A comprehensive international analysis of individual participant data from 7,750 patients treated for multidrug-resistant or rifampicin-resistant tuberculosis has demonstrated that clinical risk stratification algorithms can reliably predict unfavorable outcomes, including treatment failure and death. The study, conducted by an international research team and published in Nature Communications, established multivariable prediction models based on routine pre-treatment clinical characteristics as well as subsequent on-treatment markers such as sputum culture conversion. By categorizing patients into low-, medium-, and high-risk phenotypes, the researchers identified distinct prognostic trajectories across diverse global cohorts. Furthermore, the predictive accuracy of these algorithms was successfully confirmed using independent external validation data from 378 participants enrolled in contemporary clinical trials evaluating the all-oral six-month bedaquiline, pretomanid, and linezolid regimen.
Multidrug-resistant or rifampicin-resistant tuberculosis presents one of the most formidable challenges in global infectious disease control, arising from Mycobacterium tuberculosis strains resistant to rifampicin or to at least both isoniazid and rifampicin. Although curable with second-line antimicrobial regimens, standard therapeutic courses have historically remained lengthy, complex, and burdened by significant drug toxicities. Globally, the treatment success rate for drug-resistant tuberculosis stood at approximately 68 percent in 2023, falling well short of global targets. While the World Health Organization released consolidated guidelines in 2025 recommending shorter all-oral courses such as the six-month BPaLM regimen for eligible adolescents and adults without fluoroquinolone resistance or extensive disease, clinical practice has increasingly recognized that uniform, one-size-fits-all approaches risk either undertreating individuals with severe disease manifestations or subjecting lower-risk patients to unnecessarily prolonged, toxic regimens.
To establish an empirical foundation for precision medicine in tuberculosis management, the investigators assembled individual-level records encompassing 12,938 patients from 52 observational and experimental studies conducted across 38 countries and territories between 2009 and 2018. Following rigorous exclusion criteria that omitted participants lost to follow-up or those lacking critical records on treatment duration and time to culture conversion, the primary analysis dataset retained 7,750 individuals. Among this analyzed cohort, 5,869 patients achieved documented treatment success, representing 76 percent of the population, while 628 patients experienced treatment failure and 1,253 patients died during treatment, representing failure and mortality rates of 8 percent and 16 percent, respectively. The average participant age across the analysis cohort was 37.6 years, the mean body mass index was 20.23 kilograms per square meter, 59.6 percent were male, and 34.4 percent were living with HIV, 87 percent of whom were concurrently receiving antiretroviral therapy.
Drug resistance profiling within the analysis cohort revealed extensive phenotypic and therapeutic complexity. Approximately 24.5 percent of participants harbored baseline resistance to fluoroquinolones, while 27.0 percent demonstrated resistance to second-line injectable agents. Moreover, 77 percent of patients had at some point received treatment with a drug to which confirmed microbiological resistance was subsequently documented. Therapeutic regimens across the historical dataset exhibited wide variation, with total treatment durations spanning from 0.1 to 112 months and averaging 19.8 months. Notably, newer core medications were utilized in varying proportions: delamanid was administered to only 1 percent of individuals, whereas bedaquiline was prescribed to 21.8 percent and linezolid to 19.5 percent of participants, with the concurrent clinical use of linezolid and bedaquiline demonstrating a moderate correlation of 55 percent.
Through multivariable logistic regression modeling, the authors systematically delineated the specific baseline clinical variables driving mortality and treatment failure. Both unfavorable endpoints were independently associated with lower body mass index, prior tuberculosis treatment involving second-line drugs, the presence of lung cavities, and documented pyrazinamide resistance. Specifically, each unit increase in body mass index reduced the odds of death by an adjusted odds ratio of 0.87 and the odds of failure by an adjusted odds ratio of 0.94. Mortality was strongly elevated among patients living with HIV, who experienced an adjusted odds ratio of 2.24, as well as older individuals, male patients, and those presenting with extrapulmonary disease involvement. Conversely, microbiological treatment failure was specifically linked to baseline sputum smear positivity for acid-fast bacilli, active cigarette smoking, and confirmed baseline resistance to fluoroquinolones or second-line injectable agents.
To construct the formal risk stratification architecture, the investigators compared mixed-effects logistic regression against mixed-effects gradient tree boosting algorithms. Mixed-effects logistic regression models demonstrated superior predictive performance across repeated cross-validation and independent test splits. The baseline prognostic model, designated Model A, incorporated exclusively pre-treatment clinical and demographic characteristics, achieving an area under the receiver operating characteristic curve of 0.728 with a 95 percent confidence interval ranging from 0.726 to 0.731. Within Model A, baseline HIV seropositivity emerged as the single most potent predictor of unfavorable outcomes, carrying an adjusted odds ratio of 1.91. Although patients living with HIV who received antiretroviral therapy showed somewhat lower relative risk than untreated individuals, a substantial excess hazard of treatment failure and death persisted across all HIV-positive individuals.
