Predicting multi-resistance of bacteria in an Intensive Care Unit

Hernàndez Carnerero, Àlvar

Predicting multi-resistance of bacteria in an Intensive Care Unit

License

Citation

Hernàndez Carnerero, À. Predicting multi-resistance of bacteria in an Intensive Care Unit. 2020. handle: http://hdl.handle.net/10230/46363

This citation was generated automatically.

Abstract

This study considers the prediction of “multi-drug” resistance (MDR) of Pseudomonas aeruginosa bacterium caused by nosocomial infections in the Intensive Care Unit (ICU). An ensemble of binary classifiers implemented with different Machine Learning (ML) methods is applied for prediction using as training data health records and past sensitivity tests (antibiogram) results. This work proposes to generate two new types of features to improve predictor’s performance. The first one is based on using information of previous antibiograms of a particular patient to predict their future resistance to antibiotics. The second kind of features employs bacterial information from the rest of the patients in the ICU to predict the antimicrobial resistance for a certain patient. In addition, in the study it is suggested to use a training window with incremental size so that training set is always temporarily as near as possible to the test instances to be predicted. Some techniques such as feature selection and oversampling are also used to further improve efficiency and accuracy. Results show that using an incremental window for training improves success rates in the domain of this problem, and expose that knowing the outcomes of past antibiograms, substantially improves prediction. It is also observed that considering resistant bacteria present in the ICU is useful to anticipate antimicrobial resistance. From these results it is further inferred that resistant bacteria may be spreading among patients in the ICU within populations that rapidly mutate, which can induce non-stationary in the data distribution. It is concluded that using these contributions, experiments show promising results in MDR prediction even using simple features and limited training data.