Aims, Objectives and Background Clinical prediction models are often developed using composite outcomes, based on the implicit assumption that the predictors have similar associations with each component outcome. Using an example of a clinical prediction tool for adverse outcome in suspected COVID-19, we aimed to test this assumption and determine whether using a composite outcome led to suboptimal prediction of individual elements of the composite outcome.

Method and Design We reanalysed data from the Pandemic Respiratory Infection Emergency System Triage (PRIEST) study; data was collected from 20,891 patients attending 73 emergency departments with suspected COVID-19 and was used to develop a clinical score predicting a composite outcome of mortality or receipt of major organ support up to 30 days following attendance. In this reanalysis we created Least Absolute Shrinkage and Selection Operator (LASSO) multiple regression models to produce unrestricted prediction models for (1) the composite outcome, (2) mortality, and (3) receipt of major organ support.

Results and Conclusion Unrestricted regression models had c-statistics of 0.86 (95% Confidence Interval (CI) 0.85–0.86) for mortality, 0.78 (95% CI 0.77–0.80) for receipt of major organ support, and 0.82 (95% CI 0.82–0.83) for the composite outcome. Key variables in the clinical score (increased age, reduced performance status and reduced consciousness) predicted increased risk for mortality and the composite outcome but decreased or no significant risk for receipt of major organ support. The assumption that predictors have similar associations with individual elements of a composite outcome may not hold. Clinical prediction models may incur a ‘composite outcome fallacy’ if they are driven by predicting one element of the composite outcome but used to predict another. Further research into other clinical prediction score with composite outcomes is required.