Machine learning-driven decision support for antibiotic optimization in typhoid fever based on patient profiles

Area of Science:

• Computational biology
• Medical informatics
• Infectious disease modeling

Background:

• Typhoid fever poses a significant global health challenge, complicated by antimicrobial resistance (AMR) and patient variability.
• Selecting the optimal antibiotic for individual typhoid patients is clinically difficult.
• Machine learning (ML)-based clinical decision support systems (CDSS) can enhance diagnostic accuracy and antibiotic guidance using routine clinical data.

Purpose of the Study:

• To develop and evaluate an ML-based decision-support framework for typhoid treatment.
• To predict treatment outcomes, classify suspected typhoid cases, and estimate resistance-proxy scores.
• To simulate antibiotic selection and compare ML recommendations with clinician-prescribed treatments.

Main Methods:

• Utilized XGBoost models to predict treatment outcome, typhoid classification, and resistance-proxy scores.
• Employed AUROC and R-squared for model performance evaluation, with Brier score for probability calibration.
• Applied SHAP for feature importance analysis, patient-level explanations, and subgroup identification.
• Conducted counterfactual drug-simulation experiments to compare ML-guided and clinician-prescribed antibiotic choices.

Main Results:

• The treatment outcome classifier achieved high performance (AUROC 0.962 ± 0.010, accuracy 90%).
• The typhoid classifier demonstrated strong accuracy (AUROC 0.902 ± 0.005, accuracy 82%).
• SHAP analysis identified key predictors like platelet count, age, and hemoglobin; counterfactual simulations indicated treatment success was highest when aligned with model recommendations (72.7%).

Conclusions:

• ML models can effectively simulate antibiotic selection for typhoid fever based on patient clinical data.
• The developed framework supports antibiotic optimization under uncertainty, particularly relevant for AMR management in resource-limited settings.
• This study pioneers the integration of explainable ML with counterfactual drug simulation for typhoid treatment optimization.