Predictors of Thirty-day Mortality and Length of Stay in Operative Subdural Hematomas

The rate of postoperative morbidity and mortality after subdural hematoma (SDH) evacuation is high. The aim of this study was to compare mortality statistics from a high-volume database to historical figures and determine the most significant preoperative predictors of mortality and length of stay (LOS). The National Surgical Quality Improvement Program registry was searched (2005-2016) for patients with operatively treated SDHs, of which 2709 were identified for univariate analysis. After exclusion for missing data, 2010 individuals were analyzed with multivariable logistic regression. Primary outcome was 30-day mortality. The average patient age was 68.8 ± 14.9 years, and 64.1% were males. Upon multivariate analysis, nine variables were found to be associated with increased mortality: platelet count < 135,000 (OR 2.04, 95% CI 1.39-2.99), INR >1.2 (OR 1.87, 95% CI 1.34-2.6), bleeding disorder (OR 1.80, 95% CI 1.32-2.46), need for dialysis within two weeks preoperatively (OR 5.69, 95% CI 3.15-10.27), ventilator dependence in the 48 hours preceding surgery (OR 3.99, 95% CI 2.82-5.63), disseminated cancer (OR 2.95, 95% CI 1.34-6.47), WBC count >10,000 (OR 1.55, 95% CI 1.15-2.08), totally dependent functional status (OR 1.84, 95% CI 1.2-2.8), and each increasing year of age (OR 1.04, 95% CI 1.031-1.05). It is not surprising that chronic conditions and functional status were associated with increased mortality. However, specific laboratory abnormalities were also associated with increased mortality at levels generally considered within normal limits. More studies are needed to determine if correcting lab abnormalities preoperatively can improve outcomes in patients with intrinsic coagulopathy.


Introduction
Subdural hematomas (SDHs) are one of the most common pathologies managed by neurosurgeons. Many patients with SDHs require surgical intervention [1]. However, the aging population and increased usage of antiplatelet and anticoagulant medications complicate the clinical decision making for surgical intervention. Therefore, prognostication and management of patients with SDHs remain pertinent to modern practice.
Historically, mortality among patients with traumatic acute SDHs treated surgically is high, with a range of 30%-70% reported [2][3][4][5][6][7][8][9]. More recent studies have demonstrated a 12%-15% mortality rate for SDHs overall [1,[10][11]. Many survivors also suffer from reduced functional capacity and disability even with prompt treatment. The dilemma for neurosurgeons is attempting to decide preoperatively who can achieve a "good outcome" with surgery. Current literature has established advanced age, low Glasgow Coma Scale (GCS) at admission, and SDH location as factors that influence mortality among SDH patients [2,7,8,[12][13][14]. Additional prognostic factors of mortality include midline shift (MLS), patency of cisterns, and hematoma volume [2,[15][16][17]. The goal in writing this paper was to compare current mortality statistics from a high volume database to historical figures, and to determine the most significant preoperative predictors of mortality.

Materials And Methods
We performed a retrospective review using the American College of Surgeons National Surgical Quality Improvement Program (ACS NSQIP) database. The ACS NSQIP maintains a prospectively collected database from over 600 participating sites and includes perioperative, intraoperative, and 30-day postoperative variables for a wide variety of surgical procedures [18]. An inter-rater reliability audit of participating sites is employed to ensure high quality data. The ACS NSQIP database has been previously discussed and validated in the surgical literature [19][20]. This study was conducted in adherence to the NSQIP participant data use agreement.
A 30-day mortality was used as the primary outcome. Hospital length of stay (LOS) was a secondary outcome. All variables related to demographics, medical comorbidities, perioperative factors and laboratory results that showed variability were extracted. Standard laboratory cutoffs were used to convert the values into categorical values. Time from admission to operating room was categorized into same day or after.

Statistical analysis
Statistical analysis was performed using MATLAB and Statistics Toolbox Release 2017b (The MathWorks, Inc., Natick, Massachusetts, USA) and SPSS (IBM Corp., Armonk, New York, USA).
For univariate analysis, Student's t-tests were used for continuous variables and reported with mean and standard deviation while Pearson χ2 tests were used for categorical variables and reported as number and percentage per group. A multivariate logistic regression model for predicting mortality was fit using all significant variables on univariate analysis. A separate multivariate logistic regression model was attempted for return to the operating room. For LOS, the values were log transformed to convert the distribution into a normal distribution. LOS was treated as a continuous variable and a multivariate linear regression was fit using forward stepwise addition based on the effect of the model deviance using variables with significant effects on univariate analysis.
To identify the optimal cutoffs for preadmission laboratory values including the international normalized ratio (INR) and platelet count for mortality prediction, separate multivariate logistic regression models were fit while iteratively changing the cutoff across the range of available values. The model was optimized based off of the information criteria. See Appendix A for more detailed statistical methodology.
The ACS NSQIP and the hospitals participating in the ACS NSQIP are the source of the data used herein; they have not verified and are not responsible for the statistical validity of the data analysis or the conclusions derived by the authors.

