Bayesian Prediction of Severe Outcomes in the LabMarCS: Laboratory Markers of COVID-19 Severity - Bristol Cohort

We describe several regression models to predict severe outcomes in COVID-19 and challenges present in complex observational medical data. We demonstrate best practices for data curation, cross-validated statistical modelling, and variable selection emphasizing recent Bayesian methods. The study follows a retrospective observational cohort design using multicentre records across National Health Service (NHS) trusts in southwest England, UK. Participants included hospitalised adult patients positive for SARS-CoV 2 during March to October 2020, totalling 843 patients (mean age 71, 45% female, 32% died or needed ICU stay), split into training (n=590) and validation groups (n=253). Models were fit to predict severe outcomes (ICU admission or death within 28-days of admission to hospital for COVID-19, or a positive PCR result if already admitted) using demographic data and initial results from 30 biomarker tests collected within 3 days of admission or testing positive if already admitted. Cross-validation results showed standard logistic regression had an internal validation median AUC of 0.74 (95% Interval [0.62,0.83]), and external validation AUC of 0.68 [0.61, 0.71]; a Bayesian logistic regression (with horseshoe prior) internal AUC of 0.79 [0.71, 0.87], and external AUC of 0.70 [0.68, 0.71]. Variable selection performed using Bayesian predictive projection determined a four variable model using Age, Urea, Prothrombin time and Neutrophil-Lymphocyte ratio, with a median internal AUC of 0.79 [0.78, 0.80], and external AUC of 0.67 [0.65, 0.69]. We illustrate best-practices protocol for conventional and Bayesian prediction modelling on complex clinical data and reiterate the predictive value of previously identified biomarkers for COVID-19 severity assessment.