Predicting future hospital antimicrobial resistance prevalence using machine learning.

Area of Science:

• Epidemiology
• Infectious Diseases
• Computational Biology

Background:

• Antimicrobial resistance (AMR) is a critical global health threat.
• Nationwide prediction of AMR at the hospital level can optimize intervention strategies.
• Machine learning approaches can leverage historical AMR and antimicrobial usage data for predictive modeling.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting future antimicrobial resistance (AMR) prevalence.
• To compare the predictive performance of Extreme Gradient Boosting (XGBoost) against traditional forecasting methods.
• To identify key factors influencing AMR predictions for improved interpretability.

Main Methods:

• Utilized historical antimicrobial use and AMR prevalence data for bloodstream infections in English hospitals (FY2016-2022).
• Trained and compared Extreme Gradient Boosting (XGBoost) models against baseline methods (previous value, difference, linear trend forecasting).
• Calculated XGBoost feature importances using SHAP values for model interpretability.

Main Results:

• XGBoost models demonstrated the highest predictive performance, outperforming other methods, particularly in hospitals with greater AMR fluctuations.
• Simple forecasting methods showed comparable performance when AMR prevalence exhibited minimal year-to-year changes.
• Feature importance analysis revealed that historical resistance patterns and complex interactions between pathogen resistance and antibiotic usage significantly inform predictions.

Conclusions:

• While year-to-year AMR changes are often small, XGBoost models enhance prediction accuracy in dynamic scenarios.
• Accurate AMR prediction facilitates informed decision-making, efficient resource allocation, and targeted public health interventions.
• The study highlights the potential of machine learning in combating the global AMR crisis.