation for peripheral vascular disease, congestive heart failure (either chronic with exacerbation or new diagnosis in the 30 days preceding surgery), history of stroke or transient ischemic attack (TIA), in addition to hemodialysis, functional status 30 days prior to surgery, and surgical procedure category.

Functional status was categorized as either “independent” (not requiring assistance from another person for any activities of daily living) or “dependent” (requiring any level of assistance from another person for any activities of daily living). Surgical category was defined as the surgeon’s self-declared specialty or the surgical specialty area that best described the principal operative procedure.²¹

Data Analysis

We used the χ² test for proportional measures, and the t test or Wilcoxon rank-sum test for continuous measures where appropriate. We then calculated the incidence rate ratio (IRR) of POS for each surgical category individually. The IRR calculation for a given surgical category used a reference group comprised of procedures belonging to all other surgical categories (eg, reference for cardiac surgery included a composite of vascular, neurological, general, orthopedic, gynecologic, ophthalmologic, oral, plastic, thoracic, and urologic surgery). We first used logistic regression to determine unadjusted odds ratios (ORs) for POS for each individual surgical category, using the same reference groups as the IRR calculations, without adjusting for the clustering between patient-level factors and surgical categories. Next, to illustrate the clustering effect within each surgical group, we built an unadjusted logistic regression model that accounted for patient-level factor clustering within each surgical category. Last, we built a final multilevel logistic regression model to characterize the relationship between individual patient-level characteristics and surgical characteristics, while accounting for the clustering within each surgical category by incorporating variables that were significantly associated with the primary outcome in the unadjusted model. This allowed for the assessment of intragroup (within surgical category) and intergroup (between surgical category) differences with respect to the risk of POS, accounting for individual patient-level characteristics and surgical categories.

In both univariable and multivariable analyses, the surgical category reference was made up of gynecologic, ophthalmologic, oral, plastic, thoracic, and urologic procedures, which were combined due to their low IRRs of POS relative to all other surgical categories within the 30-day postoperative period. An α of .05 was set as the level of significance. All statistical analyses were performed using SAS (version 9.4; SAS Institute Inc, Cary, North Carolina). As the source data for the study were deidentified, it was not necessary to obtain informed consent from patients. The institutional review boards of Columbia University College of Physicians and Surgeons and Weill Cornell Medical College approved this study.

Results

Surgery-Specific Factors

We identified 729 886 surgical patients, of whom 2703 (0.3%) developed POS (Table 1). Patients who developed POS were more likely to have undergone cardiac surgery, vascular surgery, and neurosurgery when compared to patients who did not develop POS (P < .001; Table 1). In comparison to all other surgical categories, cardiac, vascular, and neurosurgical procedures were also significantly associated with an increased risk of POS, while general surgery was associated with a decreased risk of POS (Table 2). In the univariable logistic regression model, vascular, neurological, and cardiac surgical procedures all conferred an increased OR of POS in comparison to the reference group (Table 2). General surgery was associated with a decreased odds of POS compared to the reference group.

Table 1.

Patient Characteristics, Stratified by Primary Outcome.

