New Strategies in Tuberculosis Prediction: TB-CAST, an Explainable Baseline Score for Two-Month Sputum Culture...

Background: Early identification of patients who will remain sputum-culture-positive at two months could support risk-adapted tuberculosis (TB) monitoring. We developed a compact baseline score from routine admission data measured before anti-TB therapy. Methods: The cohort comprised 155 adults (62 culture-positive and 93 culture-negative at two months). Random Forest ranking and correlation clustering were used as exploratory selection steps. The complete development pipeline was evaluated by repeated fully nested cross-validation; a secondary conditional analysis held the full-cohort-selected candidate set fixed. Results: The fully nested pipeline yielded a mean test-fold AUC of 0.70 ± 0.09 (fold-level percentile interval 0.53-0.85). Conditional on full-cohort predictor preselection, mean test-fold AUC was 0.74 ± 0.08, and the patient-level mean repeated out-of-fold score AUC was 0.75 (bootstrap 95% CI 0.67-0.83). At provisional thresholds of ≥2, ≥3 and ≥4 points, conditional procedure sensitivity/specificity were 0.90/0.39, 0.80/0.60 and 0.57/0.74, respectively. The conditional repeated out-of-fold Brier score was 0.200, with calibration intercept approximately 0.00 and slope 0.74. Conclusions: TB-CAST is an exploratory score with moderate internal discrimination. The fully nested result is the primary internal estimate for the complete data-driven pipeline; the conditional results and provisional thresholds require prospective external validation before clinical use.