Predicting the future distribution of antibiotic resistance using time series forecasting and geospatial modelling

Background: Increasing antibiotic resistance in a location may be mitigated by changes in treatment policy, or interventions to limit transmission of resistant bacteria. Therefore, accurate forecasting of the distribution of antibiotic resistance could be advantageous. Two previously published studies addressed this, but neither study compared alternative forecasting algorithms or considered spatial patterns of resistance spread. Methods: We analysed data describing the annual prevalence of antibiotic resistance per country in Europe from 2012 – 2016, and the quarterly prevalence of antibiotic resistance per clinical commissioning group in England from 2015 – 2018. We combined these with data on rates of possible covariates of resistance. These data were used to compare the previously published forecasting models, with other commonly used forecasting models, including one geospatial model. Covariates were incorporated into the geospatial model to assess their relationship with antibiotic resistance. Results: For the European data, which was recorded on a coarse spatiotemporal scale, a naive forecasting model was consistently the most accurate of any of the forecasting models tested. The geospatial model did not improve on this accuracy. However, it did provide some evidence that antibiotic consumption can partially explain the distribution of resistance. The English data were aggregated at a finer scale, and expected-trend-seasonal (ETS) forecasts were the most accurate. The geospatial model did not significantly improve upon the median accuracy of the ETS model, but it appeared to be less sensitive to noise in the data, and provided evidence that rates of antibiotic prescription and bacteraemia are correlated with resistance. Conclusion: Annual, national-level surveillance data appears to be insufficient for fitting accurate antibiotic resistance forecasting models, but there is evidence that data collected at a finer spatiotemporal scale could be used to improve forecast accuracy. Additionally, incorporating antibiotic prescription or consumption data into the model could improve the predictive accuracy.