Rationale

The goal of brain tumor resection is to improve or restore quality of life and/or control disease progression. Unfortunately, the consequences of PCVA, particularly in patients with terminal illness, sometimes render the operation futile. Thus, preoperative risk stratification to select candidates with a favorable benefit/risk ratio and possibly mitigate modifiable factors (e.g., uncontrolled diabetes and atrial fibrillation) is essential. Existing literature has suggested a handful of risk factors associated with PCVA in a neurosurgical context; the current study seeks to build on these findings.^15–20 While identification of individual-level predictors of PCVA is important, there remains a critical need for the development and validation of survey instruments that integrate into the clinical workflow. Frailty, as measured by the Revised Risk Analysis Index (RAI-rev), is a measure of baseline physiological reserve and is increasingly used for preoperative risk stratification.^{16–19,21,22} However, a frailty-based predictive model for PCVA in patients undergoing intracranial tumor resection (ITR) has not been previously reported.

Objectives

The present study sought to identify and quantify the effects of preoperative characteristics such as frailty that are associated with PCVA in patients undergoing ITR using patient data extracted from the ACS prospectively collected registry, the National Surgical Quality Improvement Program (NSQIP) (specifically for 2015–2019). We strived to develop and validate a clinically translatable predictive model for PCVA for use in this patient population.

Study Design

The present study was designed as a secondary analysis of a prospectively collected database. The paper was prepared with guidelines from the Equator Network’s Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement.

Data Source and Setting

Data used for this study were gathered from the ACS NSQIP database (2015–2019), a nationally validated database with contributions from more than 700 hospitals, producing more than 1 million deidentified patient cases annually. Patient case data abstracted from medical records included more than 200 variables, including demographics, diagnoses (International Classification of Diseases, Tenth Revision [ICD-10] codes), and procedures (Current Procedural Terminology [CPT]) codes). ACS-trained specialists review these data to ensure validity and organization.²³ The study was performed under our data user agreement of the ACS and considered exempt by the institutional review board.

Participants

CPT and ICD-10 codes were used to identify adult (≥ 18 years of age) ITR patient cases from the 2015–2019 NSQIP database.

Revised Risk Analysis Index

The RAI-rev, an adaption of RAI-A, is a quantitatively robust frailty index that is readily captured by clinical surveys.^24,25 This score is established by assigning numerical values to sex, age with or without cancer, weight loss, renal failure or dialysis, heart failure, poor appetite, shortness of breath, residence type, and functional status. In the present cohort, there were no missing data within the parameters used to compute the RAI-rev score. The score ranges from 0 (robust) to 78 (most frail). Categorization for this index has been established, with “robust” as RAI-rev scores 0–10, “prefrail” as RAI-rev scores 11–20, “frail” as RAI-rev scores 21–30, and “severely frail” as RAI-rev scores ≥ 31, which is used in the present study.²⁵

Variables

Demographic data including age, sex, race, ethnicity, and BMI were captured. Additional preoperative variables included individual preexisting comorbidities, frailty index (measured by the RAI-rev), preoperative laboratory values, and tumor type (benign, primary malignant, and secondary/metastatic malignant). The primary outcome variable was PCVA. The effect of PCVA on other standardized ACS NSQIP outcomes was considered. Secondary outcomes included unplanned reintubation, sepsis, septic shock, pneumonia, deep vein thrombosis (DVT) or thrombophlebitis, myocardial infarction (MI), superficial surgical site infection (SSI), deep SSI, organ space SSI, wound disruption, cardiac arrest, operative time, length of stay, Clavien-Dindo grade IV, unplanned reoperation, unplanned readmission, 30-day mortality, and discharge to a nonhome destination.