Patient demographics and comorbidities
A total of 2709 individuals who underwent surgical evacuation of SDHs between 2005 and 2016 were selected from the ACS NSQIP database. There were 699 patients excluded from the multivariate analysis due to absence of significant data, leaving 2010 available for the prediction models.
The average patient age (± SD) was 68.8 ± 14.9 years. A disproportionate number of men were affected, 64.1% of included patients, versus 35.9% women. Most patients (80.0%) were considered functionally independent prior to surgery. Among functionally impaired patients, 11.9% were considered partially dependent and 8.2% were considered totally dependent before surgical intervention. See Table 1 for a summary of patient demographics and Table 2 for a summary of mortality rates in patients with various comorbidities.   *Bleeding disorder defined as: patients with any chronic, persistent, active condition that places the patient at risk for excessive bleeding (e.g., vitamin K deficiency, hemophilia, thrombocytopenia, chronic anticoagulation therapy that has not been discontinued prior to surgery), and patients with active heparin-induced thrombocytopenia (HIT), and patients who has a past medical history of thrombocytopenia and a low platelet count at the time of the principal operative procedure. The following cases are not included: patient on chronic aspirin therapy; patient on nonsteroidal anti-inflammatory drugs (NSAIDs); When medications are prescribed for prophylactic use, for the principal operative procedure only; patient with a history of HIT in the past which is not deemed active.

Prediction of mortality
First, we defined the optimal cutoff for INR and platelet count that best dichotomized outcomes. Interestingly, an INR >1.2 and a platelet count <135 were the optimal values based on Bayesian information criterion (BIC) and Akaike information criterion (AIC). Next, we used multivariate analysis that found several significant factors related to mortality. Of note, the need for dialysis within two weeks prior to surgery showed the greatest increase in risk of mortality  Figure 1 caption), WBC >10, increasing age, and a totally dependent functional status were also associated with an increased mortality in the multivariate analysis. See Table 3 for a summary of the regression model. See Figure 1 for a comparison of some of the most significant predictors of mortality.

FIGURE 1: Predictors of mortality.
All of the above were included as dichotomous variables on multivariate analysis and found to be statistically significant predictors of mortality. Ventilator dependence was defined as the patient requiring ventilator-assisted respiration at any time during the 48 hours preceding surgery. Patients who were "totally dependent" required total assistance with all ADLs. "Bleeding disorder" includes patients with any chronic, persistent, active condition that places the patient at risk for excessive bleeding (e.g., vitamin K deficiency, hemophilia, thrombocytopenia, chronic anticoagulation therapy that has not been discontinued prior to surgery).
Platelet count <135,000 was strongly associated with mortality in both the univariate and multivariate models (OR = 2.04, 95% CI 1.39-2.99). Patients with a platelet count <135,000 had a 32.1% mortality, whereas patients with a platelet count greater than 135,000 had a mortality of only 14.4%. Patients with a platelet count of >135,000 were also shown to be associated with a decreased LOS (β = 0.03, p < 0.001). Figure 2 shows the association between platelet count and mortality.  confidence intervals.
Every year of increasing age was associated with a small but significant incremental increase in mortality as demonstrated by the positive beta coefficient (OR 1.043, CI 1.031-1.055, Table 3). See Figure 7 for a breakdown of mortality by age broken into decades.

FIGURE 7: Univariate analysis of mortality by age.
Error bars reflect standard error of the means.

Length of stay
The average LOS (± SD) was 10.4 ± 13.1 days among all patients. Patients discharged to skilled care, acute care, or rehab had an average LOS of 13.8 ± 11.2 days. Slightly less than half (48.8%) of the patients stayed one week or less. The longest hospitalization was 112 days, but greater than 90% of patients stayed less than three weeks.
A logistic regression model for LOS showed WBC, platelet count, INR, functional status, ventilator use, dialysis, operation time, sepsis, ASA class, days from admission until operation, and need to return to the operating room as significant predictors. See Table 4 for a summary.  Model evaluation: R-squared = 0.214, F-stat = 29.9, p = 1.23E-75.

