« Back to Farmacia Journal 5/2026

MACHINE LEARNING-BASED PREDICTION OF ANTIBIOTIC SUSCEPTIBILITY TO SUPPORT PERSONALISED PHARMACOTHERAPY IN HOSPITALIZED PATIENTS: EVIDENCE FROM AN INDONESIAN REFERRAL HOSPITAL

IKHWAN YUDA KUSUMA ¹*, MUHAMMAD SYAIFUL ALIIM ², FIQIH NURKHOLIS ¹, MUH.AKBAR BAHAR ³

Download Full Article PDF

Antimicrobial resistance (AMR) is a major global health threat, particularly in resource-limited settings such as Indonesia, where empirical antibiotic selection is often unguided due to delayed diagnostics. This study developed and evaluated a multi-label, multi-class machine learning model to predict antibiotic susceptibility using routine clinical and microbiological data. A retrospective dataset of hospitalized patients in 2024 receiving systemic antibiotics classified under Anatomical Therapeutic Chemical (ATC) code J01 in an Indonesian secondary referral hospital was analyzed. Predictors included age, sex, International Classification of Diseases, Tenth Revision (ICD-10) diagnosis, hospital unit, specimen type, and microorganism. A multi-output Random Forest model was trained using an 80:20 stratified split and evaluated using accuracy, precision, recall, F1-score, subset accuracy, and area under the receiver operating characteristic curve (AUC-ROC). Across the evaluated antibiotics, model performance varied substantially when assessed using macro-averaged metrics and majority-class baselines. Meropenem showed high overall accuracy (0.94), but its macro-F1 score was low (0.32), and the accuracy lift above the no-information rate was minimal (0.005), indicating that the predominance of susceptible isolates largely drove accuracy. The summary AUC-ROC for meropenem was 0.70, whereas the resistant-class one-vs-rest AUC was 0.81. Therefore, model performance was interpreted primarily using macro-F1, class-wise metrics, AUC-ROC, no-information rate, and accuracy lift rather than accuracy alone. Misclassification mainly occurred between susceptible and resistant β-lactam classes, while intermediate categories showed lower AUC due to imbalance. Key predictors included age, microorganism, and diagnosis. The model shows potential to support empirical prescribing, although prospective validation is needed.