Variability in Palliative Care Use after Intracerebral Hemorrhage at US Hospitals: A Multilevel Analysis

Intracerebral hemorrhage (ICH) is a leading cause of mortality, with case-fatality rates as high as 50% [1, 2]. Palliative care (PC) is a multi-disciplinary approach to specialized medical and nursing care aimed at improving quality of life in the setting of life-threatening illnesses. Because of the high mortality rates after ICH, often within days or weeks after onset, early integration of PC into clinical care for ICH patients has been endorsed by major professional societies [3-5], and access to palliative and end-of-life care has been recognized as an important component of high-quality stroke care [4, 6].

Previous studies aiming to identify predictors of PC use after stroke and ICH have largely focused on differences between individuals by analysis of individual patient/discharge records alone [7-9]; however, hospitals caring for ICH patients differ in numerous geographical, structural, administrative, and local care culture characteristics. Between-hospital variability in PC use after ICH has not previously been investigated, and, at present, the variability in PC use after ICH in the US is unknown. In addition, it not known to what degree, if any, hospital characteristics contribute to explain between-hospital differences in PC use.

In the present study, we sought to understand the between-hospital variation in PC use after ICH among US hospitals. We employed a multilevel modeling approach to quantify between-hospital differences in PC use. Thereby, we aimed to (1) quantify between-hospital variability after ICH; (2) identify patient-level and hospital-level characteristics associated with PC use; and (3) determine whether between-hospital variation in PC use after ICH is driven by individual or hospital characteristics. Identifying hospitals underutilizing PC resources and determining barriers to PC use at the hospital level may inform future strategies aimed at improving universal access to PC resources for stroke patients.

Data were obtained from the Nationwide Inpatient Sample (NIS) [10]. The NIS is the largest all-payer inpatient database in the United States, representing a 20% stratified sample of all admissions to non-federal US hospitals. All diagnoses and procedures are recorded using International Classification of Diseases version 9 Clinical Modification (ICD9-CM) codes. This study was exempted from Institutional Review Board approval.

We identified adult cases with a primary diagnosis of non-traumatic ICH by using ICD9-CM code 431 between 2008 and 2011. We excluded cases with a secondary ICD9-CM code for arteriovenous malformation, traumatic brain injury, malignant brain tumor, and those undergoing aneurysm clipping and coiling to restrict our population to those with primary ICH. Only admissions to hospitals with at least 12 annual ICH cases were included. For hospitals that were sampled in multiple years, only the most recent sampling instance was included. The unit of observation in NIS is discharge after hospitalization. In order to prevent double counting of transferred patients, cases transferred to another hospital were excluded, while cases transferred from another hospital were included. This algorithm identifies ICH with high sensitivity and specificity [11]. Cases with missing information on age, sex, race/ethnicity, insurance status, and hospital characteristics were excluded (online suppl. Fig. 1; for all online suppl. material, see www.karger.com/doi/10.1159/000500276).

The outcome of interest was the use of PC resources as identified by ICD9-CM code V66.7. This code identifies documented use of PC measures [12], irrespective of the delivery mode (i.e., via a PC consultation service or integrated into routine clinical practice by the care team). The Coding Clinic first addressed the code V66.7 in 1996, and in 1998, the Coding Clinic [12] provided additional clarification regarding the use of code V66.7 (Vol. 15, No. 1, p. 11): “Terms such as comfort care, end-of-life care, and hospice care are all synonymous with PC and these, or similar terms, need to be written in the record to support the use of code V66.7. The physician should be queried if the treatment record seems to indicate that PC is being given, but the documentation is unclear. The care provided must be aimed only at relieving pain and discomfort for the PC code to be applicable” [12]. ICD9-CM code V66.7 has previously been shown to accurately identify care withdrawal in stroke patients with 81% sensitivity and 97% specificity [13].

Comorbidities were measured using a modified Charlson Comorbidity Index (CCI) [14, 15]. Case severity was determined using the All Patient Refined-Diagnosis Related Groups (APR-DRGs), derived from age, primary and secondary diagnoses, and procedures. The APR-DRG algorithm is a validated and reliable indicator of mortality, and is commonly used as a severity indicator in studies relating to stroke [16].

The following level 1 patient characteristics were included: age, sex, race, modified CCI, APR-DRG severity subclass, and insurance status. Hospital-level (level 2) covariates included hospital region (Northeast, Midwest, South, and West), hospital location (rural vs. urban), hospital teaching status, hospital bed size, hospital control (Government, not-for-profit private, and investor-owned private), hospital annual ICH case volume, hospital proportion of minority ICH patients, and hospital rates for prolonged intubation, gastrostomy, tracheostomy, and ventriculostomy per 100 ICH cases.