Characteristica	No POSb (n = 727 183)	POS (n = 2703)	P Value
Age in years, mean (range)	58.0 (16.0-89.0)	71.0 (16.0-89.0)	<.001
Male	43 (43.5)	1320 (48.9)	<.001
Race			.052
American Indian	1929 (0.8)	7 (0.8)
Asian or Pacific Islander	4310 (1.8)	12 (1.4)
Non-Hispanic black	23 854 (10.2)	89 (10.3)
Hispanic, black	191 (0.1)	0 (0)
Hispanic, unknown	10 277 (4.4)	29 (3.3)
Hispanic, white	5795 (2.5)	10 (1.2)
Unknown	19 962 (8.5)	62 (7.1)
Non-Hispanic white	168 223 (71.7)	659 (75.9)
Diabetes mellitus	129 063 (17.8)	743 (27.5)	<.001
Hypertension	374 644 (51.5)	2,121 (78.5)	<.001
Active smoking	155 379 (21.4)	726 (26.9)	<.001
Alcohol use	21 641 (3.0)	128 (4.7)	<.001
Congestive heart failure	10 395 (1.4)	148 (5.5)	<.001
Myocardial infarctionc	7597 (1.0)	132 (4.9)	<.001
Peripheral vascular disease	42 133 (5.8)	322 (11.9)	<.001
Transient ischemic attack	26 296 (3.6)	397 (14.7)	<.001
Stroked	22 068 (3.0)	432 (16.0)	<.001
Chronic obstructive pulmonary diseasee	44 603 (6.1)	377 (14.0)	<.001
Percutaneous coronary interventionf	48 991 (6.7)	364 (13.5)	<.001
Hemodialysis	16 546 (2.3)	172 (6.4)	<.001
Surgical category			<.001
Cardiac surgery	6424 (0.9)	146 (5.4)
General surgery	497 269 (68.4)	1092 (40.4)
Gynecologic surgery	23 483 (3.2)	10 (0.4)
Neurosurgery	15 116 (2.1)	113 (4.2)
Ophthalmologic surgery	1 (0.0)	0 (0.0)
Oral surgery	4 (0.0)	0 (0.0)
Orthopedic surgery	45 550 (6.3)	86 (3.2)
Plastic surgery	3723 (0.5)	6 (0.5)
Thoracic surgery	6229 (0.9)	20 (0.7)
Urologic surgery	14 543 (2.0)	35 (1.3)
Vascular surgery	108 899 (15)	1182 (43.7)
Other	6 (0.0)	0 (0.0)
Unknown	1 (0.0)	0 (0.0)
Preoperative functional status			<.001
Independent	658 785 (90.6)	1901 (70.3)
Dependent	68 221 (9.4)	800 (29.6)
Unknown	174 (0.0)	2 (0.1)

Abbreviation: POS, postoperative stroke.

^aAll data are presented as number (%), unless otherwise specified.

^bPostoperative stroke.

^cDefined as myocardial infarction occurring during the 6 months preceding surgery.

^dDefined as any patient-reported history of stroke, irrespective of chronicity.

^eDefined as chronic obstructive pulmonary disease resulting in either functional disability, hospitalization at any time, chronic bronchodilator use, or forced expiratory volume (FEV1) less than 75% of predicted value.

^fDefined as previous percutaneous coronary intervention at any time, excluding valvuloplasty.

Table 2.

Incidence Rate Ratios and Univariable Odds Ratios for Primary Outcome, Stratified by Procedure Category.

Category	IRRa	95% CI	OR	95% CI
Cardiac surgery (N = 6570)	6.38	5.37-7.55	6.46	5.41-7.58
Vascular surgery (N = 110 081)	4.41	4.08-4.76	4.41	4.10-4.76
Neurosurgery (N = 15 229)	2.05	1.69-2.48	2.06	1.70-2.45
General surgery (N = 498 361)	0.31	0.29-0.34	0.33	0.29-0.34
Otherb (N = 93 690)	0.40	0.30-0.47	0.42	0.36-0.49

Abbreviations: CI, confidence interval; IRR, incidence rate ratio; OR, odds ratio.

^aReference groups for each IRR calculation refer to procedures belonging to all other surgical categories (eg, reference for cardiac surgery refers to a composite of vascular, neurological, general, orthopedic, gynecologic, ophthalmologic, oral, plastic, thoracic, and urologic procedures.)

^bRefers to orthopedic, gynecologic, ophthalmologic, oral, plastic, thoracic, and urologic surgical procedures.

In the univariable cluster-adjusted model, cardiac, vascular, neurological, and general surgery conferred an increased odds of POS in comparison to the reference group (gynecologic, ophthalmologic, oral, plastic, thoracic, and urologic surgery). In the final, multivariable cluster-adjusted model, cardiac, vascular, neurological, and general surgical procedures conferred an increased odds of POS compared to the reference group (Table 3). The only patient-level characteristic that was significantly associated with surgical category was dependent preoperative functional status (P < .001), although preoperative functional status did not modify the association between surgical category and POS (P = .61; data not shown).

Table 3.

Univariable and Multivariable Cluster-Adjusted Odds Ratios for Primary Outcome, Stratified by Patient-Level Risk Factors.

