Predictors of Outcome After Perioperative Stroke

Overview

Perioperative stroke is a devastating complication of surgery that is currently poorly characterized with limited clinical tools available to detect and prevent its occurrence. The current literature has identified that patients who experience a stroke after surgery have a higher rate of mortality, length of stay and discharge to a facility, but given the rare nature of this complication relatively little is known about which factors predict these outcomes amongst those who experience a perioperative stroke. The study objectives are to identify predictors of mortality, length of stay and discharge disposition after perioperative stroke in non-cardiac, non-neurological surgery using the prospectively-collected American College of Surgeons National Surgical Quality Improvement Program database between 2004 and 2020.

BACKGROUND Perioperative stroke is a devastating complication of surgery that is currently poorly characterized with limited clinical tools available to detect and prevent its occurrence. Perioperative stroke is a cerebrovascular event that occurs after surgery, and affects between 0.1-1.9% of patients having non-cardiac, non-neurologic surgery. Perioperative stroke is relatively understudied compared to postoperative complications of similar incidence and severity, such as cardiac complications. Over the past few decades, significant efforts have been undertaken to reduce the risk of perioperative myocardial infarction and have led to a decrease in incidence over time due to an advancements in risk stratification and perioperative management. Although the incidence of perioperative stroke is similar or higher than that of perioperative myocardial infarction, the risk of this complication has been increasing over time and perioperative stroke remains relatively neglected. The current literature has identified that patients who experience a stroke after surgery have a higher rate of mortality, length of stay and discharge to a facility, but given the rare nature of this complication relatively little is known about which factors predict these outcomes amongst those who experience a perioperative stroke. OBJECTIVES SPECIFIC OBJECTIVES 1. Derive and externally validate risk prediction models for mortality (primary outcome), adverse discharge, ICU utilization, tracheostomy/PEG/craniectomy, and length of stay after perioperative stroke. 2. Describe temporal trends in mortality after perioperative stroke between 2005 and 2020. 3. Explore mediators of mortality risk after perioperative stroke. METHODS This study is a retrospective analysis of the prospectively-collected American College of Surgeons National Surgical Quality Improvement Program database between 2004 and 2020. The derivation cohort will be extracted from the NSQIP database. The most recent year of the dataset (2020) includes data on over one million cases from 708 sites. The validation cohort will be extracted from Population Data BC, a collection of linked provincial databases providing de-identified individual level data on all residents of BC (population 5.1 million). A date range of January 1, 2002 to December 31, 2020 was chosen to reflect contemporary stroke treatment practices (e.g., EVT), and the consistent use of ICD-10-CA coding since 2002. Study population NSQIP derivation cohort: Patients with perioperative stroke in the NSQIP database between 2004 and 2017 will be included. We will exclude neurosurgical (including carotid endarterectomy) and interventional radiology procedures. The cardiac and non-cardiac surgery populations will be analyzed separately given their unique stroke risk profile. PopDataBC validation cohort: Patients with AIS will be identified using diagnosis codes I63.0-163.9 from the International Classification of Diseases (ICD), Tenth Revision, Canada (ICD-10-CA), which accurately and reliably identifies AIS in large administrative datasets, with positive predictive value and sensitivity of greater than 97%. Patients who experienced an ICD-10-CA admission diagnosis (top 3 positions) of AIS within 30 days of admission for surgery, or as a discharge diagnosis from the index surgical hospitalization. Perioperative patients will be further extracted from the overall AIS cohort if they received surgical billing code in the 30 days prior to admission for stroke. Outcome variables: Primary outcome is 30-day mortality; secondary outcomes are ICU utilization, length of hospital stay, markers of severe stroke outcome (tracheostomy, PEG, hemicraniectomy for cerebral edema), and adverse discharge (non-home facility or death). Candidate predictor variables: Outcome after perioperative stroke is potentially related to patient, surgical, and anesthetic factors, as well as characteristics of the stroke. Candidate predictor variables will include patient characteristics (age, sex, comorbidities including history of stroke), surgical characteristics (specialty, complexity, type, emergency status), anesthetic technique (general vs regional/neuraxial), and stroke characteristics (timing relative to operation, readmission for stroke vs inpatient stroke). Continuous variables will be considered for transformation using fractional polynomials to allow a continuous non-linear association. STATISTICAL ANALYSIS Objective 1: Derivation and validation of risk prediction models. Multivariable models to predict 30-day mortality following perioperative stroke will be created separately for cardiac and non-cardiac surgery, given their unique mechanisms of stroke. To avoid over-fitting, we will undertake a