A Bayesian Approach to Predicting Outcomes During the Initial COVID-19 Outbreak

ABSTRACT

Study Objectives: At the peak of the initial COVID-19 outbreak in Italy, providers were required to make decisions related to resource rationing due to a shortage of medical equipment. Identifying COVID-19 positive patients who were high-risk for severe illness early in their course could have assisted in determining the most appropriate medical management in many cases. Currently, few models exist to predict the outcome of COVID-19 positive patients. Among those that do, none to our knowledge utilize Bayesian logistic regression. The goal of this study was to generate a model that would dynamically estimate the probability of severe disease in patients who test positive for COVID-19 during their initial emergency department (ED) visit. Methods: This model initially utilized a Bayesian approach with prior data based on the literature at the time, and after one week employed logistical regression using retrospective data from our own patient set. In total, data from 428 RT-PCR-confirmed COVID-19 patients who presented between March 4th and May 7th of 2020 was incorporated. Priors included female sex, O2 Saturation, lymphocytes, LDH, and CRP. Data acquired during the patients’ encounter included co-morbidities, temperature, MAP, HR, ferritin, d-dimer, hs-troponin, platelets, total bilirubin, hgb, lactate, albumin, and SOFA score. Single imputation was utilized to address patients with missing data points. Our primary outcomes were vasopressor requirement, intubation, and death. Results: Utilizing these data points, a risk calculator for vasopressor requirement, intubation, and/or death was developed with a C-statistic of 0.85. See the supplementary materials for a comprehensive list of the regression coefficients, their betas, and standardized betas (Table 1) and a graph of our predicted primary outcomes compared to actual primary outcomes (Figure 1). Conclusion: A model predictive of vasopressor use, intubation, and death in COVID-19 positive patients was derived. By initially incorporating Bayesian logistic regression and prior data, this model could have theoretically been utilized in medical decision-making early in US outbreak the event that resource rationing had to be pursued at our institution. [Formula presented] [Formula presented]