A Novel Method for Handling Pre-Existing Conditions in Multivariate Prediction Model Development for COVID-19 Death

Abstract Many mathematical models have been proposed to predict death following COVID-19; all started with comorbidity subsets for this still-little understood disease. Thus, we derived a novel predicted probability of death model (PDeathDx) upon all diagnostic codes documented in the Department of Veterans Affairs. We present the conceptual underpinnings and analytic approach in estimating the independent contribution of pre-existing conditions. This is the largest study to-date following patients with COVID-19 to predict mortality. Cases were identified with at least one positive nucleic acid amplification test (NAAT). Starting in 1997, we use diagnoses from the first time a patient sought care until 14 days before a positive NAAT. We demonstrate the clear advantage of using an unrestricted set of pre-existing conditions to model COVID-19 mortality, as models using conventional comorbidity indices often assign little weight or usually do not include some of the highest-risk conditions; the same is true of conditions associated with COVID-19 severity. Our findings suggest that it is risky to pick comorbidities for analysis without a systematic review of all those experienced by the cohort. Unlike conventional approaches, our comprehensive methodology provides the flexibility that has been advocated for comorbidity indices since 1993; such an approach can be readily adapted for other diseases and outcomes. With our comorbidity risk adjustment approach outperforming conventional indices for predicting COVID-19 mortality, it shows promise for predicting outcomes for other conditions of interest.