The research team subsequently enhanced model discrimination by incorporating therapeutic and microbiological variables into sequential algorithms. Model B integrated baseline clinical features with the specific antimicrobial agents administered throughout therapy, elevating the receiver operating characteristic area under the curve to 0.746. Within this model, the clinical receipt of linezolid, bedaquiline, and levofloxacin was significantly associated with reduced odds of unfavorable outcomes, yielding adjusted odds ratios of 0.73, 0.45, and 0.75, respectively, whereas para-aminosalicylic acid administration was associated with increased odds of an unfavorable outcome. Model C further incorporated on-treatment sputum culture conversion at six months. The achievement of culture conversion by six months proved to be the single strongest negative predictor of adverse clinical outcomes, yielding an adjusted odds ratio of 0.097 and driving the area under the receiver operating characteristic curve to 0.813 among patients remaining on treatment beyond six months.
Translating these statistical equations into discrete risk phenotypes, the authors defined patient risk tiers based on predicted probabilities of unfavorable outcome derived from Model A: low risk was defined as a predicted probability below 10 percent, medium risk as 10 to 15 percent, and high risk as greater than 15 percent. In the primary dataset, Kaplan-Meier curves confirmed that the cumulative probability of unfavorable outcome differed significantly across these tiers over time since treatment initiation. Furthermore, baseline risk classifications exhibited strong stability: 91 percent of individuals initially assigned to the low-risk category remained low risk at six months, with only 8 percent shifting into the high-risk classification. In contrast, among medium- and high-risk patients who achieved culture conversion, substantial proportions were safely re-stratified downward into lower risk categories during interim therapeutic evaluations.
External validation was conducted using an independent cohort of 378 participants pooled from the Nix-TB, ZeNix, and TB-PRACTECAL clinical trials, all of whom received the standard six-month BPaL regimen. Baseline demographic characteristics closely matched the primary analysis cohort, with a mean age of 36 years, average body mass index of 21 kilograms per square meter, and 32 percent HIV prevalence. Across the validation cohort, overall treatment success reached 90 percent. When applying the risk algorithms, Model A stratified patients into groups with statistically distinct outcome trajectories: participants in the low-risk phenotype attained a 97 percent treatment success rate at completion, compared to 91 percent among individuals in the combined medium-to-high-risk phenotype. When applying Model C incorporating interim culture conversion, low-risk participants similarly achieved 97 percent success, whereas medium-to-high-risk participants achieved an 88 percent success rate.
The study authors highlighted several methodological considerations and study limitations inherent to retrospective individual patient data pooling. The historical analysis dataset relied on studies published through 2018 with treatment enrollments concluding in 2016, preceding widespread programmatic adoption of the newest six-month regimens. Additionally, post-treatment follow-up data on recurrence were available for less than one percent of historical participants, restricting primary model derivation to on-treatment endpoints, and individuals lost to follow-up were necessarily excluded. However, external validation against prospective Phase 3 trial data encompassing 12-to-18-month post-treatment relapse monitoring confirmed the durability of the stratification rules. The validated risk framework provides an empirical basis for prospective clinical evaluation, directly informing the Phase 2C/3 PRISM-TB clinical trial designed to test risk-stratified treatment durations for fluoroquinolone-susceptible multidrug-resistant tuberculosis.
Subject of Research: Patient risk stratification algorithms and treatment outcomes in drug-resistant tuberculosis
Article Title: Risk stratification for MDR/RR-TB: a comprehensive integrated patient-level analysis from TB programs and contemporary clinical trials
Article References: Garcia-Cremades, M., Campbell, J. R., Strydom, N., Solans, B. P., Mouksassi, S., Li, Z., Phillips, P. P. J., Velásquez, G. E., Nahid, P., Menzies, D., & Savic, R. M. (2026). Risk stratification for MDR/RR-TB: a comprehensive integrated patient-level analysis from TB programs and contemporary clinical trials. Nature Communications, 17(1), Article 11337. https://doi.org/10.1038/s41467-026-77931-0
Image Credits: AI Generated
DOI: 10.1038/s41467-026-77931-0
Keywords: multidrug-resistant tuberculosis, rifampicin-resistant tuberculosis, risk stratification, patient risk phenotypes, bedaquiline, pretomanid, linezolid, mixed-effects logistic regression, ROC AUC, culture conversion, treatment success, HIV coinfection, cavitary disease, BPaL regimen, clinical prediction models, individual participant data, global tuberculosis, SMART4TB consortium
News Source: Ophelia Keating. (November 16, 2026). Patient Risk Stratification Models Accurately Predict Multidrug-Resistant Tuberculosis Outcomes. Scienmag.



