Machine Learning Classification of High Antibiotic Consumption Using the Drug Resistance Index: A Comparison of Random Forest, Support Vector Machine and Logistic Regression in Indian Tertiary-Care Hospitals

Abstract

Background: Hospitals record antibiotic consumption and antimicrobial resistance but rarely
analyse them together. The Drug Resistance Index (DRI) combines both by weighting resistance
to each antibiotic by that drug’s share of total use. This study compared three machine learning
classifiers for identifying high antibiotic consumption from the DRI and routine surveillance
variables.
Methods: Antibiotic consumption records with corresponding DRI values from tertiary-care
hospitals in India were classified as high or low consumption. Random forest (RF), support vector
machine (SVM) and logistic regression (LR) models were trained on five predictors: hospital,
year, antibiotic, antibiotic class and DRI. Performance was evaluated on a hold-out test set of
856 records (401 high and 455 low consumption) using accuracy, precision, recall, F1-score and
confusion matrices. The contribution of each predictor was estimated using RF impurity-based
feature importance.
Results: RF correctly classified 846 of the 856 test records (accuracy 98.8%; macro-averaged
F1-score 0.99). SVM and LR achieved accuracies of 80% and 57%, with macro-averaged F1-
scores of 0.80 and 0.57, respectively. In the RF model, antibiotic (importance 0.33) and antibiotic
class (0.31) were the strongest predictors, followed by hospital (0.18), DRI (0.12) and year (0.06).
The median DRI was 0.51 in both the high- and low-consumption groups.
Conclusions: RF distinguished high from low antibiotic consumption considerably better than
SVM and LR, but it relied mainly on the antibiotic and its class rather than on the DRI. Grouped
and external validation are needed before such models are used to guide antimicrobial
stewardship.