The stakes of this question are far from academic. Fluoroquinolones, including drugs such as levofloxacin and moxifloxacin, form the backbone of the shorter, all-oral treatment regimens that the World Health Organization now endorses for patients whose tuberculosis does not respond to rifampicin, a key first-line antibiotic. When a physician knows that a patient’s tuberculosis strain is susceptible to fluoroquinolones, a shorter regimen becomes a realistic option, sparing the patient months of toxic therapy. When the strain is resistant, that regimen would fail, and an alternative must be assembled. The problem is that in many of the places where drug-resistant tuberculosis burdens are heaviest, laboratory capacity for rapid drug susceptibility testing—the gold standard for detecting resistance—is limited or absent, leaving clinicians to make consequential decisions with incomplete information.

Prediction models have emerged as one proposed bridge across this diagnostic gap. The idea is elegant in principle: by feeding routinely collected patient information—age, sex, treatment history, comorbidities, social risk factors—into a statistical algorithm, a model could estimate the probability that a patient’s tuberculosis is already resistant to fluoroquinolones, even without laboratory confirmation. Clinicians could then weigh that probability when selecting a regimen, much as they weigh other forms of clinical judgment. But most prediction models in this field have been developed and evaluated within a single country, using data from one health system and one circulating population of tuberculosis strains. Whether such models generalize to different epidemiological settings has remained an open and largely untested question.

To answer it, the research team—Tianfang Shao, Mariana R. Neves, Molly Franke, Carole Mitnick, Jennifer Furin, Ted Cohen, and Reza Yaesoubi—turned to the TB Portals, an open-access data-sharing platform curated by the National Institute of Allergy and Infectious Diseases. From this resource, they assembled records for 5,175 patients with rifampicin-resistant tuberculosis who had documented fluoroquinolone drug susceptibility testing results, drawn from eight countries: Azerbaijan, Belarus, Georgia, Kazakhstan, Kyrgyzstan, Moldova, Romania, and Ukraine, with data collected between 2012 and 2024. The burden in this cohort was substantial: 1,772 patients, or 34.2 percent, carried tuberculosis strains that were already resistant to fluoroquinolones. That high prevalence reflects both the severity of the drug resistance epidemic in Eastern Europe and Central Asia and the clinical urgency of identifying resistant cases before committing patients to a regimen that depends on these drugs.

Methodologically, the study was designed to stress-test generalizability in a way that few previous analyses have attempted. The researchers built prediction models using three distinct algorithms spanning a range of complexity: logistic regression, a transparent and widely used statistical technique; neural networks, which can capture intricate nonlinear relationships among predictors; and XGBoost, a gradient-boosted decision tree method that has become a favorite in clinical prediction research for its flexibility and performance. Each algorithm was then evaluated under three different training and validation strategies. The first strategy pooled data from all eight countries to train a single multi-country model. The second trained and tested models within individual countries using internal validation. The third—and most revealing—trained models on data from a subset of countries and then externally validated them on countries deliberately held out of training, simulating exactly what happens when a model built elsewhere is imported into a new setting.

Model performance was quantified using two standard metrics. The area under the receiver operating characteristic curve, or AUROC, measures how well a model separates resistant from susceptible cases across all possible decision thresholds, with 0.5 representing chance performance and 1.0 representing perfect discrimination. The area under the precision-recall curve, or AUPRC, is often considered more informative when the outcome of interest is relatively uncommon, because it focuses on how precisely the model identifies true positive cases. The pooled, multi-country models achieved what the authors describe as moderate discrimination after optimism correction, with AUROC values ranging from 0.70 to 0.72 and AUPRC values from 0.57 to 0.59 across the three algorithms. In practical terms, such models carry real information, but they would misclassify a substantial fraction of patients if used alone to route individuals toward or away from fluoroquinolone-based regimens.

Within-country models performed somewhat better, with AUROC and AUPRC values reaching 0.8 for some countries, suggesting that locally trained models can extract more signal from their own patient populations than a globally pooled model can. The cross-country external validation, however, exposed the fragility of geographic transfer. When a model trained on external data was applied to a held-out country, the loss in performance ranged from negligible to more than 0.1 in AUROC or AUPRC, depending on the country and the algorithm involved. A drop of that magnitude is clinically consequential: it can mean the difference between a model that usefully flags resistant infections and one that generates enough false reassurance to steer patients toward regimens destined to fail. The variability across settings indicates that no single number can capture how well a model will travel, and that each new deployment demands its own rigorous evaluation.

Just as striking as the performance differences were the differences in which predictors mattered. A limited set of variables—chiefly those related to the case definition and to a patient’s treatment history—emerged as the most consistently informative predictors across countries and algorithms. This makes biological and epidemiological sense: patients who have been treated for tuberculosis before, particularly with fluoroquinolones or other second-line drugs, face a higher likelihood of harboring strains that have already evolved resistance. By contrast, demographic characteristics, comorbidities, social risk factors, education, and employment variables contributed far more inconsistently, with their predictive weight swinging widely across countries and modeling approaches. This instability suggests that the social and structural determinants of drug resistance are woven into local contexts in ways that a model trained elsewhere cannot simply assume.

The authors are careful to note an important limitation of their analysis. The incidence of rifampicin-resistant tuberculosis and of fluoroquinolone resistance was relatively stable across the years represented in their dataset, which means their findings may not extend to settings where the dynamics of multidrug-resistant tuberculosis are changing markedly—for example, where new regimens are rapidly reshaping patterns of resistance. Prediction models are, at their core, pattern-recognition devices calibrated to the populations and time periods that produced their training data, and shifts in transmission, treatment practices, or drug exposure can quietly invalidate the relationships a model has learned. Any real-world deployment would therefore need ongoing monitoring and periodic recalibration, not just a one-time validation.

The study’s conclusions carry a clear message for the field of tuberculosis care and for clinical prediction modeling more broadly. Predicting fluoroquinolone resistance from demographic and clinical characteristics is feasible to a moderate degree, and locally developed models can offer meaningful support to clinicians working without rapid laboratory testing. But models developed for one country, or even for several countries pooled together, cannot be assumed to perform elsewhere without rigorous external validation in the destination setting. As all-oral, shorter regimens spread across the globe and diagnostic capacity lags behind, the researchers argue that the path forward lies in locally informed prediction models—tools built and validated on the populations they are meant to serve—rather than in the export of algorithms whose accuracy quietly dissolves at national borders.

Subject of Research: Cross-country validation of prediction models for fluoroquinolone resistance in rifampicin-resistant tuberculosis

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

Article References: Shao, T., Neves, M. R., Franke, M., Mitnick, C., Furin, J., Cohen, T., & Yaesoubi, R. (2026). Predicting resistance to fluoroquinolones among patients with rifampicin-resistant tuberculosis: A cross-country validation study. PLOS Medicine, 23(9), e1004965. https://doi.org/10.1371/journal.pmed.1004965

Image Credits: AI Generated

DOI: 10.1371/journal.pmed.1004965

Keywords: tuberculosis, fluoroquinolone resistance, rifampicin-resistant tuberculosis, prediction models, drug susceptibility testing, machine learning, XGBoost, external validation, TB Portals, WHO treatment regimens, multidrug-resistant tuberculosis, clinical decision-making