Statistical Analysis

Statistical analysis was performed using IBM SPSS (version 28.0.0.0, IBM Corp.). Statistical significance was determined via an a priori determined alpha of 0.05. Descriptive statistical analysis compared baseline characteristics between PCVA and non-PCVA cohorts. Categorical variables were analyzed with Pearson’s chi-square test with results stated as number (%). Continuous variables were analyzed using the Mann-Whitney U-test with results stated as median (IQR). Univariate and multivariable logistic regression analyses were used to identify preoperative factors associated with PCVA. Baseline clinical parameters with p < 0.1 and a minimum of 10 cases per cell in the univariate model progressed to the multivariable model. The goal of the multivariable model was to create a robust predictive model for PCVA after ITR. Variance inflation factor diagnostics were assessed to ensure no multicollinearity within the model. Discrimination capacity of the model was measured in area under the receiver operating characteristic curve (AUROC) analysis. The ITR-PCVA predictive model was designed as follows: log (Pi/1 − Pi) = β₀χ₀ + β₁χ₁ + β₂χ_{2 +} β_nχ_n, where Pi = predicted probability, χ = regression parameters, and beta (β) coefficient = effect size of regression parameters.

Participants

A total of 30,951 patients were queried from the ACS NSQIP (2015–2019) database who underwent ITR. The rate of PCVA was 1.7% (532/30,951).

Preoperative Demographic and Clinical Characteristics

The PCVA cohort was chronologically older, on average, than the non-PCVA cohort, with mean (SD) ages of 60 (14) years and 57 (15) years, respectively. The PCVA cohort had a greater proportion of nonelective surgery (46% vs 40%) and interhospital transfer (21% vs 18%) (both p < 0.05). Frailty, measured by the RAI-rev, was higher on average in the PCVA group than in the non-PCVA group (p < 0.001) (Table 1).

Outcome Data

Patients with PCVA had higher rates of unplanned reintubation (17.1% vs 1.7%), sepsis (5.5% vs 1.2%), septic shock (3.0% vs 0.4%), pneumonia (11.5% vs 1.8%), DVT/thrombophlebitis (8.1% vs 2.0%), pulmonary embolism (4.5% vs 1.3%), MI (0.9% vs 0.2%), deep incisional SSI (0.9% vs 0.4%), organ space SSI (2.4 vs 1.2), wound disruption (0.9 vs 0.2), Clavien-Dindo grade IV complications (31.2% vs 4.9%), unplanned readmission (28.5% vs 12.8%), unplanned reoperation (24.6% vs 5.0%), 30-day mortality (10.0% vs 0.9%), and discharge to nonhome destination (70.8% vs 21.3%) (all p < 0.05). PCVA was associated with prolonged operative time (median 236 vs 181 minutes) and length of stay (median 11 vs 4 days) (both p < 0.001) (Table 2).

ITR-PCVA Predictive Model

Univariate analysis screened the following possible risk factors for PCVA: chronological age, RAI-rev frailty, BMI, nonelective surgery, interhospital transfer, tumor type, coagulopathy, diabetes, hypertension, and abnormal preoperative laboratory values (Table 1).

Independent predictors of PCVA in this RAI-driven model were frail (vs robust) status (OR 1.7, 95% CI 1.2–2.4; p = 0.006), Black (vs White) race (OR 1.5, 95% CI 1.1–2.1; p = 0.009), nonelective surgery (OR 1.4, 95% CI 1.1–1.7; p = 0.003), preoperative elevated blood urea nitrogen (BUN) (OR 1.4, 95% CI 1.1–1.8; p = 0.014), diabetes (OR 1.5, 95% CI 1.1–1.9; p = 0.002), and hypertension (OR 1.4, 95% CI 1.1–1.7; p = 0.006) (Table 3). The ITR-PCVA RAI model achieved a discriminatory strength AUROC of 0.64 (95% CI 0.61–0.66) (Fig. 1). The predicted probability (p) of PCVA is estimated by the following equation: Logit (p) = −4.566 + 0.211 × (prefrail = 1 vs 0) + 0.502 × (frail = 1 vs 0) + 0.302 × (severely frail = 1 vs 0) + 0.434 × (Black race = 1 vs 0) + 0.052 × (other race = 1 vs 0) + 0.330 × (hypoalbuminemia = 1 vs 0) + 0.313 × (bleeding disorder = 1 vs 0) + 0.279 × (diabetes = 1 vs 0) + 0.342 × (hypertension = 1 vs 0) + 0.303 × (elevated BUN = 1 vs 0) − 0.182 × (primary malignant tumor = 1 vs 0) − 0.844 × (secondary malignant tumor = 1 vs 0).

ROC curve for the ITR-PCVA predictive model (RAI-rev based) with a concordance statistic of 0.64 (95% CI 0.61–0.66).

