# Predictive modelling of the dynamics of antimicrobial resistance: creation of a bank of renewable models based on machine learning

## TL;DR

This paper presents a machine learning-based system to predict antimicrobial resistance trends and optimize antibiotic use in Russia.

## Contribution

A renewable model bank using LightGBM and optimization methods to forecast AMR and guide antimicrobial stewardship.

## Key findings

- LightGBM achieved 66.6% validation precision for E. coli–cefotaxime AMR prediction.
- COBYLA optimization predicted 15–20% AMR reduction over 10 years.
- AMCmodel.ru platform enables real-time forecasting and decision-making.

## Abstract

The growth in antimicrobial resistance (AMR) presents a global threat, caused to a large extent by irrational antimicrobial consumption. For its part, mathematical modelling of the dynamics of AMR on the basis of data on antimicrobial consumption and historical levels of resistance may prove a promising tool for optimizing strategies to control this problem.

We used data on the consumption of systemic antimicrobials in the period 2008-22 for 82 regions of the Russian Federation and AMR levels in the period 2013-22. The data was processed with standardization of the regional names, the exclusion of antimicrobials with insignificant usage, the calculation of moving averages for AMR (with a window of 3-10 years) and antimicrobial consumption in Defined Daily Doses. To reduce dimensionality, principal component analysis was employed. On the basis of the “model pair” (microorganism-antibiotic) concept we tested machine learning algorithms: Light Gradient Boosting Machine (LightGBM), Random Forest, logistic regression, Support Vector Machines (SVMs) with linear and Gaussian kernels. We performed the calibration of the hyperparameters with cross-validation and assessed the metrics of precision and recall. We carried out predictions of AMR for optimized Constrained Optimization BY Linear Approximation (COBYLA method) and realistic Error, Trend, Seasonality (ETS model) usage scenarios.

For the model pair E. coli–cefotaxime, the LightGBM model presented the highest precision (67.5% for the training set, 66.6% for the validation set) without indications of overfitting. The key predictors of AMR would be the moving average of historical resistance and the type of infection. Optimization of consumption structure by the COBYLA method made it possible to predict a 15–20% reduction in AMR over 10 years, while a realistic scenario ETS foresaw a 5–10% growth in resistance.

The bank of models created on the basis of LightGBM provides for precise forecasting of the dynamics of AMR and the formation of strategies for the management of antimicrobial therapy. An optimization of consumption according to the results of the modelling is capable of reducing resistance by 15–20%. The AMCmodel.ru platform provides tools for real-time decision-making. An online platform AMCmodel.ru has been developed for data visualization, access to models and forecast generation.