Background: Hospitalized COIVD-19 patients are complex and heterogenous with regards to socio-demographics and comorbidities and many patients are at heightened risk for adverse clinical outcome. Early risk stratification enables clinical decision making for appropriateness and timeliness of interventions. Since the pandemic began, more than 100 models to forecast prognosis for hospitalized patients with COVID 19 have been developed, none have sufficient clinical utility to inform clinical decision making. Hence, we attempted to develop and validate a multivariable logistic regression model for predicting disease progression to organ failure and death in hospitalized patients with laboratory confirmed COVID-19.

Methods: Data from 3844 consecutive adults hospitalized with laboratory conformed COIVD-19 at Mayo Clinic in States of Arizona, Florida, and Minnesota from March 2020 to February 2021, and followed up 28 days from admission were obtained from a comprehensive review of electronic medical records. Risk prediction model was developed from 2881 patients (age 66.5 years, 58% males, and 83% whites) and validated on a randomly selected comparable 963 patients (mean age 66.9 years, 56% males, and 84% whites). Accompanying table showed baseline characteristic by model. To fit the model, we used a backward selection algorithm with a p value >=0.20 indicating exclusion and model fit was evaluated by Hosmer-Lemeshow test. Model calibration was assessed by calibration belt and diagnostic accuracy by area under the receiver operating characteristic curve (AUROC). Primary outcome was a composite of progression to organ failure (defined as need for invasive ventilation for respiratory failure or circulatory support for incident shock) or death at day 28 (whichever occurred first) from date of admission. Study is adherent to transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) guidelines and approved by Institutional review board approved the study (IRB ID 20-007620).

Results: 388 patients (10%) reached composite outcome at 28 days. The development model was developed using multivariable logistic regression with backward elimination, 15 baseline patient characteristics measured on the day of admission were retained in the development model included race (being nonwhite), social indicator (current cigarette smoking), comorbidities (heart disease, COPD, dementia, immunocompromised state, transplant recipient, liver disease), physiological variables (systolic blood pressure < 90, respiratory rate ≥30, temperature >38.0) laboratory parameters (blood urea nitrogen ≥30, c-reactive protein ≥75) and pharmacological intervention (dexamethasone) were independent predictors of composite outcome. Model demonstrated good discrimination in predicting 28-day composite event with area under the receiver operating characteristic curve (AUROC) of 0.7899 (95% Confidence Interval [CI] 0.75294 – 0.82019) in development model, AUROC 0.7606 (95% CI 0.70741 – 0.81369) in the internal validation cohort. The calibration plot and Hosmer-Lameshow goodness-of-fit test demonstrated good calibration of the model (Figure).

Conclusions: The prediction model developed on contemporary COVID-19 patient population with a range of comorbid conditions from 3 geographically dispersed states provide robust discrimination of patients at risk for progression to organ failure or death at day 28 after hospitalization for COVID-19. External validation is warranted.

IMAGE 1: Table. Baseline characteristics by development and validation cohorts

IMAGE 2: Figure.Receiver operating characteristics, development and validation models and calibration plot