An Explainability Study Associated with Fluid Creep Administration During the First 24 Hours of ICU Admission

Giulia Carpani1, Maximiliano Mollura1, Edoardo Maria Polo2, Stefano Finazzi3, Francesca Baroncelli4, Alessia Paglialonga5, Riccardo Barbieri1
1Politecnico di Milano, 2Sapienza University of Rome, 3Mario Negri Institute of Pharmacological Research, 4San Giovanni Bosco Hospital, 5CNR-IEIIT


Abstract

Intravenous fluid therapy is one of the most common in- terventions for hospitalised patients admitted to intensive care. In particular, patients may be subject to a progres- sive and dangerous accumulation of fluids. In this con- text, we can define Fluid Creep (FC) as those fluids used to dilute drugs and nutritions and to maintain catheter pa- tency. This single-center, retrospective study was carried out on the MargheritaTre database and included 4786 pa- tients with an average of 1606 ml (1 quartile 849-3 quar- tile 2000) of FC in the first 24 hours of intensive care unit admission. The objective of this analysis is to identify vari- ables significantly associated with FC, initially by means of a linear model and subsequently by means of a classifi- cation model aimed at identifying patients at risk of receiv- ing high FC using explainable artificial intelligence (AI) techniques. After comparing the performance of seven ma- chine learning models, logistic regression was found to be the model with the best accuracy on the test set of 0.76. Therefore, the SHAP (SHapley Additive exPlanations) al- gorithm was applied to conduct an explainable AI analy- sis, with the aim of interpreting the behaviour of the model and determining the most relevant variables in classifying the risk of high FC.