Improving Interpretable Prediction Models for Antimicrobial Resistance

Abstract

One of the major problems of healthcare institutions is the treatment of infections caused by bacteria that are resistant to antimicrobials. The early prediction of such infections can improve the patient's evolution as well as minimise the spread of antimicrobial resistance. The creation of effective prediction models is particularly limited due to the high dimensionality of data, the imbalanced datasets and the concept drift problem. In this paper, we face these challenges from a machine learning perspective, considering the interpretability of the resulting models as essential. In particular, we present a study of multiple techniques focused on the mitigation of these problems, that are used in combination with interpretable models. Our results indicate that the use of oversampling along with sliding windows can improve the resulting AUC of models (up to reaching a mean AUC of 0.80 in our dataset), and FCBF can be used to drastically reduce the number of predictors, obtaining simpler models with a slight AUC reduction (from a mean number of predictors of 69.78 to 16.28, achieving a mean AUC of 0.76). According to our results, we show that the combination of multiple techniques for dealing with the aforementioned data-mining problems can clearly improve the performance of prediction models for antimicrobial resistance.