Introduction

Tuberculosis (TB) remains one of the most pressing global health threats. In 2024, an estimated 10.7 million people developed TB, and 1.23 million died from the disease [1,2]. Despite progress in diagnosis and treatment, the burden of TB has been exacerbated by the rise of drug-resistant TB, particularly multidrug- or rifampicin-resistant TB (MDR/RR-TB) [3–5]. MDR/RR-TB is difficult to treat (with a success rate of 71% in 2022) and should be treated with second-line regimens, which include multiple antibiotics [1,6]. The selection of antibiotics for treating MDR/RR-TB should ideally be guided by DST. However, access to rapid and comprehensive DST remains limited in many high-burden settings, resulting in substantial delays in identifying resistance to second-line agents such as fluoroquinolones (FQs) [7]. FQs are a cornerstone of most all-oral, shorter regimens endorsed by the World Health Organization for the treatment of RR/MDR-TB [6]. However, FQ resistance continues to challenge efforts to treat MDR/RR TB; in 2024, an estimated 18% of MDR/RR TB cases globally were resistant to any FQs tested [1,8].

Recent studies have shown the potential of machine-learning approaches for predicting resistance to specific antibiotics using patient-level data. In Moldova, a clinical risk-based model for predicting resistance to FQs demonstrated reasonable discrimination (area under the receiver operating characteristic curve >0.8) using basic patient characteristics, including age, urban residence, and TB type [9]. Similarly, a study in Brazil developed prediction models for tuberculosis drug resistance using clinical and social characteristics from patients in São Paulo [10]. Together, these studies suggest that clinical and demographic information may contain usable signals for predicting resistance and that prediction models could enhance antibiotic selection in settings with limited or no access to DST.

A key limitation of existing studies is that they develop and evaluate resistance prediction models within a single country. Hence, little is known about whether these models generalize to other epidemiological contexts [9–11]. Geographic differences in social determinants, comorbidities, diagnostic pathways, and circulating resistance patterns may lead to substantial variation in predictor–outcome relationships, potentially limiting the transportability of models trained in one setting [12,13]. Understanding the extent to which prediction models can be effectively transferred across countries is essential before these tools can be meaningfully integrated into global TB programs.

To address this gap, we used records of patients with RR-TB from eight countries in the broader Eastern Europe and Central Asia region, provided by the TB Portals database [14], to systematically evaluate the within-country performance and cross-country transportability of three machine-learning models for predicting FQ resistance. Our analysis quantifies the impact of geographic heterogeneity on prediction performance and evaluates the transportability of these risk prediction models across geographic regions. Our findings offer insight into the feasibility and limitations of using clinical risk-based models to guide second-line therapy globally.

Methods

Data sources and study populations

We conducted a secondary analysis of de-identified clinical data from TB Portals, an open-access multinational repository of tuberculosis case data curated by the National Institute of Allergy and Infectious Diseases (NIAID) [14]. TB Portals aggregates standardized, multi-domain patient records—including demographic, clinical, and DST data—from 13 countries, with an emphasis on drug-resistant TB cases. Data are collected as part of routine practice in TB clinics, research studies, and clinical trials. Some of the data originate from historical records, while others are collected prospectively. All data adhere to the HL7 Fast Healthcare Interoperability Resources standard and are validated by clinicians at entry, though variability in original collection protocols persists.

Our initial dataset included 17,297 TB cases from TB Portals. For our analysis, we focused on patients with confirmed RR-TB who had definitive DST results for FQs. Detailed resistance classification rules and preprocessing steps are provided in the S1 Appendix. The number of patients with RR-TB and FQ DST method and coverage among patients with RR-TB varied substantially across countries and over time (Tables A–B and Fig A in S1 Appendix). For this analysis, we included only country-level annual data with a relatively stable prevalence of RR-TB and FQ-DST coverage above 80%. The final sample included 5,175 RR-TB patients from eight countries: Kazakhstan, Belarus, Georgia, Moldova, Kyrgyzstan, Azerbaijan, Romania, and Ukraine, with registration dates between 2012 and 2024 (Table 1). Time series for the percentage of patients with RR-TB and FQ DST coverage among patients with RR-TB included in our analysis are displayed in Fig B in S1 Appendix.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Country-specific inclusion of RR-TB patients with available fluoroquinolone DST results.

