Machine Learning–Driven Drug Optimization for Typhoid Fever Based on Patient Profiles

Typhoid fever remains a major Global public health concern, with treatment outcomes dependent on antimicrobial resistance (AMR) and patient variability. Clinically determining the best medication for a certain patient can be difficult. Machine learning–based clinical decision support systems (CDSS) offer a promising avenue for improving diagnostic accuracy and guiding antibiotic selection using routinely collected clinical data We developed a multi-task predictive framework using XGBoost models to predict (i) treatment outcome/success (binary), (ii) treatment duration (days), (iii) a resistance-proxy score, and (iv) suspected typhoid using clinical data; we assessed discrimination (AUC), calibration (Brier), and R² where appropriate, and used SHAP for explainability to interpret feature importance, derive patient clusters, and generate individualized model explanations. We further implemented a counterfactual drug-simulation experiment to compare actual prescribed antibiotics with predicted best options for each patient The treatment outcome and suspected typhoid classification models achieved an AUROC of 0.9998 ± 0.0001 and 0.9834 ± 0.0007 respectively, accurately distinguishing between successful and failed therapies and typhoid and non-typhoid cases respectively. The treatment-duration and resistance-proxy regression models demonstrated strong explanatory capacity, accounting for 80% (R² = 0.8027 ± 0.0054) and 90% (R² = 0.9014 ± 0.0036) of the variance in therapy length and inferred resistance likelihood, respectively. SHAP analysis consistently highlighted the severity group, biochemical markers as dominant predictors across models and revealed biologically coherent patient subgroups. Counterfactual drug simulations suggested that alternative antibiotics could yield higher predicted success probabilities for a subset of patients, indicating the potential value of model-informed treatment ranking This study demonstrates the feasibility of machine learning for simulating drug choice in typhoid treatment using patient clinical profiles.