Medical informatics approaches for decision support in antimicrobial stewardship

Abstract

Health-care organisations are promoting a rational use of antimicrobials, also known as antimicrobial stewardship, with the aim of maximizing their clinical outcomes while limiting the rise in antimicrobial resistance. The University Hospital of Getafe, Spain participated in the development of the Wise Antimicrobial Stewardship Programme Support System (WASPSS) project: A Clinical Decision Support System (CDSS) focused on assisting the multidisciplinary teams which are responsible for antimicrobial stewardship in hospitals. The aim of this PhD thesis is to prove that production rules are a suitable approach by which to address the key challenges of antimicrobial stewardship from a Medical Informatics perspective. We confront the problems of: i) increasing the effectiveness of antimicrobial treatments, ii) facilitating the use of Clinical Practice Guidelines (CPGs) and iii) predicting infections caused by resistant microorganisms. We use WASPSS as the platform on which to implement and test our approaches. We first focus on improving the results of antimicrobial susceptibility tests (AST) to increase the effectiveness of antimicrobial treatments. We decided to use expert rules to infer new resistance patterns from those obtained in a laboratory. The main challenge is to model knowledge based on rules and defined over complex taxonomies. We use ontologies to translate those taxonomies into a multi-hierarchical definition of concepts and generate rules to link each term with its definitions. We tested our approach with AST results obtained over a year, obtaining 26.4% of new resistance patterns. Using this approach, we have implemented a new WASPSS module that extends the available AST results and alerts clinicians to possible microorganism misclassification or improper treatments. We then deal with the problem of integrating CPGs related to antimicrobial administration into CDSSs to facilitate their use in daily hospital practice. We propose the use of BPMN and DMN to model and visualise the complex processes and decisions included in these guidelines. Moreover, we use production rules derived from these models to estimate the adherence of a CPG to a specific patient. We put our approach into practice by modelling a guideline for vancomycin administration. As a result, we have implemented a new module for WASPSS which provides contextualised information concerning the current task. This approach also facilitates both guideline visualisation and task scheduling. Finally, we confront the clinical problem of predicting infections caused by Vancomycin-Resistant Enterococci (VRE). This kind of prediction models, also known as Clinical Prediction Rules (CPRs), must deal with several challenging problems, such as concept drift, imbalanced datasets and the high number of predictors. We combine different strategies to deal with these problems and develop a CPR with which to predict VRE infections. We obtained a final model with an AUC of 0.82, by combining a 30-month sliding window, oversampling, Fast Correlation Based Filter and LASSO. We then implemented a new WASPSS module that provides decision support by visualising the predicted outcome for a patient and by alerting physicians to patients at high risk of VRE infection. In conclusion, we have proved our hypothesis that production rules can be used in CDSSs for antimicrobial stewardship, being a basis on which to incorporate different kinds of knowledge. In addition, we have confronted their limitations. Ontologies along with rules can be used to incorporate knowledge based on complex taxonomies. The use of BPMN and DMN along with rules can improve the task of modelling and visualising processes and complex decisions. Finally, production rules can be used to incorporate the results of prediction models into CDSSs, while it is necessary to combine different datamining techniques to deal with the intrinsic problems of this scenario.