Risk Factor	Unadjusted		Adjusted
	OR	95% CI	OR	95% CI
Surgery-specific
Cardiac surgery	15.44	12.07-19.75	11.25	8.52-14.87
General surgery	1.39	1.15-1.66	1.40	1.15-1.70
Neurological surgery	5.13	3.95-6.67	4.60	3.48-6.08
Vascular surgery	7.43	6.18-8.93	4.75	3.88-5.82
Othera	Ref	Ref	Ref	Ref
Patient-specific
Age	1.05	1.05-1.05	1.03	1.03-1.04
Female sex	0.77	0.71-0.84	1.06	0.97-1.16
Diabetes	1.84	1.66-2.04	1.06	0.97-1.16
Smoking	1.45	1.33-1.59	1.20	1.08-1.32
Myocardial infarction	5.71	4.70-6.95	1.44	1.15-1.80
Percutaneous coronary intervention	2.29	2.03-2.58	1.01	0.88-1.15
Dependent functional status	6.32	5.64-7.09	4.11	3.60-4.70
Congestive heart failure	5.00	4.16-6.00	1.28	1.04-1.58
Chronic obstructive pulmonary disease	2.54	2.24-2.87	1.39	1.21-1.59
Hypertension	3.45	3.12-3.81	2.11	1.89-2.36
Peripheral vascular disease	2.42	2.12-2.76	0.72	0.62-0.84
Transient ischemic attack	5.21	4.64-5.84	2.49	2.19-2.82
Stroke	6.52	5.78-7.34	2.35	2.06-2.70
Acute renal failure	5.40	4.32-6.75	2.35	1.85-2.99

Abbreviations: CI, confidence interval; OR, odds ratio; Ref, Reference.

^aComprised of orthopedic, gynecologic, ophthalmologic, oral, plastic, thoracic, and urologic surgical procedures.

Patient-Level Factors

Patients who developed POS were significantly more likely to be older, male, and have vascular risk factors than patients who did not develop POS (Table 1). In the univariable analysis, tobacco smoking, diabetes mellitus, myocardial infarction, PCI, congestive heart failure, history of stroke or TIA, dependent functional status, severe COPD, hypertension, peripheral vascular disease, and acute renal failure all increased odds of POS. Compared to male sex, female sex was associated with decreased odds of POS (Table 3). In the final, adjusted multivariable model, dependent preoperative status, history of stroke, history of TIA, acute renal failure, active smoking, hypertension, myocardial infarction, and COPD were associated with an increased odds of POS. After adjustment for medical comorbidities and surgical category clustering, diabetes, sex, and PCI were no longer significantly associated with POS, while peripheral vascular disease was associated with decreased odds of POS (Table 3).

Discussion

In a multilevel analysis of a large, nationwide, inpatient surgical cohort from the United States, we found that patients undergoing cardiac, vascular, and neurosurgical procedures were significantly more likely than patients undergoing other categories of surgery to develop POS within 30 days. Although general surgery was associated with a lower IRR and OR of POS than the nonneurological, noncardiac, nonvascular reference group in unadjusted analyses, general surgery was associated with an increased, yet much more modest, risk of stroke when compared to relatively lower risk procedures for POS in the final multivariable analysis. We also found that patients with ischemic cardiac disease, history of stroke or TIA, and COPD and acute renal failure were more likely to develop POS, independently of the degree of clustering between patient-level factors and surgical category, although some previously described patient-level factors for POS were not shown to be significantly associated with the primary outcome.

Previous noncluster-adjusted studies confirm the association between POS and both cardiac and vascular surgical procedures.^{1,9–11,14–16,20} These results are biologically plausible due to the fact that the aorta, cardiac valves, and carotid arteries are frequently manipulated in such procedures, thereby resulting in disruption of atherosclerotic plaques or embolic phenomena from local thrombus formation.²⁴ Extended use of cardiopulmonary bypass during cardiac surgery, which is associated with POS,^9,25 may further explain the particular increased risk of POS associated with cardiac procedures. In addition to direct manipulation of the aorta, the mechanistic underpinnings include relative hypotension and hypoperfusion leading to border-zone cerebral infarctions, as well as formation of thrombi in the setting of extracorporeal circulation or cardiac hypokinesis.

Neurosurgical—and particularly neurovascular—procedures are also associated with postoperative cerebral ischemia. In a NSQIP-based retrospective study, 7.3% of patients undergoing cerebrovascular surgery developed POS within 30 days, conferring an OR of 7.18 (95% confidence interval [95% CI]: 3.95-13.06) for developing POS.²⁶ Postoperative ischemic lesions on diffusion-weighted brain imaging are common after resection of newly diagnosed gliomas^27,28 and POS has been described following epilepsy surgery.²⁹ Plausible mechanistic explanations for these findings include surgical disruption of dural and parenchymal vessels, with subsequent thrombosis or hemorrhage.

