Calibrated Risk Prediction for Loss to Follow-Up in Drug-Susceptible Tuberculosis: A Decision-Analytic Machine-Learning Framework Developed on a Parameter-Anchored Synthetic Cohort

Hadi, Muhammad Abdul
Falahuddin, Qoonita Dzakiyya

Jurnal Ragam Pengabdian
Vol. 3 No. 2 (2026): Mei-Agustus, Sustainable Development Goals (SDGs): Multidisciplinary Perspectiv

Loss to follow-up is the principal leak in Indonesia’s tuberculosis care cascade, yet adherence support is rationed by staffing rather than by need. We developed a risk model for loss to follow-up in drug-susceptible pulmonary tuberculosis on a synthetic cohort of 20,000 patients whose parameters were anchored to published estimates, principally a national analysis of the Indonesian Tuberculosis Information System covering 71,665 patients. Penalised logistic regression, random forest and gradient boosting were compared on a held-out test set of 6,000 patients, with isotonic recalibration, bootstrap confidence intervals, decision curve analysis and a capacity-constrained triage rule. Observed loss to follow-up was 18.6 per cent. Discrimination was closely similar across learners, with areas under the curve of 0.735 for logistic regression and 0.739 for calibrated boosting, whereas calibration slopes ranged from 1.480 to 1.017. Flagging the highest-risk fifth of the cohort captured 46.0 per cent of events at a positive predictive value of 42.5 per cent, a 2.29-fold enrichment. Algorithm choice proved inconsequential; calibration and explicit decision analysis determined whether the model could be used. The estimates describe a simulation and require validation on real surveillance data.