FIG 3.

Impact of feature engineering approach and machine learning algorithm on performance of machine learning models for polymyxin resistance prediction. Mean performance with 95% confidence intervals is shown across different performance metrics. The algorithms used were those that achieved the highest area under receiver-operator curve and can be found in Table 3. Histograms show how performance is impacted by the feature engineering approach (A) and the choice of machine learning algorithm (B). Abbreviations: AUROC, area under receiver-operator curve; bACC, balanced accuracy; CUIMC, Columbia University Irving Medical Center; GBTC, gradient boosted trees classifier; GWAS, genome-wide association study; SVC, support vector machine classifier.