data reduction strategy and exclude variables with greater than 10% missing data or less than 20 observations. Where \>1% but \<10% data are missing, we will consider multiple or mean imputation; variables with \<1% missing data will be handled through complete case analysis. Pre-specified predictor variables will be used to construct a logistic regression model where coefficients will be shrunken using elastic net penalization. This method for variable selection and penalization was chosen because we anticipate possibility of a large number of potential predictors relative to the number of outcomes and that several important predictors will be collinear. By using elastic net regularization, we can employ a balance between least absolute shrinkage and selector operator (LASSO) penalization and RIDGE regression. RIDGE regression will allow us to keep important collinear predictors, but shrink their coefficients (which could be inflated due to collinearity). LASSO methods will allow the coefficients of some non-contributory predictors to zero, therefore, eliminating them from the model. Included variables will be assessed for interactions and additional terms considered to improve model performance as needed. Specifically, we will a priori examine the following interactions: age\*gender, surgical complexity (WRVU)\*age, WRVU\*specialty, and specialty\*anesthetic technique. Model discrimination will be evaluated using the area under the receiver operating characteristic curve (c-statistic). Model calibration will be assessed with a loess-smoothed plot of observed vs predicted risks over the risk spectrum. A similar analysis will be used to create a prediction model for length of stay (which will be a log-transformed linear model). As death is a competing outcome for discharge disposition, adverse discharge will be modeled as an ordinal outcome (home, non-home discharge, or death). Internal validation: Following derivation, 5,000 bootstrap samples will be used for internal validation and to generate an estimated optimism. Sensitivity analyses will test model performance in different time epochs to assess the impact of year of surgery (e.g., 2 to 3-year epochs, depending on the number of observations available in each year). Following derivation of the final model to predict mortality, we will derive a simplified predictive index by converting the regression coefficients to points that reflect their relative weights. The predictive accuracy of the resulting tool will be assessed in a similar fashion to above. External validation: The final logistic model will be externally validated in the distinct dataset obtained from PopDataBC after mapping NSQIP data elements to the PopDataBC database. Regression coefficients will be applied to PopDataBC from the NSQIP-derived model to calculate expected outcome probabilities and compare predicted to observed outcomes. Discrimination and calibration will be measured, with an area under the receiver operating characteristic curve \>0.70 considered acceptable model discrimination and Hosmer-Lemeshow test with p\>0.05 indicating acceptable goodness of fit. Objective 2: Temporal trends in mortality. We will perform an exploratory unadjusted ordinary least squares regression model with annual mortality rate after perioperative stroke as the dependent variable and year as the predictor to estimate the yearly change in mortality rate over time. A multivariable linear regression model will be specified, adjusting for important predictors. Objective 3: Mediators of mortality risk. To explore possible effect modification of pre-stroke mortality risk by post-stroke events, we will specify a series of logistic regression models with death as the dependent variable and the probability of death from our primary model as a linear predictor. We will then add pre-specified predictors (above), each as an additional variable, to explore whether the probability of death increases or decreases based on addition of the postulated effect modifying variable. Gender/Sex-based analysis: We will explore the relationship between gender and outcomes after perioperative stroke by performing subgroup analyses on women and men, and characterizing and comparing the baseline features and outcomes of these two populations. We will further include gender a priori as a predictor in the adjusted model, and examine interactions with other variables. We are unable to examine non-binary gender as NSQIP provides only "male", "female", or "unknown". Sample size and power: We used the methods of Riley et al to estimate the required sample size to derive a stable logistic regression model that minimizes overfitting. Using 30 model parameters and an estimated outcome prevalence of 25% mortality in people with stroke,3 we estimated the required minimal sample sizes based on potential model discrimination (c-statistic) and explained variance (Cox-Snell R2). If our model achieves a c-statistic of 0.7 (R2 0.09), we require 2,754 participants. Stronger discrimination (c-statistic 0.8, R2 0.2) would reduce the required sample size to 1,145. The NSQIP database 2004-2017 contains 6.6 million patients, and based on our preliminary analysis of the 2015 dataset, 99.5% will be non-cardiac surgery with a 0.19% incidence of stroke, and 0.5% will be cardiac surgery with a 1.76% incidence of stroke. We therefore estimate a total of at least 12,434 and 5,808 non-cardiac and cardiac perioperative strokes, respectively. A p value \<0.05 will be considered significant for all analyses and all data analysis will be performed using STATA 17 (StataCorp, Texas, USA).