Scholarlink

Introduction <p>Typhoid fever remains a major Global public health concern, with treatment outcomes strongly influenced by antimicrobial resistance (AMR) and inter-patient variability. Determining the most appropriate antibiotic for an individual patient remains clinically challenging. Machine learning–based clinical decision support systems (CDSS) offer a promising avenue for improving diagnostic precision and guiding antibiotic selection using routinely collected clinical data.</p> Methods <p>We developed a machine learning-based decision-support framework using XGBoost models to predict (i) treatment outcome (binary), (ii) suspected typhoid classification, and (iii) a resistance-proxy score from clinical and engineered features. Model performance was evaluated using AUROC for classification tasks and R<sup>2</sup> for regression, alongside probability calibration analysis using the Brier score. SHAP was used to interpret feature importance, generate patient-level explanations, and identify latent patient subgroups. A counterfactual drug-simulation experiment was further implemented to compare clinician-prescribed antibiotics with model-recommended alternatives.</p> Results <p>The treatment outcome classifier demonstrated strong generalization performance, achieving a test AUROC of 0.962 ± 0.010 and an overall accuracy of 90%. The suspected typhoid classifier achieved an AUROC of 0.902 ± 0.005 with an overall classification accuracy of 82%. The resistance-proxy regression model showed moderate predictive capacity (R<sup>2</sup> = 0.588 ± 0.011). SHAP analysis identified platelet count, age, hemoglobin, calcium, potassium, and severity score as dominant predictors across models and revealed biologically coherent patient subgroups through attribution-based clustering. Counterfactual drug simulations showed that the model’s top recommendation matched the clinician-prescribed drug in 37.1% of cases and appeared as the second-rankedt option in 28.2% of cases. Treatment success was highest when prescriptions aligned with the model’s primary recommendation (72.7%) and lowest when no alignment was observed (32.6%).</p> Conclusion <p>This study demonstrates the feasibility of using machine learning to simulate antibiotic selection in typhoid treatment using patient-level clinical profiles. It presents a machine learning-based decision-support framework for antibiotic optimization under uncertainty, with explicit relevance to antimicrobial resistance management in resource-limited settings. To our knowledge, this is among the first studies to integrate explainable machine learning with counterfactual drug simulation for antibiotic optimization in typhoid fever.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

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

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月12日

查看全文
  • Charles Ssemuyiga,
  • Elminah Saru,
  • Yusuf Abbas Aleshinloye

摘要

Introduction

Typhoid fever remains a major Global public health concern, with treatment outcomes strongly influenced by antimicrobial resistance (AMR) and inter-patient variability. Determining the most appropriate antibiotic for an individual patient remains clinically challenging. Machine learning–based clinical decision support systems (CDSS) offer a promising avenue for improving diagnostic precision and guiding antibiotic selection using routinely collected clinical data.

Methods

We developed a machine learning-based decision-support framework using XGBoost models to predict (i) treatment outcome (binary), (ii) suspected typhoid classification, and (iii) a resistance-proxy score from clinical and engineered features. Model performance was evaluated using AUROC for classification tasks and R2 for regression, alongside probability calibration analysis using the Brier score. SHAP was used to interpret feature importance, generate patient-level explanations, and identify latent patient subgroups. A counterfactual drug-simulation experiment was further implemented to compare clinician-prescribed antibiotics with model-recommended alternatives.

Results

The treatment outcome classifier demonstrated strong generalization performance, achieving a test AUROC of 0.962 ± 0.010 and an overall accuracy of 90%. The suspected typhoid classifier achieved an AUROC of 0.902 ± 0.005 with an overall classification accuracy of 82%. The resistance-proxy regression model showed moderate predictive capacity (R2 = 0.588 ± 0.011). SHAP analysis identified platelet count, age, hemoglobin, calcium, potassium, and severity score as dominant predictors across models and revealed biologically coherent patient subgroups through attribution-based clustering. Counterfactual drug simulations showed that the model’s top recommendation matched the clinician-prescribed drug in 37.1% of cases and appeared as the second-rankedt option in 28.2% of cases. Treatment success was highest when prescriptions aligned with the model’s primary recommendation (72.7%) and lowest when no alignment was observed (32.6%).

Conclusion

This study demonstrates the feasibility of using machine learning to simulate antibiotic selection in typhoid treatment using patient-level clinical profiles. It presents a machine learning-based decision-support framework for antibiotic optimization under uncertainty, with explicit relevance to antimicrobial resistance management in resource-limited settings. To our knowledge, this is among the first studies to integrate explainable machine learning with counterfactual drug simulation for antibiotic optimization in typhoid fever.