Retrospective Analysis of Tuberculosis Treatment Duration and Outcomes Using Explainable Machine Learning on Malaysian Patient Data

Background/Objectives: This study utilizes machine learning (ML) models integrated with explainable artificial intelligence (XAI) to perform a retrospective analysis of factors associated with tuberculosis (TB) treatment duration and outcomes. Methods: Two datasets comprising completed TB patient records from a Malaysian hospital were analyzed. Significant features were identified using Pearson correlation for treatment duration and the Boruta algorithm for treatment outcome. Support Vector Regressor (SVR) was used to evaluate treatment duration, while Logistic Regression, Random Forest, and XGBoost were evaluated for treatment outcome classification. SHapley Additive Explanations (SHAP) was integrated to enhance model interpretability. Results: On the primary dataset (Dataset A), SVRPearson demonstrated the best performance for treatment duration (MAE = 44.18 days, R2 = 0.54), although performance decreased on the external dataset (Dataset B: MAE = 71.33 days, R2 = 0.11), partly due to missing feature imputations. For treatment outcome classification, XGBoostBoruta achieved an accuracy of 99.20% on Dataset A and 97.34% on Dataset B. SHAP feature attribution indicated that treatment outcome was the primary contributor to treatment duration, whereas treatment status and treatment duration were the dominant features contributing to the treatment outcome model. Conclusions: These findings demonstrate the value of explainable machine learning for characterizing factors associated with TB treatment duration and outcomes in completed treatment episodes and supporting retrospective clinical audit and post hoc case interpretation.