Chest X-ray TB triage AI adapted for Uzbekistan performs well internally but not better than baseline

Locally adapted chest X-ray AI for tuberculosis triage in Uzbekistan reached held-out ROC AUC 0.9902 (95% CI 0.9694–1.0000) but did not outperform the public baseline (p=0.4864); only 15 positive held-out cases limit conclusions.

• Primary held-out ROC AUC was 0.9902 (95% CI 0.9694–1.0000) with F1 0.9375; only 15 TB-positive cases in that set, so estimates remain imprecise.
• Site-stratified internal AUCs were 0.9844 at a district hospital and 0.9973 at a TB dispensary, but these were not institution-held-out validation.
• Limitations include no complete microbiological confirmation, enriched cohort, no external/prospective validation, and no statistical superiority over the public model (paired AUC difference 0.0054, p=0.4864).

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.