An integrated monitoring system for antimicrobial consumption and resistance forecasting using dynamic models

Objective.

Development and validation of a suite of mathematical models to predict national-level antimicrobial resistance (AMR) trends by integrating antimicrobial consumption (AMC) and current AMR data.

Materials and Methods.

Data on systemic antimicrobial consumption across 82 regions of the Russian Federation (2008–2022; source: IQVIA) and AMR levels (2013–2022; source: AMRmap.ru) were analyzed. The workflow included regional name normalization, exclusion of antimicrobials with low clinical relevance, calculation of moving averages (3–10 years) for AMR parameters, and AMC quantification in Defined Daily Doses (DDD). Principal Component Analysis (PCA) was applied for feature dimensionality reduction. Within the «model pair» framework (microorganism–antibiotic), machine learning algorithms were tested: LightGBM, Random Forest, logistic regression, and Support Vector Machines (SVM) with linear and Gaussian kernels. Hyperparameter optimization was performed via k-fold cross-validation, with model performance assessed using accuracy and recall metrics. AMR forecasting was conducted under two scenarios: optimized (COBYLA) and realistic (ETS model).

Results.

For the «Escherichia coli / cefotaxime» pair, the LightGBM model achieved the highest accuracy (67.5% training, 66.6% validation) without overfitting. Key predictors included historical AMR moving averages and infection type (community-acquired/hospital-acquired). Optimized consumption patterns projected a 15–20% reduction in AMR over a decade, whereas the realistic scenario predicted a 5–10% increase, particularly among community isolates. An online platform (AMCmodel.ru) was developed for data visualization, model access, and forecast generation.

Conclusions.

The developed system enables accurate AMR trend forecasting and supports antimicrobial stewardship strategy design. Consumption optimization, as demonstrated by simulations, could reduce resistance by 15–20%. The AMCmodel.ru platform provides real-time decision-making tools. Further work will involve extending the suite of models to new microorganism-antibiotic pairs and incorporating international data to increase the predictive power of the models.