https://doi.org/10.1371/journal.pmed.1004965.t001

Results

Our analysis included 5,175 individuals with RR-TB and confirmed FQ DST results from eight countries, of whom 1,772 (34.2%) had confirmed FQ resistance (Table 2). The prevalence of FQ resistance varied substantially across countries, ranging from 13.1% in Kyrgyzstan to 57.8% in Ukraine. In our final sample, most individuals were male (69.7%), newly diagnosed with TB (65.9%), and had pulmonary TB (92.8%). Compared with individuals with confirmed FQ susceptibility, those with FQ resistance were more often classified as treatment failure cases (21.8% versus 3.8%) and less often classified as newly diagnosed TB cases (36.8% versus 65.9%) (Table 2). Several demographic, clinical, and social risk factors were available for modeling, although multi-response variables, such as social risk factors and comorbidities, had substantial numbers of unknown or unspecified responses (Table 2 and Table C in S1 Appendix). Overall, the characteristics of individuals with RR-TB (N = 5,357) were broadly similar to those with RR-TB and DST results (N = 5,175) (Table C in S1 Appendix).

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Summary of clinical and demographic characteristics of patients with RR-TB who had a drug susceptibility testing (DST) result for FQ resistance in our dataset.

https://doi.org/10.1371/journal.pmed.1004965.t002

Pooled logistic regression, neural network, and XGBoost models, which are trained on the entire data from all included countries, showed moderate optimism-corrected discrimination (Fig 1). Optimism-corrected AUROC ranged from 0.70 to 0.72 across the three algorithms, while optimism-corrected AUPRC ranged from 0.57 to 0.59. Pooled logistic regression had the highest optimism-corrected AUROC and AUPRC, although differences across algorithms were small. The difference between the optimism-corrected and 80/20 split estimates was negligible (Fig C in S1 Appendix).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Apparent and optimism-corrected performance of pooled models trained on the entire dataset.

Error bars represent 95% confidence intervals of the optimism-corrected estimates. AUROC: area under the receiver operating characteristic curve; AUPRC: area under the precision-recall curve. The performance of a model trained on 80% of the data and evaluated on the remaining 20% is presented in Fig C in S1 Appendix.

https://doi.org/10.1371/journal.pmed.1004965.g001

When the pooled model’s performance was examined by country, AUROC was relatively similar across countries (orange dots in Fig 2A–2C), whereas AUPRC varied more substantially between countries (orange dots in Fig 2D–2F). This difference was expected because AUPRC is more sensitive to the underlying prevalence of FQ resistance. Countries with lower FQ resistance prevalence, such as Kyrgyzstan and Kazakhstan, tended to have lower AUPRC values, even when AUROC was not markedly lower. In contrast, countries with higher FQ resistance prevalence, such as Ukraine and Azerbaijan, showed higher AUPRC values across several modeling strategies.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Comparing the performance of models that are trained on the entire dataset (the pooled model), trained and evaluated on the data from a single county (within-country models), and trained on data with one country excluded and assessed on the data from the excluded country (cross-country models).

Error bars represent 95% confidence intervals. AUROC: area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve.

https://doi.org/10.1371/journal.pmed.1004965.g002

Across evaluation strategies, model performance varied by country, algorithm, and metric (Fig 2). Within-country models were often comparable to or slightly higher than the pooled models, particularly for logistic regression and XGBoost (comparing green dots with orange dots in Fig 2A and 2C). This pattern was less consistent for the neural network model, where the green, orange, and blue dots were closer together for several countries and confidence intervals were wider in some settings, such as Romania and Kazakhstan (Fig 2B). In the AUPRC panels, differences between strategies were more pronounced (Fig 2D–2F). Within-country models often had higher AUPRC than pooled or cross-country models, especially in countries with higher FQ resistance prevalence, such as Azerbaijan, Belarus, and Ukraine. However, this pattern was not uniform across all countries and algorithms; for example, in some lower-prevalence settings, including Kazakhstan, Kyrgyzstan, and Georgia, AUPRC remained relatively low across strategies. Cross-country models, evaluated on data from a country not included in the training set, performed slightly worse than the pooled models across all countries and modeling approaches (compare blue and orange dots in Fig 2).

