Methods
We analyzed hospitalization records from January 1, 2003 through December 31, 2017 that were contained in the Nationwide Inpatient Sample (NIS). The NIS datasets constitute the largest all-payer, publicly available inpatient database in the US and are made available by the Healthcare Cost and Utilization Project (HCUP). Although the systematic sampling process used to select the hospitalizations to be included in the NIS has changed over time, the result is an approximate 20% sample of hospital discharges from participating states. The sampling strategy ensures that hospitalizations in the NIS are representative of the population on important factors including month of admission, primary reason for hospitalization, and hospital size, location, ownership, and teaching status. The NIS contained approximately seven million inpatient hospitalizations each year (35 million when weighted) from around 47 participating states.
Our study sample included hospitalizations among transition-aged patients within the age range of 16 to 24 years. Diagnoses and procedures are coded using International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnosis codes till the 3rd quarter of 2015, after which HCUP transitioned to ICD-10-CM format. To assess the study’s primary exposure, we first scanned the diagnosis codes (the principal diagnosis and up to 29 secondary diagnoses) in each patient’s discharge record for an indication of SCD (ICD-9-CM - 282.41, 282.42, 282.61, 282.62, 282.63, 282.64, 282.60, 282.68, 282.69; ICD-10-CM - D57.4x, D57.0x, D57.2x, D57.1x, D57.8x). These sickle cell-related encounters were then sub-divided into four mutually exclusive categories based on ICD-9-CM and ICD-10-CM codes: HbSS (ICD-9-CM - 282.61, 282.62; ICD-10-CM - D57.0x), HbSC (ICD-9-CM - 282.62, 282.63; ICD-10-CM - D57.2x), HbS/β-thalassemia (ICD-9-CM - 282.41, 282.42; ICD-10-CM - D57.4x), and ‘Other SCD’ (ICD-9-CM - 282.60, 282.68, 282.69; ICD-10-CM - D57.1x, D57.8x)
The covariates in this study included individual‐level sociodemographic and hospital characteristics. We categorized patients’ ages in groups of 16 to 18 years, 19 to 21 years and 22 to 24 years. Ethnicity was first stratified on the basis of reported ethnicity (Hispanic, non‐Hispanic); and the non‐Hispanic group was subdivided (White, Black, or other). The primary payer for the hospitalization was classified into Medicare, Medicaid, private, self‐pay and other (including under/uninsured). As a proxy for socioeconomic status, the Healthcare Cost and Utilization Project provides zip‐code‐level estimates of median household income, grouped into quartiles based on the patient’s residence. Hospital factors included census region (Northeast, Midwest, South, West), bed size (small, medium, large), and type (rural, urban‐nonteaching, urban‐teaching).
Descriptive statistics were utilized to represent the prevalence of SCD among all transition-aged hospitalizations and to demonstrate the prevalence of sub-types of SCD among all SCD hospitalizations for each year of the study period. We conducted bivariate analyses to assess the association between socio-demographic characteristics of the patients and SCD and its sub-types. Furthermore, to describe the trends in rates of in-hospital mortality among patients with SCD over the study period, we conducted joinpoint regression analysis, which is a statistical modeling approach specifically designed to evaluate and describe the extent to which the rate of a condition changes over time. The model first fits the annual rates of outcome of interest (i.e. in-hospital mortality) to a model with the minimum number of joinpoints (zero), suggesting that a straight line and single trend best fits the annual prevalence data. Then, more joinpoints are added iteratively to test the statistical significance of the various models using the Monte Carlo permutation method. Once the final (best-fitting) model with the optimal number of joinpoints has been selected, each joinpoint represents a statistically significant change in the trend, and each distinct trend is characterized using an annual percent change (APC) measure and its 95% confidence interval (CI). Lastly, we conducted adjusted survey logistic regression model to assess the association between various patient characteristics and in-hospital mortality among patients with and without a diagnosis of SCD. Analyses were performed using R (version 3∙6∙1) and RStudio (Version 1∙2∙5001) and the Joinpoint Regression Program, version 4.7.0.0 (National Cancer Institute); we assumed a 5% type I error rate for all hypothesis tests (two-sided). This study was deemed exempt by the IRB of Baylor College of Medicine as the study was performed on secondary de-identified data.