Supervised Learning Model Systems to Predict and Identify Drivers of AMR in Africa

Background The threat of antimicrobial resistance (AMR) is a critical and persistent challenge to global health and modern health care, especially in Africa. To address this challenge, we conducted a comparative analysis using statistical modelling to identify the predicting variables that impact AMR in Africa and identified the patterns surrounding AMR surveillance in the continent and leveraged existing AST/WGS data to develop models for predicting resistance gene profiles, which provide cost-effective means to gain genotypic insights and help prioritize isolates for WGS. Methods Patient metadata (n = 917,049) from the Pfizer’s Antimicrobial Testing Leadership and Surveillance (ATLAS) dataset were used in training two predictive algorithms. For the AMR gene prediction model, patient metadata was combined with antibiotic resistance profiles and corresponding minimum inhibitory concentration (MIC) values as input parameters. A Random Forest cluster model was trained to predict the status of specific genotypes. In parallel, a LightGBM model was developed to predict the resistance status of antibiotics to respective isolates, leveraging both metadata and gene profiles as input variables. Linear regression models were used to explore the factors influencing AMR across both categories, with MIC as the dependent variable. Pearson's chi-square tests examined the relationship between geographic region (Africa versus Others) and AMR status. Results The RF model achieved high accuracy and F1-scores for most beta-lactamase genotypes. This indicated effective classification across the categories. The LightGBM model achieved an accuracy of 81% in predicting overall antimicrobial susceptibility. Within African countries, the resistance status and country showed the highest significant association. The linear regression model revealed that 24.64% of the variance in MIC values in the continent is dependent and significantly impacted by the predictors selected. Conclusion Our models approach potentially enables prioritization of isolates to WGS in resource-limited settings and enables effective public health response to outbreaks.