Finally, while patients who underwent general surgery had reduced IRR and noncluster adjusted OR when compared to all nongeneral surgical procedures, general surgical patients had a slightly increased odds of POS in the final, adjusted analysis compared to surgical categories that were associated with the lowest IRR of POS. The difference in the IRR and cluster-adjusted OR associated with general surgery likely reflects the choice of denominator used in the calculation of the IRR, which included surgical categories associated with a higher risk of POS than the relatively lower risk surgical categories chosen as reference for the OR calculation.

With respect to patient-level risk factors, our findings mostly mirrored those from previous unadjusted analyses, which describe prior stroke^5,11,12,16 or TIA,⁶ hypertension,^7,25 acute renal failure,^6,30 ischemic cardiac disease,⁸ and COPD³¹ as notable risk factors for POS. The association between dependent presurgical functional status and postsurgical stroke has been described, particularly within the setting of carotid endarterectomy.^16,31–33 While age,^4–6,30 diabetes mellitus,^4,25 and peripheral vascular disease⁹ are associated with POS, in the final adjusted analysis, age was only modestly associated with POS. Moreover, diabetes mellitus, while showing a significant association in univariable analysis, was not associated with POS after adjusting for covariates, and peripheral vascular disease was associated with reduced adjusted odds of POS. While patients with peripheral vascular disease may reasonably be expected to harbor systemic atherosclerosis, one possible explanation for this result is that peripheral vascular disease may have served as a marker for antiplatelet or dual-antiplatelet therapy, which may have affected the risk of POS. Alternatively, peripheral vascular disease may also have contributed to greater perioperative mortality, which could have reduced the number of patients with POS during the follow-up period. Additional known risk factors for POS, such as atrial fibrillation,³⁴ aortic atherosclerosis,¹⁰ and high transfusion requirements⁹ were not data points collected by the NSQIP, and therefore, their impact on the risk of POS could not be determined in our study.

The disappearance of diabetes and sex as significantly associated risk factors for POS in the multivariable analysis may have been due to the relatively stronger association between POS and other factors, such as functional dependence, surgical category, history of stroke, or acute renal failure. Moreover, the discrepancy compared to existing studies may also reflect a degree of sampling bias, whereby the patients included within the NSQIP cohort may have had a lower incidence of vascular risk factors than the populations that were analyzed in many previous studies. However, this would not fully explain our results, which did show a degree of association between other vascular risk factors, such as history of stroke, TIA, and hypertension, and POS.

Our study benefits from its use of a large multicenter sample comprised of patients undergoing a wide-ranging variety of surgical procedures. The use of such a cohort allows for multilevel modeling and provides a more generalizable analysis of the rates and risks of POS than some of the extant studies. In addition, we employed a multilevel model to adjust for the effect of clustering of patient-level factors with specific surgical categories on the risk of POS, thereby controlling for the degree of clustering between medical comorbidities and surgical categories.

However, our study has multiple limitations. First, neither clinician adjudication nor neuroimaging confirmation was required to make the determination of the primary outcome. Therefore, the basis on which POS was determined is potentially inaccurate. Second, many established clinical risk factors for POS, such as atrial fibrillation, were not collected in NSQIP data, therefore raising the possibility that unmeasured risk factors could have partially explained the associations between specific surgical categories and POS. Third, the timing of the primary outcome event was not available in the NSQIP data, and we were therefore unable to differentiate whether POS events occurring outside the immediate perioperative period may have been due to unmeasured nonsurgical factors. Fourth, our dichotomization of functional status did not allow a more granular analysis of the relationship between functional status and POS. Finally, the results do not entirely exclude the possibility of confounding by indication, whereby patients who underwent particularly high-risk surgical procedures, such as carotid endarterectomy, may have done so as a result of a stroke.

Our results confirm that cardiac, vascular, and neurosurgical procedures are associated with a high risk of POS. Although we were not able to adjust for all POS risk factors, the results of our analysis suggest that the increased risk of POS after cardiac, vascular, and neurosurgical procedures may reflect factors related to the surgical procedures themselves rather than the medical comorbidities that often accompany patients undergoing these same high-risk procedures. Additionally, after adjusting for the effect of clustering, some well-established patient-level risk factors were no longer significantly associated with POS. These results should be interpreted with caution and should be confirmed in additional analyses of multiple multicategory surgical cohorts that address the limitations of our current study. Furthermore, given that silent POS also occurs in among surgical patients in the postoperative period, future areas of inquiry could include similar multilevel analyses as the present study, stratified by subclinical or clinical POS.