Association of Do-Not-Resuscitate Patient Case Mix With Publicly Reported Risk-Standardized Hospital Mortality and Readmission Rates

This cross-sectional study examines the association between rates of mortality and readmission used by Medicare to determine financial penalties and to rank hospital quality with the case mix of patients with do-not-resuscitate status at US hospitals.

The inverse was true among the readmission cohorts, with the highest quintiles of DNR prevalence Meaning Results of this study suggest that risk-standardized hospital outcomes should control for do-notresuscitate status to avoid potential biases in pay-for-performance programs and hospital quality rankings.

+ Invited Commentary + Supplemental content Introduction
The Centers for Medicare and Medicaid Services (CMS) has invested more than $130 million to develop, refine, and disseminate risk-standardized quality outcome models. 1 These models are considered the criterion standard for risk standardization and have been adopted by professional societies in the US. 2,3 Currently, the condition-specific 30-day risk-standardized mortality rate (RSMR) model is used under value-based purchasing to calculate financial reimbursement for 3 conditions: acute myocardial infarction (AMI), heart failure (HF), and pneumonia. 4 Similarly, the condition-specific 30-day risk-standardized readmission rate (RSRR) model is used to establish Hospital Readmissions Reduction Program (HRRP) penalties, 5 which are levied annually against most US hospitals, for 4 conditions: AMI, HF, chronic obstructive pulmonary disease (COPD), and pneumonia. 6 In addition to the financial implications, 30-day RSMR and RSRR together account for 44% weight in the CMS Overall Hospital Quality Star Ratings. The 30-day RSMRs and RSRRs are adjusted for approximately 20 to 40 condition categories, which include common comorbidities such as cancer, diabetes, renal failure, and psychiatric disorders. However, a patient's do-notresuscitate (DNR) status is not included in risk adjustment despite the greater severity of illness and mortality risk among patients with DNR status present on admission (POA). [7][8][9][10] Evidence is limited on the association between the prevalence of POA DNR status and the riskstandardized outcome rates at hospitals. Although hospital-level prevalence of POA DNR status varies, the DNR status of patients is documented reliably enough for inclusion in risk-standardization models and analyses, suggesting that adjustment for patient-level DNR status improves the predictive accuracy of 30-day RSMRs for AMI, HF, and pneumonia. [11][12][13] To our knowledge, however, the association between hospital-level DNR prevalence and 30-day RSMRs and RSRRs has not been examined using national data from all Medicare-eligible US hospitals.
In this cross-sectional study, we assessed the variation in hospital-level POA DNR status across the condition-specific cohorts used in value-based purchasing reimbursement, HRRP penalty, and the CMS Overall Hospital Quality Star Ratings. Our aim was to identify the association between the prevalence of POA DNR status and condition-specific 30-day RSMRs and RSRRs. We also evaluated the implication of POA DNR status prevalence for the HRRP financial penalties.

Methods
This cross-sectional study was deemed exempt by the Mayo Clinic Institutional Review Board because the data used have been previously collected and deidentified. Informed consent was not required for this secondary research as contact with participants was not possible. We followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline.

Data Source and Study Population
We obtained data from the 2015 to 2018 CMS Limited Data Set Inpatient Standard Analytical File, which includes 100% of Medicare fee-for-service inpatient encounters across 4484 US hospitals.
The condition-specific cohorts for mortality were AMI, HF, stroke, pneumonia, and COPD. To form these condition-specific 30-day mortality cohorts, we applied International Classification of Diseases, Ninth Revision (ICD-9) codes (from July 1, 2015, to September 30, 2015 and International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10) codes (from October 1, 2015, to October 31, 2018) to the index admissions, adhering to the inclusion criteria specified in the CMS inpatient quality-reporting, condition-specific mortality documentation method, 14 and we excluded incoming acute care transfers and patients who were discharged against medical advice. In the AMI, pneumonia, and HF mortality cohorts, we also excluded patients who were discharged alive within 1 day of admission who were not transferred to an acute care facility. We did not exclude patients who arrived from hospice or who were discharged to hospice.
The condition-specific cohorts for readmission were AMI, HF, pneumonia, and COPD. To form these condition-specific 30-day readmission cohorts, we followed the CMS inpatient qualityreporting 30-day readmission measures method 15 and excluded all patients who died before discharge, were discharged against medical advice, or were discharged or transferred to an acute care facility. We did not exclude patients who arrived from hospice or were discharged to hospice. An important distinction is that CMS attributes 30-day readmissions to the discharging hospital,

Exposure and Outcomes
We defined DNR status as an ICD-9 diagnosis code of V49.86 (before October 1, 2015) or an ICD-10 diagnosis code of Z66 (beginning October 1, 2015) if the DNR order was POA per the corresponding POA indicator on the inpatient claim. Patients without a POA DNR status but who acquired DNR status after admission were included in the analysis. The outcomes were the hospital-level conditionspecific 30-day RSMRs for AMI, HF, COPD, pneumonia, and stroke and the 30-day RSRRs for AMI, HF, COPD, and pneumonia as reported on the Hospital Compare website in August 2019. Briefly, the CMS calculates hospital-specific RSMRs and RSRRs by taking each hospital's predicted vs expected (P/E) 30-day mortality or readmission rates from hierarchical mixed-effects risk-standardization models and multiplying these P/E ratios by the national mean 30-day mortality or readmission rate. We matched at the hospital level the values on the Hospital Compare website to the Inpatient Standard Analytical File data using the 6-digit CMS provider numbers. Hospitals without reported RSMR or RSRR data were excluded. diagnosis related group. We then used the Healthcare Cost and Utilization Project Comorbidity Index program to calculate the Elixhauser comorbidity index score from the weighted sum of these 29 comorbidity categories. 19,20 The Elixhauser comorbidity index score is an integer ranging from -32 (lowest) to 99 (highest) comorbidity burden.

Association Between DNR Status and Comorbidity Burden
We hypothesized that DNR status would be associated with a patient's comorbidity burden.
However, we also wanted to assess the shape of this association to ensure that an unexpected decline in DNR status did not occur at the sickest or healthiest patient extremes, which might suggest systematic bias in documentation. To test this hypothesis, we estimated a logistic regression model by pooling the patient-level data from the 5 condition-specific 30-day mortality cohorts, with DNR status as the dependent variable regressed on Elixhauser comorbidity index and age (both modeled using 5-knot restricted cubic spline functions) 21 and with a condition indicator as a third independent variable and an interaction term between the condition and the Elixhauser comorbidity index score.
We visually assessed the association between estimated probability of DNR prevalence (y-axis) and comorbidity burden (x-axis) (eFigure 1 in the Supplement).

Variation in Hospital-Level DNR Prevalence
From the logistic regression model, we calculated hospital-level observed vs expected DNR ratios and visually assessed the variation in both unadjusted condition-specific hospital-level DNR prevalence and the observed vs expected value using density plots. This visualization allowed us to consider the extent of interhospital differences in DNR documentation.

Association Between RSMR or RSRR and DNR Prevalence
Within each condition-specific cohort, we used the hospital-level RSMR or RSRR reported by the CMS as dependent variables in distinct inverse-variance-weighted linear regression models with the following independent variables: (1) quintiles of hospital-level DNR prevalence, with prevalence defined as the percentage of patients with POA DNR status; (2) mean hospital-level age modeled with a 5-knot restricted cubic spline function; and (3) mean hospital-level Elixhauser comorbidity index score modeled with a 5-knot restricted cubic spline function.
The inverse-variance weights were estimated from the CIs of the RSMR and RSRR in the Hospital Compare data, where SE = [(Upper CI -Lower CI) / (2 × 1.96)] and weight = 1 / SE. 2 The adjusted least-squared mean RSMR and RSRR (and 95% CIs) for each quintile of hospitallevel DNR prevalence were reported along with the overall Wald P value for DNR quintiles. We also compared the condition-specific mean RSMR and RSRR of the fifth (highest) quintiles with those of the first (lowest) quintiles using a Bonferroni-corrected P value. We conducted a sensitivity analysis using a more inclusive definition of DNR prevalence, in which patients were considered to have DNR status regardless of whether the DNR order was POA.