Baseline characteristics of individual (level 1) and hospital (level 2) variables were compared using Pearson chi-square test for categorical variables and Wilcoxon rank sum test for continuous variables, stratified by PC use in hospitals and individuals. To explore the between-hospital variability of PC use, we generated the frequency distribution of PC use among all included hospitals [17]. Multilevel multivariable logistic regression modeling was employed to (1) account for non-random clustering of individuals within hospitals; (2) characterize the between-hospital variability of PC after ICH; and (3) estimate whether differences in PC use between hospitals are due to individual-level (level 1) or hospital-level (level 2) characteristics. First, an intercept-only (null) model was fitted with a random intercept for hospitals, but without any other covariates (Model 0). This model serves as baseline model to assess the variability of PC use due to between-hospital variability and allows comparison to subsequent models containing potential explanatory covariates. Model 1 includes individual-level (level 1), and Model 2 includes hospital-level (level 2) variables; covariates were added as fixed effects variables to assess the residual variation between hospitals. The fully adjusted model includes both level 1 and level 2 covariates (Model 3).

Measures of between-hospital variation in PC use include the intraclass correlation coefficient (ICC), the proportional change in variance, and median OR (MOR). The ICC is the proportion of variance of the outcome attributable to between-hospital variability [18]. The proportional variance change is the percent change of between-hospital variance compared to the null model (Model 0) [19]. The MOR is defined as the median of a set of ORs that could be obtained by comparing 2 patients with identical patient-level characteristics from 2 randomly chosen hospitals [18-20]. The MOR can be interpreted as the median increase in odds of PC use when moving from a hospital with lower to a hospital with higher propensity for PC, adjusting for other covariates in the model.

Statistical analysis was performed using Stata version 15 (Stata Statistical Software: Release 15. College Station, TX, USA). A p value of <0.05 was considered statistically significant, and 95% CIs are reported.

A total of 26,791 admissions in 629 hospitals with at least 12 annual ICH cases met inclusion criteria (online suppl. Fig. 1). Among the 629 included hospitals, the median rate of PC use was 9.1 (interquartile range 1.5–19.3) per 100 ICH admissions. Figure 1 shows the rank order of PC use among all included hospitals with at least 12 annual ICH cases from lowest to highest. Of all hospitals, 150 (23.9%) hospitals had no recorded PC use, and an additional 199 (31.6%) hospitals had PC use rates below the study mean.

Compared to hospitals with PC use below the median, hospitals using PC at or above the median had a lower median percentage of ethnic minority patients (24.3 vs. 35.6%, p < 0.001; Table 1). Similarly, hospitals with higher use of PC had lower median rates of gastrostomy (7.1 vs. 9.1 per 100 cases, p < 0.001), tracheostomy (2.4 vs. 3.5 per 100 cases, p = 0.006), and prolonged intubation (7.4 vs. 8.3 per 100 cases, p = 0.010) when compared to hospitals with lower PC use, while median mortality rates were higher (25.4 vs. 24.0 per 100 cases, p = 0.006). Other hospital characteristics and procedure rates, stratified by hospital PC use, are shown in Table 1. Online supplementary Figure 2 shows a graphic representation of the association between hospital PC use and hospital-level utilization of common inpatient procedures and mortality after ICH (1 hospital with 100% PC use was excluded for this analysis).

A total of 11,866 ICH cases received care at hospitals with PC use below the median, while 14,925 patients were treated in hospitals at or above the median hospital PC use. ICH cases receiving care in hospitals at or above the median rate of hospital PC use were more likely to be -female (50.2 vs. 48.2%, p = 0.002), white (67.4 vs. 59.0%, p < 0.001), and had higher inpatient mortality (26.6 vs. 24.9%, p < 0.001) compared to those who were treated at hospitals with PC use below the median (Table 2). Other characteristics of patients, stratified by hospital PC use are presented in Table 2.

Among all patients, 12.5% received PC (95% CI 11.5–13.5). Patient characteristics, stratified by individual-level PC use, are presented in online supplementary Table 1.

Table 3 shows the results of multilevel modeling of PC use, starting with an empty random intercept model (without covariates; Model 0) allowing for variability by hospital. The ICC of Model 0 without covariates was 0.274 (Table 3), suggesting that 27.4% of the overall variability in PC use was due to between-hospital variability. After adjusting for the sampling year and the addition of level 1 (patient-level) covariates age, sex, race, insurance status, modified CCI, and -APR-DRG Severity, there was only a mild decrease in the ICC (0.265; Model 1). After adding only level 2 (hospital-level) variables (Model 2), the ICC decreased to 0.219, suggesting that hospital--level rather than individual-level covariates explain between-hospital variability in PC use. This model accounted for 25.8% of the between-hospital variance observed in the null model (Model 0), with 74.2% of between-hospital variance remaining unexplained.

In the fully adjusted model (Model 3), level 2 covariates associated with PC were hospital geographic region (OR 1.55, 95% CI 1.11–2.17 in the Midwest; OR 1.40, 95% CI 1.05–1.87 in the South; and OR 1.94, 95% CI 1.41–2.65 in the West, compared to Northeast region), hospital ethnic minority composition (OR 0.91, 95% CI 0.87–0.95 per 10% increase in the percentage of ethnic minority patients at a given hospital), and annual hospital ICH case volume (OR 1.17, 95% CI 1.02–1.35 per increase in 50 annual ICH cases; Model 3; Table 3). The median OR for the model including both individual-level and hospital-level variables was 2.63 (95% CI 2.41–2.89), indicating that if patients moved from a hospital to another hospital with a higher propensity of PC use, there was a 2.63-fold increase in the odds of receiving PC.

