Predicting resistance to fluoroquinolones among patients with rifampicin-resistant tuberculosis: A cross-country validation study

Background Fluoroquinolones (FQs) are a cornerstone of most all-oral, shorter regimens endorsed by the World Health Organization for the treatment of rifampicin-resistant or multidrug-resistant tuberculosis (RR/MDR-TB). Knowledge of resistance to FQs can help guide regimen selection at the point of care. In settings where rapid testing for FQ resistance is unavailable, prediction models could support treatment decisions by identifying FQ resistance based on patient characteristics observable at the point of care. These prediction models have been typically developed and evaluated within a single country, and their generalizability across different geographic settings is unclear. Methods and findings We used data from 5,175 patients with RR-TB and available FQ drug susceptibility testing (DST) results submitted to the TB Portals, an open-access data-sharing platform curated by the National Institute of Allergy and Infectious Diseases, from eight countries (Azerbaijan, Belarus, Georgia, Kazakhstan, Kyrgyzstan, Moldova, Romania and Ukraine) between 2012 and 2024. Among these patients, 1,772 (34.2%) had FQ-resistant TB. We developed prediction models for FQ resistance using logistic regression, neural networks, and XGBoost. Models were evaluated under three strategies: (1) pooled models trained on multi-country data; (2) within-country models trained and evaluated using internal validation; and (3) cross-country models trained on subsets of countries and externally validated on held-out countries. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). Across the three algorithms, pooled models showed moderate optimism-corrected discrimination, with AUROC ranging from 0.70 to 0.72 and AUPRC ranging from 0.57 to 0.59. Within-country models demonstrated slightly better discrimination, with AUROC and AUPRC reaching 0.8 for some countries. Cross-country external validation showed that performance loss from using a model trained on external data could be negligible to >0.1 AUPROC or AUROC, depending on the country and algorithm. A limited set of predictors, including case definition and treatment-history-related variables, were among the most consistently informative predictors, whereas demographic, comorbidity, social-risk, education, and employment variables showed more variable contributions across countries and algorithms. A limitation of our study is that the incidence of RR-TB and FQ resistance was relatively stable in our analysis dataset. Hence, the results may not generalize to scenarios with marked changes in MDR-TB dynamics. Conclusions Predicting FQ resistance using demographic and clinical characteristics showed moderate ability to identify FQ resistance among patients with RR-TB, but their performance and predictor patterns varied across countries. Models developed for one or several countries cannot be assumed to generalize to other settings without rigorous external validation. These findings highlight the limitations of globally trained prediction models for RR/MDR-TB and underscore the need for locally informed prediction models to support clinical decision-making.