Predicting antimicrobial resistance with machine learning algorithms

• Applicants: Dr. Raphael Koch; Professor Dr. Frieder Schaumburg; Professor Dr. Julian Varghese
• Subject Area: Epidemiology and Medical Biometry/Statistics ; Medical Microbiology and Mycology, Hygiene, Molecular Infection Biology
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 544920804

The emergence and spread of antimicrobial resistance is a global concern as it is causing an increasing number of deaths (approx. 4.95 million in 2019). To curb this development, the use of antimicrobials must be reduced or more targeted. Currently, the rational use of antimicrobials is limited by slow diagnostic methods. Rapid antimicrobial susceptibility testing methods are therefore needed. The risk of infection with antimicrobial-resistant pathogens is not only determined by the pathogen itself (e.g. resistance genes), but also by the host/human (e.g. drugs, pre-existing diseases) and the environment (occupational exposure in farmers, nursing home residences). We therefore want to investigate whether holistic information from pathogen, patient, and environmental data can be used to accurately predict antimicrobial resistance using machine learning (ML) methods. This prediction would be available immediately after species identification, 18–24 h earlier than culture-based methods. Our project also differs from other work in that we aim to predict the exact minimum inhibitory concentration (MIC) rather than just resistance. The MIC is independent of the steadily changing MIC breakpoints at which a pathogen is considered resistant, so our proposed models could be used in a generalized fashion. We will create a project-specific retrospective database of patients treated for bacterial infection at Münster University Hospital (2021–2023). We will fit and compare different ML models for each combination of species and antimicrobial, including random forests, extreme gradient boosting, support vector machines, regularized linear regressions, and neural networks. Each model consists a pipeline including preprocessing, variable selection, etc., whose hyperparameters are optimized. The goal is to identify the best ML model and the most important influencing variables and to approximate it by a less complex (sparse) model that can be used in clinical practice. We will validate the generalizability of the predictive ability of the final ML models on a separate dataset from another geographic region in Germany. In a retrospective patient population, we will test whether predicting antimicrobial resistance using ML algorithms would have improved antimicrobial prescribing. For this purpose, the predicted MICs will be translated into the categories "susceptible", and "resistant" using EUCAST breakpoints. We will compare the ML-based prediction with the prescription of the medical staff. Using the proposed study, we can contribute to the extent to which ML algorithms can be integrated into microbiological diagnostics to improve antimicrobial prescribing by using pathogen, patient, and environmental data.

• DFG Programme: Research Grants