Explainable machine learning for predicting clinical outcomes in HIV/TB co-infection: a comparative retrospective...

Area of Science:

• Medical Informatics
• Public Health
• Machine Learning in Medicine

Background:

• HIV/TB co-infection significantly increases mortality and treatment failure compared to tuberculosis alone.
• Machine learning (ML) offers advanced methods for early identification of high-risk patients.
• Early risk stratification is crucial for improving patient outcomes in co-infected populations.

Purpose of the Study:

• To develop and validate a machine learning model for predicting outcomes in patients with HIV/TB co-infection.
• To identify key clinical and immunological predictors of unfavorable outcomes.
• To assess the potential of ML tools in clinical practice for resource allocation and treatment optimization.

Main Methods:

• Retrospective analysis of 359 HIV/TB co-infected patients.
• Utilized six ML classifiers (Random Forest, XGBoost, LightGBM, SVM, Extra Trees, CatBoost) with SMOTE for class imbalance.
• Model performance evaluated using AUC, accuracy, precision, recall, F1-score, and ranked with TOPSIS; leading model interpreted with SHAP.

Main Results:

• The LightGBM model achieved the highest performance (AUC=0.771, accuracy=84.72%, F1=0.522).
• SHAP analysis identified age, CD4/CD8 counts, BMI, and occupation as significant predictors.
• Lower BMI, severe immunosuppression, and older age were associated with unfavorable outcomes.

Conclusions:

• A gradient-boosted model (LightGBM) with SHAP interpretation reliably predicts outcomes in HIV/TB co-infection.
• The model highlights clinically actionable risk factors, enabling earlier identification of high-risk patients.
• Integration into clinical workflows can enhance resource allocation and improve TB treatment success.