Using a permutation-based feature importance method, we identified that for pooled models, case definition variables were the most consistently important predictors across algorithms, particularly new TB and treatment failure. This finding was supported by logistic regression odds ratios and SHAP analyses for the neural network and XGBoost models (Figs D–F in S1 Appendix). In the logistic regression models, treatment failure and other previously treated case definitions were associated with higher odds of FQ resistance, whereas new TB was generally associated with lower odds (Fig D in S1 Appendix), a conclusion also supported by SHAP values for neural network and XGBoost models (Figs E–F in S1 Appendix). Other variables, including BMI, number of daily contacts, number of children, comorbidity indicators, and selected social risk factors, showed smaller or less consistent contributions (Figs 3A and D–F in S1 Appendix). Country-specific models showed more heterogeneous patterns of feature importance. The case definition remained an important predictor in country-specific models, but the relative contributions of other predictors varied across countries (Figs 3B–3F and D–F in S1 Appendix). For example, the number of daily contacts contributed more strongly in Azerbaijan, whereas BMI and age of onset were more prominent in Ukraine (Fig 3, panels B and F). Some comorbidity, education, employment, and social risk variables also contributed in selected countries, although these patterns were not consistent across all algorithms or settings.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Permutation-based feature importance for pooled (Panel A) and within-country models (Panels B–F).

For each potential predictor, bars indicate the mean decrease in area under the precision-recall curve (AUPRC) after exclusion. Larger decreases indicate a greater contribution of the feature to model discrimination. Georgia, Kazakhstan, and Kyrgyzstan were not included in the country-specific feature-importance panels because their country-specific model performance was less stable, making feature-importance interpretation less reliable.

https://doi.org/10.1371/journal.pmed.1004965.g003

Discussion

In this multi-county analysis, we evaluated the performance and generalizability of prediction models that use a patient’s basic demographic and clinical characteristics to identify resistance to FQs. We found that pooled models, trained on data from all countries, demonstrated moderate overall performance in detecting resistance to FQs (Fig 1). However, their performance varies across countries, particularly for AUPRC, which is sensitive to the underlying prevalence of FQ resistance (Fig 2). Models trained within individual countries achieved moderately better discrimination than models trained on pooled multi-country data or applied across countries. The performance loss from using a model trained on external data ranged from negligible to >0.1 AUPROC or AUROC, depending on the country and algorithm (Fig 2). These findings suggest limitations in the cross-country applicability of prediction models for identifying FQ resistance based on basic patient characteristics available at the point of care. Although these patient characteristics contain some information for identifying FQ resistance, their predictive power is shaped by the country context. Setting-specific factors, such as healthcare systems and epidemiological contexts, may influence the relationships between patient characteristics and FQ resistance. These setting-specific factors are typically not directly incorporated into individual-level risk prediction models, which may limit their transportability.

The countries included in our analysis are within the broader Eastern Europe and Central Asia region. Yet, our analysis indicated that FQ resistance prediction is context-dependent. Our analysis of permutation-based feature importance and SHAP values for pooled and within-country models showed that across all countries and algorithms, case definition (new or treatment failure) emerged as the only predictor that was consistently identified as highly informative (Figs 3, D–F in S1 Appendix). In contrast, the relative importance of other predictors, including BMI, number of daily contacts, comorbidity indicators, education, and employment status, varied between countries and algorithms (Figs 3, D–F in S1 Appendix). These differences suggest that the contribution of individual patient characteristics to resistance prediction may be influenced by local factors not directly incorporated into the risk prediction models. As such, individual-level risk prediction models may not provide universally stable predictors for MDR-TB.

Although these countries share several epidemiologic characteristics, including a high burden of RR/MDR-TB, the observed limitations in transportability may reflect multiple sources of heterogeneity. These include differences in the relationships between measured predictors and FQ resistance, variation in unmeasured factors such as prior fluoroquinolone exposure and treatment history, temporal changes in resistance patterns and treatment practices over the study period, and differences in resistance ascertainment methods. Because our analysis was not designed to distinguish among these mechanisms, the observed decline in cross-country performance should be interpreted as evidence that predictor–FQ resistance relationships are not fully stable across settings rather than as evidence for any single source of heterogeneity.

The limited advantage of neural network and XGBoost models over logistic regression is notable. One possible explanation is that the available predictors were relatively coarse clinical and demographic variables, rather than high-dimensional microbiological or genomic features. When the predictor set contains only a few strong signals, such as case definition or prior treatment history, flexible models may have little additional structure to learn. In addition, several predictors had substantial missingness or heterogeneous coding across countries, which may further reduce the benefit of complex algorithms. Therefore, our findings do not necessarily indicate that machine-learning methods are unsuitable for drug-resistance prediction, but rather that their added value may be limited when only routinely collected intake variables are available.

