Predicting unfavorable tuberculosis outcomes using machine learning: a prospective cohort.

Area of Science:

• Biostatistics
• Computational Biology
• Infectious Disease Epidemiology

Background:

• Tuberculosis (TB) remains a leading infectious cause of death globally, with significant treatment failure rates.
• Conventional statistical methods struggle with complex, nonlinear predictor interactions in TB prognosis.
• Previous machine learning (ML) studies for TB prediction were limited by small sample sizes and lack of external validation.

Purpose of the Study:

• To develop and externally validate machine learning models for predicting unfavorable tuberculosis outcomes.
• To identify key clinical predictors of treatment failure, relapse, or death in drug-susceptible pulmonary TB patients.
• To compare the performance of ML models against traditional statistical approaches.

Main Methods:

• A multicenter prospective cohort study in Korea (2016-2024) with 1580 participants.
• Development of five ML models, including XGBoost, with optimized thresholds for prediction.
• Performance evaluation using AUROC and AUC-PR, with feature importance assessed by SHAP values.

Main Results:

• The XGBoost model achieved an AUROC of 0.818, identifying 73.3% of high-risk patients.
• Serum albumin, hemoglobin, lymphocyte count, and age were identified as significant predictors by SHAP analysis.
• Serum albumin and hemoglobin were independent protective factors in multivariate logistic regression.

Conclusions:

• Externally validated ML models, particularly XGBoost, demonstrate high accuracy and generalizability in predicting TB outcomes.
• Serum albumin and hemoglobin are primary predictors, with ML models also highlighting lymphocyte count's prognostic value.
• Nutritional and immune health indicators are crucial for predicting tuberculosis progression.