Answers are grounded in this paper and cite their sources.
Background/Objectives: Unsuccessful tuberculosis (TB) treatment outcomes impede disease control. This study identified associated factors and evaluated machine-learning models using Malaysian national registry data. Methods: A retrospective cohort study used Malaysian National Tuberculosis Information System records registered from 2014 to 2021 with an evaluated treatment outcome (n = 193,862); registrations from 2022–2023 were excluded because follow-up was incomplete. Missing predictor values were retained as a separate ‘Not recorded’ category. Multivariable logistic regression with 17 prespecified predictors quantified adjusted associations. Logistic regression, decision tree, k-nearest neighbours, random forest, gradient boosting, XGBoost, LightGBM and a stacked ensemble were developed on 2014–2019 registrations (n = 149,322) using patient-grouped cross-validation tuned for the area under the precision–recall curve (PR-AUC) and evaluated on a 2020–2021 temporal test set (n = 44,540) with unsuccessful outcome as the positive class, using ROC-AUC, PR-AUC, calibration, decision curves and classification metrics at an exploratory threshold of 0.20. SHAP was used for explainability. Results: Unsuccessful outcomes occurred in 39,172 episodes (20.2%), of which 55.6% were deaths. The strongest adjusted associations were observed for TB meningitis (aOR = 5.15 (95% CI: 4.72–5.61)), having no DOT supervisor (aOR = 4.72 (95% CI: 4.19–5.30)) and HIV-positive status (aOR = 3.36 (95% CI: 2.97–3.81)). The higher odds among non-Malaysian patients (aOR = 2.15) depended on classifying transfers out as unsuccessful. On the test set, ROC-AUC was 0.741 (95% CI: 0.735–0.746) for conventional logistic regression and 0.753 (95% CI: 0.747–0.758) for gradient boosting, the primary model; XGBoost, LightGBM and the stacked ensemble were within 0.001 of gradient boosting. At the 0.20 threshold, gradient boosting had a sensitivity of 0.696, specificity of 0.675, positive predictive value of 0.365 and F1 score of 0.479. Calibration was good (slope 0.997 for logistic regression and 0.959 for gradient boosting). SHAP identified ethnicity/citizenship, HIV status, age, sex and chest radiography as the largest contributors to model predictions. Conclusions: Routinely collected national registry variables provided moderate discrimination for unsuccessful TB treatment outcomes, and complex machine-learning models added only about 0.01 in ROC-AUC over logistic regression. Combining multivariable regression with interpretable machine learning provides complementary epidemiological and predictive insights; external validation and fairness evaluation are required before individual-level implementation.
No takes yet. Share an insight, caveat, or question.
Ahmad et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: