National Trends and Factors Associated With Inpatient Mortality in Adult Patients With Opioid Overdose

Brittany N. Burton, MHS, MAS; Timothy C. Lin, MAS; Engy T. Said, MD; Rodney A. Gabriel, MD, MAS


Anesth Analg. 2019;128(1):152-160. 

In This Article


Data Collection

Data were obtained from the publicly available dataset, the National (Nationwide) Inpatient Sample (NIS) database of the Healthcare Cost and Utilization Project (HCUP). HCUP was developed through a multiorganization partnership sponsored by the Agency for Healthcare Research and Quality. NIS is the largest multi-state all-payer inpatient health care database in the United States and approximates a 20% stratified sample of medical records from the US hospitals. NIS includes deidentified data of several variables (ie, demographic characteristics, clinical variables, payment source, and hospital factors). In 2012, NIS was redesigned to improve national estimates.[8] The new design of 2012 affected the numbers and types of discharges for prior years.[8] To address this, HCUP recommends users use "trend" discharge weights for NIS files before 2012 to minimize the effects of the redesign on estimated trends that cross 2012.[9] For years before 2012, we used the trend weight data element "TRENDWT" in place of the original discharge weight data element "DISCWT" to create national estimates for trends analysis that are consistent across 2012.[9] The NIS redesign deidentified hospitals and instead provided an NIS unique hospital identification number. As such NIS still retains hospital-level data elements for all hospitals. We used our analysis that was not affected as we used the "HOSP_NIS" and "HOSPID" data elements to identify hospitals. Finally, HCUP recommends users to use HOSP_NIS or HOSPID in mixed-effect models account for within-hospital correlation in regression models. NIS meets the criteria of the Health Insurance Portability and Accountability Act to protect personal information and therefore was exempt from the consent requirement by the University of California, San Diego Institutional Review Board. This retrospective analysis adheres to the STROBE checklist for cross-sectional studies.

Patient Population

Hospitalizations of patients ≥18 years of age from 2010 to 2014 with a diagnosis of POD or IOD were included. The primary objective was to explore national trends in opioid overdose (POD and IOD) from 2010 to 2014. The secondary objectives were to explore factors associated with inpatient mortality, defined as death during the inpatient stay, and to present an exploratory analysis of unadjusted rates among opioid overdose cohorts. We used the International Classification of Disease, Ninth Revision, Clinical Modification (ICD-9-CM) and External Cause of Injury codes to identify POD and IOD, as well as patient comorbidities and inpatient interventions. Here, POD was defined with the following ICD-9-CM codes: 9650.00 (poisoning with opium), 965.02 (poisoning by methadone), 965.09 (poisoning by other opiates and related narcotics), E850.1 (accidental poisoning by methadone), and E850.2 (accidental poisoning by other opiates and related narcotics). IOD was defined as 965.01 (poisoning by heroin) and E850.0 (accidental poisoning by heroin). We extracted medical history from patients with any diagnosis of IOD or POD; such extraction has been shown to be the most sensitive definition of opioid overdose.[10] Sociodemographic and hospital variables supplied in the NIS include race, sex, age, median household income, insurance status, hospital ownership, hospital location and teaching status, hospital region, and weekend hospital admission. Table 1 lists the ICD-9-CM codes used to identify medical history for each case.

Statistical Analysis

R, a software environment for statistical computing (R version 3.3.2; R Foundation for Statistical Computing, Vienna, Austria), was used to perform all statistical analysis. We weighted all patient records to produce national trends; we used data elements "TRENDT" before 2012 and "DISCWT" for 2012 and onward.[9,11] TRENDT and DISCWT were used to weight each patient record, and these weights are stored in each patient record. When the discharge weights are applied to the unweighted NIS data, the result is an estimate of the number of admissions/discharges for all inpatient discharges from community hospitals in the United States.[9,11] The stratification and weighting was used to provide national estimates for Table 2 and Table 3. We specifically used the "survey" package in R version 3.3.2.

For bivariate analyses, the P value for the comparisons between 2 cohorts (POD versus IOD) was derived from the Pearson χ 2 and Wilcoxon rank sum test for categorical (with continuity correction) and non-normally distributed continuous variables, respectively. Time trend analysis was performed using logistic regression for binary outcomes with year of hospital admission as the independent continuous variable. We also adjusted our time trend analysis for race, gender, age, median household income, and insurance status. The binary outcomes in the time trend analysis included the following: (1) IOD admission among all patients in NIS; (2) POD admission among all patients in NIS; (3) mortality among all patients admitted for IOD; and (4) mortality among all patients admitted for POD. To assess the association of potential risk factors with inpatient mortality, we first performed a mixed-effects univariable logistic regression followed by a mixed-effects multivariable logistic regression. All cases that had missing values for any of the variables were removed from the final analysis. Mixed-effect logistic regression analysis was done on complete cases. The random effect was "hospital identification number" (a unique value assigned in the NIS database for a specific institution). The mixed-effect logistic regression analysis was not weighted. Using this data element allowed us to account for clustered observations within hospitals.

In the initial mixed-effect multivariable logistic regression model, we included all covariates with P < .2 from the univariable analysis. We included the following 32 covariates in the univariate analysis: sex, race, age, median household income, source of payment, hospital location and teaching status, weekend hospital admission, type of opioid overdose (POD versus IOD), opioid abuse history (yes/no), respiratory complications (mechanical ventilation or tracheostomy), cardiopulmonary resuscitation, cardiac evaluation, hemodialysis, packed red blood cell transfusion, central venous catheter placement, any naloxone use, alcohol rehabilitation and therapy, drug rehabilitation and therapy, alcohol abuse, pneumonitis, metabolic acidosis, metabolic alkalosis, rhabdomyolysis, acute kidney injury, sepsis, pneumonia, hypotension, solid tumor malignancy, chronic pain, and cocaine, amphetamine, and hallucinogen poisoning. Backward elimination was then performed by stepwise removal of covariates with the largest P value until all covariates in the model were P < .05. The odds ratios (ORs), 95% confidence intervals (CIs), and Wald test P value were reported for each independent variable. We used a Bonferroni-corrected P value for 21 covariates (final model). Two-sided P < .002 was considered statistically significant. We assessed multicollinearity with variation inflation factor statistic, in which a value <5 was deemed adequate with no collinearity. We evaluated model discrimination with area under the receiver operating characteristic curve.

Based on the work by Peduzzi et al,[12] we estimated the minimum sample size needed for our study. Per Peduzzi et al,[11] p is the smallest of the proportions of negative or positive cases in the population and k is the number of covariates (the number of independent variables), then the minimum number of cases to include is N = 10 k/p. Given that the proportion of patients with inpatient mortality was 0.026 and we have 32 covariates in the final model, the minimum number of cases required is 12,308.