Abstract
In the intensive care unit (ICU), accurate dosing of antimicrobials is crucial to not only benefit the patient but also society by reducing the burden of antimicrobial resistance, considered a major global health threat. Piperacillin-tazobactam (TZP) and meropenem (MEM), two beta-lactam antimicrobials, are frequently prescribed in ICUs worldwide. Current dosing of these antimicrobials has been shown to not reach therapeutic plasma concentrations in many patients. Furthermore, it is believed that appropriate antimicrobial dosing decreases patient morbidity, decreases mortality, and avoids drug toxicity. Therapeutic drug monitoring guided dosing is advocated to optimize the usage of these antimicrobials. However, adoption in clinical practice is limited due to its time-consuming and costly nature. We, therefore, propose a hybrid machine-learning model that incorporates pharmacokinetics into a Catboost regressor model for personalized dosing of TZP and MEM in ICU patients. The resulting model is a single model for both antimicrobials that learns the patient’s drug clearance rate for determining their plasma concentrations and suggesting doses. The model was trained and temporally validated on two different datasets from different periods at the Ghent University Hospital ICU each with|without having access to a previous plasma concentration. The model reached R2 test scores of 0.869|0.807 on the internal dataset and 0.727|0.752 on the time validation dataset. This work established the feasibility of predicting plasma concentrations for an antibiotic group enabling personalized antimicrobial therapy in critically ill patients using machine learning. As such this work contributes to facilitating truly personalized antimicrobial therapy for critically ill patients.
Funding Information
  • Research Foundation Flanders (1S59522N)
  • FWO (G085920N)

This publication has 26 references indexed in Scilit: