ObjectivesTo develop cross-validated prediction models for severe outcomes in COVID-19 using blood biomarker and demographic data; Demonstrate best practices for clinical data curation and statistical modelling decisions, with an emphasis on Bayesian methods. DesignRetrospective observational cohort study. SettingMulticentre across National Health Service (NHS) trusts in Southwest region, England, UK. ParticipantsHospitalised adult patients with a positive SARS-CoV 2 by PCR during the first wave (March - October 2020). 843 COVID-19 patients (mean age 71, 45% female, 32% died or needed ICU stay) split into training (n=590) and validation groups (n=253) along with observations on demographics, co-infections, and 30 laboratory blood biomarkers. Primary outcome measuresICU admission or death within 28-days of admission to hospital for COVID-19 or a positive PCR result if already admitted. ResultsPredictive regression models were fit to predict primary outcomes using demographic data and initial results from biomarker tests collected within 3 days of admission or testing positive if already admitted. Using all variables, a standard logistic regression yielded an internal validation median AUC of 0.7 (95% Interval [0.64,0.81]), and an external validation AUC of 0.67 [0.61, 0.71], a Bayesian logistic regression using a horseshoe prior yielded an internal validation median AUC of 0.78 [0.71, 0.85], and an external validation median 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 AUC of 0.74 [0.67, 0.82], and external validation AUC of 0.70 [0.69, 0.71]. ConclusionsOur study reiterates the predictive value of previously identified biomarkers for COVID-19 severity assessment. Given the small data set, the full and reduced models have decent performance, but would require improved external validation for clinical application. The study highlights a variety of challenges present in complex medical data sets while maintaining best statistical practices with an emphasis on showcasing recent Bayesian methods.