Creating a prediction model for invasive candidiasis in the intensive care unit using a case control design: a...

Invasive candidiasis (IC) has a high attributable morbidity and mortality in patients in the intensive care unit (ICU). Current diagnostic tools lack sensitivity, introduce delay or have not been validated for regular use. As early treatment has proven vital for survival, multiple prediction models have been proposed but have not been validated for multinational implementation. In this study we propose to find factors predisposing the ICU patient to develop IC. We hope to develop an alternative prediction model using a large international dataset.

Using ICU-acquired IC as primary endpoint we retrieved retrospective information about 285 cases and 285 matched controls from the EUCANDICU database. Data about comorbidities, severity of illness and known risk factors for IC were available. We identified 31 independent risk factors using univariate analysis. A random subset of 80% of the observations were used to find the optimal prediction model. The selection of predictors was done using the LASSO technique, using λ = 1SE as regularization parameter. This choice for λ implies that a small amount of precision of the prediction is sacrificed to improve the external validity. The remaining 20% of cases were used to assess the predictive performance of the model.

Among other factors SAPS II score, SOFA score, past infection, renal impairment and the presence of multiple Candida colonization sites were all independently associated with an increased risk of developing IC. We incorporated 22 of 31 variables in a LASSO regression analysis which showed an AUROC of 0.7433.

Predicting which ICU patient will develop invasive candidiasis remains challenging, despite using an alternative methodology in a large multinational database. The performance of this prediction model is not good enough to be used in clinical practice.