In the present study of US hospitals, we investigated between-hospital differences in PC use after ICH by using a multilevel modeling approach. We show that the percentage of ICH patients receiving PC varied substantially between hospitals, even after accounting for patient-level and hospital-level characteristics. This is consistent with previously reported wide variation in do-not-resuscitate orders after ICH among hospitals in California [21, 22], and substantial variability in prognosis and treatment recommendations among providers caring for ICH patients [23].

Although hospital-level characteristics explained just over 25% of the between-hospital variability in our null model, most of the between-hospital variability remained unexplained suggesting that other unmeasured variables may account for the remaining hospital-to-hospital variability. Uncertainty regarding prognosis and life expectancy may influence differences in timing and threshold for triggering PC services [21, 22, 24]. Similarly, knowledge, preferences, attitudes, and perceptions on quality of life as well as information on decision-making regarding life-sustaining therapies by providers and patients and their surrogates were not available [25-27]. However, hospital-level use of prolonged intubation, gastrostomy, tracheostomy, and ventriculostomy were not significantly associated with PC use in our study and did not explain between-hospital variation. Hospital ownership was accounted for in our models; however, other structural differences not captured in our data, including specialists on staff and resource availabilities, may drive hospital-level differences by determining the local practice patterns. Although individual-level variables were associated with PC use, they did not explain between-hospital variability in PC use, and the ICC decreased only when adding hospital-level covariates into the model.

Hospital minority racial composition of ICH patients was associated with decreased use of PC, even after accounting for individual-level patient race. Therefore, care location rather than individual patient race may be a particularly relevant barrier to racial equity in ICH-related care, since ethnic minority ICH patients tend to cluster at relatively few hospitals [28]. Future studies may determine whether differences in funding or staffing explain the observed lower use of PC at ethnic minority hospitals [29, 30]. Minority patients have higher use of other life-sustaining procedures after ICH, including gastrostomy [31-33]; however, hospital racial composition remained significantly associated with PC use after accounting for hospital-level life-sustaining procedures, suggesting that hospital racial composition determines the use of PC resources beyond a general increase in use of life-sustaining procedures.

The relatively low rate of discharge to hospice in our data is consistent with the notion that the hospital is still the predominant location of death for ICH patients, regardless of whether PC was utilized or not. The reasons for this potential underutilization of hospice resources cannot be derived from this study, but may include patient/surrogate preferences, provider prognostic uncertainty, and logistical challenges to access such resources in a timely manner. For example, it is possible that the rapidity in which death ensues after withdrawal of care in some cases outpaces the ability to arrange for discharge to hospice.

In addition to previously mentioned limitations, our study has other potential limitations inherent to administrative datasets [34], including the potential for miscoding. ICD9-CM code V66.7 identifies PC services use with high specificity [13]. However, there is a possibility of under-reporting or under-coding patients who actually received PC, especially among patients receiving PC for pain or symptom management in the absence of care withdrawal [35], or among patients who receive PC by the primary team as opposed to a specialty consultation team. Furthermore, NIS does not collect information on the trigger, timing, and mode of PC delivery. We acknowledge that data were collected from 2008 to 2011 when PC services and knowledge were relatively nascent. In addition, code V66.7 does not capture the quality or the impact of the actual care delivered. We attempted to mitigate the absence of clinical and physiological stroke data in NIS by adjusting all models for the CCI, a validated measure of patient comorbidities in ICH [14], as well as medical complications. Lastly, we were unable to investigate the contribution of implicit bias, provider attitudes, and individual patient preference to the observed differences in PC resource use.

Our data demonstrate a wide variability in PC use for ICH patients among US hospitals, and this variability is insufficiently explained by measured individual- or hospital-level characteristics. Further studies are needed to better understand the drivers of between-hospital variability in PC use, including the differences in attitudes and preferences by providers and patients alike as determinants of local practice patterns. Such efforts may include mandatory reporting of quality measures of PC as part of the stroke center certification and quality assurance process, similar to the reporting of other stroke-related measures that are captured in Get With The Guidelines. Transparency in reporting on PC measures by hospitals may better characterize the drivers of differences in local PC practice, and may identify gaps in care delivery. A standardized approach to implement PC services in the care of ICH patients is needed to ensure equity in care and adherence to recommendations in guidelines set forth by professional associations.

None.

This study was exempted from institutional review board approval.

Dr. Roland Faigle reports no conflicts of interest. Dr. Rebecca F. Gottesman is an Associate Editor for Neurology.

Dr. Roland Faigle is supported by a Career Development Award from the National Institute of Neurological Disorders and Stroke (K23NS101124). Dr. Rebecca F. Gottesman is supported by a grant from the National Institute on Aging (K24AG052573).

R.F. designed the study, analyzed the data, interpreted the data, and wrote the manuscript. R.F.G. interpreted the data, supervised the study, and edited the manuscript.