An explainable machine learning pipeline using XGBoost and SHAP to identify factors associated with tuberculosis treatment adherence in a district health center

Tuberculosis (TB) treatment adherence is critical for preventing disease relapse and the emergence of multidrug-resistant strains. Understanding factors associated with non-adherent behavioral profiles can guide targeted counseling interventions, particularly in resource-constrained public health settings. We conducted an exploratory cross-sectional study utilizing survey data from 379 TB outpatients at a district-level medical center. Following the 80/20 split, mode imputation parameters and Boruta feature selection were derived exclusively from the training partition, with training-derived mode values subsequently applied to missing observations in the hold-out test set. SMOTE was applied exclusively to the training data. The performance of an eXtreme Gradient Boosting (XGBoost) classifier was compared against a SMOTE-balanced Logistic Regression baseline on the independent hold-out test set. The proposed XGBoost + SMOTE model achieved an AUC-ROC of 0.768 (95% CI: 0.623-−0.896), a Sensitivity of 47.1% (95% CI: 23.1-−73.3%), and a Specificity of 89.8% (95% CI: 82.0-−96.7%). McNemar’s test ( $$p = 0.4545$$ ) indicated no statistically significant difference in classification performance compared to the SMOTE-balanced Logistic Regression. However, SHAP (SHapley Additive exPlanations) analysis provided deeper interpretable insights, revealing that patients’ comprehension of periodic testing schedules and core treatment principles were most strongly associated with adherence behaviors. By identifying non-linear interactions between sociodemographic factors and specific knowledge deficits, the explainable AI framework maps associative risk profiles. This empowers healthcare workers to precisely target educational gaps and prioritize behavioral interventions, serving as a foundation for future longitudinal predictive systems.