Key Results

To the best of our knowledge, this is the first study to propose a predictive model for PCVA in patients undergoing ITR. The ITR-PCVA risk model based on RAI-rev robustly predicts PCVA with a C-statistic of 0.64. The findings suggest that a risk model created using a combination of frailty score, demographics, comorbidities, type of tumor, and preoperative laboratory values can reliably predict PCVA that is associated with devastating consequences. Of note, factors influencing PCVA rates included frailty, race, nonelective surgery status, preoperative elevated BUN, diabetes, and hypertension.

Interpretation

Neurosurgical intervention is the mainstay treatment for patients with intracranial tumors causing symptoms and/or with high-risk features concerning for rapid disease progression. Particularly in patients with acceptable quality of life preoperatively and nonoperative alternatives available (e.g., radiation therapy), it is critical to anticipate and prevent serious peri- and postoperative complications. Stroke, both ischemic and hemorrhagic, is a known contributor to morbidity and mortality for patients undergoing brain tumor surgery.^{12–14,20,26,27} The incidence of PCVA after intracranial tumor surgery was 1.7% in our cohort of 30,951 cases, which is expectedly higher than rates after hemicolectomy (0.7%), hip replacement (0.2%), and lung resection (0.6%).⁴ A recent large validated ACS NSQIP study identified a 0.1% risk of perioperative stroke in a large cohort of patients who underwent noncardiac, nonmajor vascular, and nonneurological surgeries.⁵

PCVA is a distinct pathology associated with unique causes and risk factors compared with strokes in nonsurgical patients.²⁸ However, published data on this topic are sparse and limited to specific tumor and stroke types, reducing the utility of these data in general preoperative decision-making.^29–32 Previously discovered risk factors for perioperative stroke in neurosurgical patients include diabetes mellitus, ventilator dependence, previous neurological deficits such as hemiparesis and impaired sensorium, and history of stroke.^15,29 The present study adds to existing literature by identifying and quantifying independent risk factors for PCVA and proposing a clinically translatable risk model. Frailty was the most robust risk factor, which supports prior literature across the spectrum of neurosurgical patient populations.^{16,17,19,21,22,33} The current study provides increased data regarding specific aspects of a patient’s demographic.

Generalizability

The present study of 30,951 cases, derived from more than 700 hospitals in the United States, is widely generalizable from a socioeconomic, geographic, and hospital system level. The goal of the present study was to ultimately improve neurosurgical patient outcomes with improved preoperative risk assessment and provide a foundation for future research studies. While the presented predictive model provides a considerable foundation for clinical utility, we recognize the modest AUROC (0.64) of this predictive model, which is comparable to the AUROC generated by the original Spetzler-Martin AVM scoring system (0.66).³⁴ Nonetheless, the current model identifies several independent predictors of PCVA after ITR. We encourage continued reevaluation and prospective critique of the proposed predictive model as the total quantity of PCVA events increase with newer available data sets. Future research will need to address the mechanistic pathways between the predictor variables and outcome of PCVA.

Study Limitations

A nationwide database study using the ACS NSQIP allows for a large sample size and impactful statistical power as well as generalizability. However, there a number of limitations that warrant discussion. Studies of observational design are inherently limited by data quality, observer bias, and granularity of clinical specifics, among other limitations. However, the excellent data quality assurance measures taken by ACS NSQIP mitigate these limitations that are common to other nationwide databases while still providing a large sample size and multicenter perspective. The ACS NSQIP is curated from data entered prospectively by ACS-trained surgical clinical reviewers, which maximizes consistency, reliability, and granularity.

Patient outcomes are tracked for the first 30 days postoperation, and thus long-term outcomes were not reported. PCVA adhered to a standardized definition provided and tracked by the ACS, which promotes generalizability but may not always correlate with clinically relevant stroke.

Furthermore, selection criteria were based on billing codes such as CPT and ICD-10, which hold some inherent bias and inaccuracy because of categorization and institutional coding protocols. More specific patient case details, including family history and disease severity, are not included in the NSQIP and therefore were unable to be analyzed. Lastly, occurrence of PCVA among this study cohort was rare (1.7%), which may have hindered the discriminative power of the predictive model, but frailty and other preoperative factors were nonetheless identified as independent predictors of PCVA.