Previous studies on predicting drug-resistant tuberculosis have primarily focused on single-country settings and have relied on internal validation, limiting insights into model generalizability across populations [9,10]. In contrast, studies examining cross-country generalizability have primarily relied on genomic features rather than patient characteristics [23]. Our findings extend this literature by systematically comparing within-country, pooled multi-country, and cross-country prediction strategies using harmonized patient-level data from multiple countries. By evaluating transportability, our study demonstrates that while risk prediction models can perform well locally, their predictive relationships do not readily generalize across settings, highlighting an important limitation that has not been previously examined in a multi-country context.

Our study has several limitations. First, the incidence of RR-TB and FQ resistance was relatively stable in our analysis dataset, and hence, our findings may not generalize to scenarios with marked changes in MDR-TB dynamics. Second, although the analysis used a large multi-country dataset, the number of eligible RR-TB patients varied substantially across countries and years, and some country-specific estimates were based on smaller samples. This may have affected the stability of within-country performance estimates and feature-importance patterns. This concern is particularly relevant for neural networks and XGBoost, which may be more prone to overfitting smaller training sets. Third, we did not incorporate pathogen genomic data. While such data have been shown to improve resistance prediction, genomic sequencing remains unavailable or costly in many high-burden, resource-constrained settings. Our study intentionally focused on clinical and demographic predictors that are collected during patient intake, enabling treatment decisions at the point of care without requiring molecular testing infrastructure. Fourth, although the data were drawn from multiple countries, the final modeling workflow was restricted to a harmonized predictor set available across the selected countries. This restriction enabled pooled modeling and cross-country validation using a consistent feature space, but resulted in within-country models not utilizing all the variables that may have been available for individual countries. Fifth, among the 5,357 patients with RR-TB in our dataset, about 96.6% (N = 5,175) had DST results for FQ resistance. If DST is performed only for selected groups of patients (e.g., those with specific clinical or epidemiological risk factors), prediction models developed on this sample may not generalize to the entire population of patients with RR-TB. However, we observed a similar composition of patient characteristics among patients with RR-TB and patients with RR-TB and DST results (Table C in S1 Appendix), which could mitigate this issue. Finally, although FQ DST coverage was high in the selected analytical sample (Table 1), the diagnostic tests used for rifampicin and fluoroquinolone resistance determination differed across countries (Table A in S1 Appendix). Because testing practices varied by country, resistance ascertainment may not have been fully comparable across settings, potentially affecting generalizability.

Despite these limitations, the consistent patterns observed across countries strengthen the conclusion that predicting FQ resistance using patient characteristics alone can achieve reasonable accuracy but is inherently context-dependent. As such, prediction models trained and evaluated for a subset of countries may not be transportable to other countries. Future work can consider modeling approaches that explicitly account for geographic heterogeneity, such as hierarchical modeling frameworks that borrow strength across settings while preserving locally relevant risk relationships [24,25]. Future models may also benefit from combining routine demographic and clinical variables with microbiological, genomic, radiographic, treatment-history, and health-system-level predictors. These may include more granular prior TB and DR-TB treatment history, prior treatment outcomes, prior exposure to non-TB fluoroquinolones, hospitalization history, contact history with patients with pre-XDR or XDR-TB, and setting-level patterns of community- or over-the-counter-use of fluoroquinolones. Where rapid FQ DST is widely available, the incremental clinical utility of prediction models based only on demographic and clinical variables is likely limited. Their potential role may be greatest in settings where rapid second-line DST is unavailable, delayed, or inconsistently implemented, as interim triage tools while awaiting DST results, or for epidemiological risk stratification and program planning. More broadly, similar prediction approaches may also be relevant for newly introduced anti-TB drugs, particularly during early implementation when validated DST is not yet widely available, and the molecular basis of resistance remains incompletely understood [26].

Supporting information

Acknowledgments

Data were obtained from the TB Portals (https://tbportals.niaid.nih.gov), which is an open-access TB data resource supported by the National Institute of Allergy and Infectious Diseases (NIAID) Office of Cyber Infrastructure and Computational Biology (OCICB) in Bethesda, MD. These data were collected and submitted by members of the TB Portals Consortium (https://tbportals.niaid.nih.gov/Partners). We gratefully acknowledge the patients whose clinical, laboratory, and radiological data made this research possible, as well as the clinicians and site staff involved in their care and data collection.

Disclaimer: The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.