Association Between DNR Prevalence and HRRP Penalty
To assess the association between hospital-level POA DNR status and HRRP penalty, we pooled patient-level data from the 4 readmission cohorts included in the HRRP penalty calculation and assessed hospital-level POA DNR status. Next, we used logistic regression, adjusting for age and Elixhauser comorbidity index score (using restricted cubic spline functions), to identify the association between POA DNR prevalence (independent variable) and avoidance of HRRP penalty (dependent variable).

Statistical Analysis
All analyses were performed with SAS, version 9.4 (SAS Institute Inc). Two-sided P < .05 indicated statistical significance. Data were analyzed from July 2019 to March 2020. Condition-specific patient-level characteristics are presented in Table 1.
Across all condition-specific mortality cohorts, the Elixhauser comorbidity index score showed a statistically significant and generally nonlinear association with DNR status (eFigure 1 in the Supplement). Visual assessment indicated minimal potential systemic bias from lack of documentation of DNR status among patients at the extremes (sickest or healthiest).
Across all condition-specific 30-day mortality cohorts, variation was found in observed to expected DNR prevalence ratios (eFigure 2 in the Supplement), with a median (IQR) observed vs expected DNR prevalence ranging from 0.79 (0.38-1.36) in the HF readmission cohort to 0.94 (0.61-1.38) in the AMI mortality cohort (  , and an association between RSMR and DNR prevalence quintile was found across all conditions (all Wald P < .001) (Figure 1). In all condition-specific readmission cohorts, the RSRR was lower in the fifth DNR prevalence quintile than in the first quintile (AMI: 15  across all conditions (all Wald P < .001) (Figure 2). The sensitivity analysis in which we removed the POA requirement for patients with DNR status showed similar results (eTable in the Supplement).

JAMA Network Open | Statistics and Research Methods
Among the 2948 hospitals included in the analysis of the association between hospital-level POA DNR prevalence and HRRP penalty, 200 (6.8%) hospitals avoided an HRRP penalty from fiscal years 2016 to 2018. After adjustment for age and Elixhauser comorbidity index score, POA DNR prevalence was associated with HRRP penalty avoidance (odds ratio per 1% absolute increase in POA DNR prevalence, 1.06; 95% CI, 1.04-1.08; P < .001) (Figure 3).

