Comparing conventional and Bayesian workflows for clinical outcome prediction modelling with an exemplar cohort study of severe COVID-19 infection incorporating clinical biomarker test results

Table 3 Predictive performance of the individual biomarker models in Table 2 as described by the median area under the curve (AUC) in receiver operating curve (ROC) analysis and median difference between an Age and Gender reference model and the same model (negative values indicate the reference has worse performance) with the particular biomarker included (except univariate age and gender models)

^*Biomarkers not included in subsequent models due to small sample size, and recorded only in ICU (PCT)
Regressions were fit using all associated dummy variables for a given biomarker (e.g. mild, moderate, severe) and using only complete cases of training data (n=590), i.e. not using a variable for ‘Test not taken’. 95% inter-quantile ranges calculated via 5-fold cross-validation with 20 repeats (100 models total). Categorical variables use a reading of ‘Normal’ as a reference in the fitted model, except ‘Male’ used as the reference category for gender

ISSN: 1472-6947