Arm Groups

Detailed Description

Patients with critical illness and/or respiratory failure tend to display better outcomes at hospitals that manage higher volumes of these patients. However, during a COVID-19 pandemic, this relationship may have been altered by the strain of higher patient volume on resources and personnel. A retrospective cohort study will be conducted using a large administrative database to understand the volume-outcome relationship among patients hospitalized with COVID-19. Understanding the COVID-19 volume-outcome relationship in U.S. hospitals may help identify the optimal care setting and transfer thresholds, and inform triage policies and risk-adjustment. The Premier Healthcare Database (PHD), an all-payer administrative data repository covering approximately 20% of U.S. hospitalizations from 48 states will be queried for adult (18+ years) in-patient encounters that recorded diagnosis coding for COVID-19 March 1st through August 31st, 2020 at continuously-reporting hospitals. Hospitals with <15 COVID-19 encounters over the study period will be excluded from the primary cohort to preserve statistical reliability. A surge index will be calculated for each hospital month that incorporates care-complexity-weighted COVID-19 case burden and will be normalized for nominal (2019) bed capacity and non-COVID-19 case burden and will represent the primary covariate. Hospital months will be stratified by the percentile of the surge index (e.g. <50th, 50-75th, 75-90th, 90-95th and ≥95th percentile). The impact of COVID-19-case burden (surge index) on the risk-adjusted odds ratio (aOR) of mortality (or discharge to hospice) among overall COVID-19 cohorts will be determined using a hierarchical generalized linear mixed model controlling for case-mix, treatment-related, time-varying and other hospital-level factors. Variables will be selected based on their representing "baseline" status upon presentation to the hospital to minimize effect of mediators. Variables for case mix adjustment: age, gender, race/ethnicity, comorbidity burden, insurance status, admission source and acuity, acute organ failure score and code status present-on-admission, use of pharmacologic agents (remdesivir, corticosteroids, hydroxychloroquine and azithromycin), severity of hypoxia (acute respiratory failure codes, need for non invasive positive pressure ventilation, need for invasive mechanical ventilation), vasopressor use within two days of admission respectively. Variables for temporal and hospital-level adjustment: hospital teaching status, urbanity, geographic region, bed capacity, hospital proportion of Medicaid/uninsured admissions, proportion of COVID-19 patients intubated within 2 days of admission, hospital's technologic index (stratified by availability of extracorporeal membrane oxygenation, more than one ICU and continuous renal replacement therapy), hospital availability of remdesivir, attending physician to patient ratio, pre-COVID-19 mechanical ventilation volume, proportion of overall admits tested for COVID-19 and admission month (1st vs 2nd wave). Effect modification by (1st vs. 2nd ) pandemic wave, the previous month's surge index and rate of acute hospital transfers and/or tracheotomies will assessed using interaction terms. Sensitivity analyses will be performed using alternative iterations/strata of the surge index, excluding fewer patients for statistical reliability, models with and without treatment-related variables and using Elixhauser comorbidity index instead of US Centers for Disease Control and Prevention-defined poor prognostic underlying conditions in COVID-19 to test robustness of results. Analyses will be repeated using in-hospital mortality (without reported discharges to hospice). The magnitude of an unmeasured confounder that would be necessary to alter the direction of statistically significant findings will be assessed. More selective subgroups of patients in ICU and mechanical ventilation will not be used as a primary study population given the potentially dynamic nature of admission and intubation criteria over the course of the pandemic and its relationship to volume. Depending on availability at the time of analysis, an de-identified curated electronic health record-based dataset will be used for clinical validation of select variables in administrative data (e.g. reliability of respiratory failure codes in representing patients requiring high-flow oxygen alone i.e. not non-invasive or invasive mechanical ventilation.)