Discussion
To our knowledge, this study is the first to use national Medicare claims data to assess the association between hospital-level prevalence of patients with POA DNR status and 30-day RSMRs and RSRRs across 5 conditions. A higher prevalence of patients with POA DNR status was associated with a higher RSMR. Conversely, a higher prevalence of patients with POA DNR status was associated with a lower RSRR. The 30-day RSMR and RSRR models do not adjust for POA DNR status, which may be a factor in unearned HRRP financial penalties and inaccurate CMS Overall Hospital Quality Star Ratings. 30-Day RSMRs were significantly higher for hospitals in the highest present-onadmission DNR prevalence quintiles vs the lowest quintiles (all P < .001). The points represent the Elixhauser Comorbidity Index-and age-adjusted 30 day least square mean for RSMR in each of the 5 DNR quartiles. The error bars represent the Wald 95% CIs. 1 Indicates the lowest quintile and 5 the highest quintile; AMI, acute myocardial infarction; COPD, chronic obstructive pulmonary disease; and HF, heart failure. The highest present-on-admission DNR prevalence quintiles had the lowest 30-day RSRR (all P < .001). The points represent the Elixhauser Comorbidity Index-and age-adjusted 30 day least square mean for RSRR in each of the 5 DNR quartiles. The error bars represent the Wald 95% CIs. 1 Indicates the lowest quintile and 5 the highest quintile; AMI indicates acute myocardial infarction; COPD, chronic obstructive pulmonary disease; and HF, heart failure.
The association between hospital DNR case mixes and quality outcomes has been examined previously, notably by Walkey and colleagues. 22,23 Specifically, Walkey and colleagues 22,23 found that hospitals with higher rates of DNR prevalence were more likely to be outliers for pneumonia mortality, and these authors hypothesized that the lack of adjustment for DNR status was associated with biased penalization of hospitals that in reality were providing high-quality, patient-centered care. 22,23 Results of the present analysis not only support this hypothesis and provide expanded evidence of this pattern across multiple conditions and quality outcomes linked with CMS reimbursements but also offer the first direct evidence of the implications of this association for HRRP penalties. In addition, these results are consistent with findings from 2 studies of approximately 300 hospitals in the California State Inpatient Database (CA-SID). The CA-SID study reported an increase in the C statistic from 0.75 to 0.78 when comparing a CMS-based riskstandardized AMI mortality model to an improved risk model that adjusted for DNR status 11 and a decrease in readmission rates for patients with pneumonia who had a DNR status. 12 In the AMI analysis, Bruckel and colleagues 11  The omission of DNR status as a risk-adjustment variable was attributed to concerns about its availability and documentation in claims data, 25 including whether DNR status acquired during hospitalization represented a preventable complication or natural history of disease. However, the same rationale exists for many of the clinical risk factors in the CMS models, and the variation may reflect true prevalence or accuracy of documentation. One example from the CMS models is proteincalorie malnutrition, a comorbidity for which reported hospital prevalence may reflect documentation rather than case mix. 26 Furthermore, no evidence is available to suggest that hospitals could take advantage of the system by overdocumenting DNR status. Abusing the system

Strengths and Limitations
This study has some strengths. It used a large national sample of CMS claims data from more than 4000 US hospitals. It also closely adhered to the CMS condition-specific cohort inclusion and exclusion criteria.
This study has some limitations. First, it had an ecological design, which makes it possible that patients with POA DNR status were not direct factors in the increased mortality rates at hospitals with high DNR prevalence. However, given that patients with a DNR status are known to have greater mortality risks, there is strong reason to suspect that these patients may be direct contributors to higher mortality rates. Second, some of the variation in DNR status may be associated with interinstitutional coding differences, physician preferences and comfort with discussing DNR status, and patient preferences and values. Given these differences, misclassification of DNR status may have occurred that could have biased the results to the null, particularly among several hospitals with 0% reported DNR prevalence. Third, although the hospital-level DNR prevalence rates in this study were similar to those in the CA-SID study, 11 which was validated with medical record reviews, we did not validate DNR documentation in this analysis. Furthermore, the inclusion of hospice patients may have introduced bias because these patients were excluded from the CMS metrics. Nevertheless, less than 2% of all encounters were hospice exclusions under the CMS method, so it is unlikely that this small percentage of encounters changed the results. 14 We did not assess the association between in-hospital DNR status acquisition and RSMR or RSRR, which could be an important area for future research given the increasing consideration of patient-centered palliative care in the hospital setting.
Fourth, socioeconomic and demographic factors were likely associated with patients' DNR status.
We did not adjust for these factors in these analyses, but they may help explain the associations between hospital DNR case mixes and RSMR or RSRR.

Conclusions
Under the current CMS risk-standardization models for 30-day mortality and readmission, mortality rates appeared to be worse among hospitals with greater proportions of patients with POA DNR status, whereas readmission rates appeared to be better among hospitals with greater proportions of patients with POA DNR status. Findings from this study suggest that controlling for DNR status would improve estimates for 30-day RSMRs and RSRRs and would have direct implications for HRRP financial penalty and hospital performance rankings.