Discussion
This review of the NSQIP database showed a similar overall mortality rate (16.5%) of patients with surgically treated SDHs compared to that found by Lukasiewicz in 2016 (18%) [21]. Similarly, increasing age, higher ASA class, and bleeding disorders were associated with higher mortality. Additional factors in the current study that were significantly correlated with mortality were INR>1.2, platelet count <135,000, dialysis dependence, ventilator dependence within 48 hours prior to surgery, disseminated cancer, totally dependent functional status, and WBC >10,000.
While this retrospective review utilized the same data source as Lukasiewicz et al., the database now contains laboratory information and clinical variables not previously reported. Additionally, sample size for the current study is over four times larger than was available for the previous study. This adds strength to the congruent findings and sheds light on clinically relevant variables that were not available for the prior study.
The prior study reported that "bleeding disorders" were associated with increased mortality. "Bleeding disorder" is a broad term that can apply to a number of different pathological or pharmacologically induced bleeding diatheses. However, the label is of limited clinical utility, from both prognostic and therapeutic standpoints. More specific parameters, i.e., platelet count and INR, can give clinicians more quantifiable information preoperatively. Both were shown to have an association with increased mortality. In contrast, Senft et al. have shown that patients on oral anticoagulants with SDHs who had their pharmacologic coagulopathy reversed in a timely manner did not differ significantly in terms of overall outcome when compared to historical cohorts [22]. One interpretation of this discrepancy could be that thrombocytopenia may be a marker of worse overall health that is not affected by giving platelets, as opposed to aspirin-induced platelet dysfunction that can be partially overcome by platelet transfusion. Likewise, an elevated INR can be the result of taking anticoagulants (i.e., coumadin), or as a surrogate marker for poor liver function. Patients' use of coumadin was not divulged in this dataset, prohibiting direct comparison between the groups.
Thrombocytopenia can be seen in patients with poor renal function [23]. However, as can be seen on the Kaplan-Meier curves in Figure 4, the mortality of patients who were dialysis dependent was greater than that of patients with isolated thrombocytopenia. In fact, dialysis dependence had the highest OR for mortality (OR 5.69, 95% CI 3.15-10.27) on multivariate analysis.
Additionally, ventilator dependence within the 48-hour period preoperatively was associated with an OR for mortality of 3.99 (CI 2.82-5.63). The reason for mechanical ventilation was not disclosed in the dataset. As patients with severe traumatic brain injury (TBI) are generally intubated for airway protection, it is possible that many of the ventilated patients also had more severe head injuries that contributed to higher mortality. However, GCS was not available for this dataset.
Disseminated cancer had an OR of 2.95 (CI 1.34-6.47) for mortality, supporting the evidence that the patient's general medical condition has a large impact on mortality.
Age was found to have a small but significant effect on mortality. Every increasing year of age was associated with an incremental increase in the risk of mortality (OR 1.04, 95% CI 1.031-1.05).
It has been established in the literature that SDHs have a high associated morbidity [2][3][4][5][6][7][8][9]. However, larger collections of retrospective data on this patient population are allowing us to determine variables that make patients with SDHs more likely to have a poor outcome. Carefully analyzing these data may aid clinicians in deciding which patients may benefit from operative intervention.
There are several potential limitations of this study. First, the data were collected retrospectively. Second, there was no indication of the severity of the TBI based off clinical exam. Also, anticoagulant and antiplatelet use were not reported. Additionally, for patients with an elevated INR, the use of reversal agents, the types of reversal agents used, and the timing of reversal for anticoagulation were not available. Finally, mechanism of injury was not available in this dataset. One would expect higher mortality in the subset of SDHs secondary to high speed motor vehicle accidents (MVAs) or high impact blunt trauma that often cause SDHs in the younger population, given the underlying brain injury. Alternatively, lower velocity trauma like ground level falls that are often responsible for SDHs in the older population often have little underlying parenchymal injury. However, even in a cohort of only high velocity trauma, it would be difficult to control for the differing degrees of brain injury.

Conclusions
This study showed that general medical conditions and preoperative functional status were associated with increased mortality and LOS in patients with surgically treated SDHs, as expected. However, specific laboratory abnormalities (platelet count <135,000, INR > 1.2) were also shown to contribute to increased mortality, and the quantitative values at which coagulopathy and thrombocytopenia were shown to be significantly associated with mortality were in a range that many hospital laboratories consider within normal limits. More studies are needed to determine if correcting lab abnormalities preoperatively can improve outcomes in patients with intrinsic coagulopathy (i.e., not due to an anticoagulant medication).