Use of an EHR to inform an administrative data algorithm to categorize inpatient COVID-19 severity

Abstract

Importance: Algorithms for classification of inpatient COVID-19 severity are necessary for confounding control in studies using real-world data (RWD). Objective: To explore use of electronic health record (EHR) data to inform an administrative data algorithm for classification of supplemental oxygen or noninvasive ventilation (O2/NIV) and invasive mechanical ventilation (IMV) to assess disease severity in hospitalized COVID-19 patients. Design: In this retrospective cohort study, we developed an initial procedure-based algorithm to identify O2/NIV, IMV, and NEITHER O2/NIV nor IMV in two inpatient RWD sources. We then expanded the algorithm to explore the impact of adding diagnoses indicative of clinical need for O2/NIV (hypoxia, hypoxemia) or IMV (acute respiratory distress syndrome) and O2-related patient vitals available in the EHR. Observed changes in severity categorization were used to augment the administrative algorithm. Setting: Optum de-identified COVID-19 EHR data and HealthVerity claims and chargemaster data (March - August 2020). Participants: Among patients hospitalized with COVID-19 in each RWD source, our motivating example selected dexamethasone (DEX+) initiators and a random selection of patients who were non-initiators of a corticosteroid of interest (CSI-) matched on date of DEX initiation, age, sex, baseline comorbidity score, days since admission, and COVID-19 severity level (NEITHER, O2/NIV, IMV) on treatment index. Main Outcome and Measures: Inpatient COVID-19 severity was defined using the algorithms developed to classify respiratory support requirements among hospitalized COVID-19 patients (NEITHER, O2/NIV, IMV). Measures were reported as the treatment-specific distributions of patients in each severity level, and as observed changes in severity categorization between the initial procedure-based and expanded algorithms. Results: In the administrative data cohort with 5,524 DEX+ and CSI- patient pairs matched using the initial procedure-based algorithm, 30% were categorized as O2/NIV, 5% as IMV, and 65% as NEITHER. Among patients assigned NEITHER via the initial algorithm, use of an expanded algorithm informed by the EHR-based algorithm shifted 54% DEX+ and 28% CSI- to O2/NIV, and 2% DEX+ and 1% CSI- to IMV. Among patients initially assigned O2/NIV, 7% DEX+ and 3% CSI- shifted to IMV. Conclusions and Relevance: Application of learnings from an EHR-based exploration to our administrative algorithm minimized treatment-differential misclassification of COVID-19 severity.