Interpretable machine learning model stratifies latent tuberculosis risk using low-cost data

Researchers in China have built an interpretable machine learning model that can separate active pulmonary tuberculosis from latent tuberculosis infection using nothing more than questionnaire-based epidemiological information, and then rank people with latent infection into three risk tiers to help decide who needs closer follow-up. The multicenter case-control study, published open access in BMC Public Health on 22 October 2026, enrolled 782 participants at 21 designated tuberculosis institutions across Henan Province. The cohort included 401 individuals with latent tuberculosis infection and 381 patients with newly diagnosed, treatment-naive pulmonary tuberculosis.

Because roughly a quarter of the world's population is estimated to carry latent tuberculosis infection, and because conventional screening tools have limited sensitivity while existing prediction models often depend on costly laboratory biomarkers, the team deliberately restricted itself to inexpensive, widely collectable predictors that grassroots health services could realistically gather. Methodologically, the investigators collected epidemiological predictors through standardized face-to-face questionnaires and randomly divided the dataset into training and test sets at a 7:3 ratio. Eleven predictors survived screening by ten-fold cross-validated LASSO regression, and the team then constructed and compared seven machine learning algorithms using a battery of metrics that included the area under the receiver operating characteristic curve, F1 score, Brier score, calibration curves, and decision curve analysis.

The eXtreme Gradient Boosting model, known as XGB, achieved the highest test-set performance, with an AUC of 0.818 (95% CI 0.763–0.873), an F1 score of 0.764 (95% CI 0.698–0.822), and a Brier score of 0.176 (95% CI 0.148–0.204). To translate the statistical output into something usable in a clinic, the researchers defined a three-tier risk framework based on a percentile strategy. Using probability cutoffs of greater than 0.763 for high risk and less than 0.328 for low risk, the model sorted participants into three tiers, and the proportion of actual pulmonary tuberculosis cases rose steeply across them, from 20.00% in the low-risk tier to 84.29% in the high-risk tier.

That gradient is the study's central practical claim: rather than treating everyone with latent infection as an undifferentiated mass, the model concentrates diagnostic suspicion and potential preventive intervention on a much smaller subgroup in which active disease is far more prevalent. The authors emphasize that the goal is prioritization for further evaluation, not a diagnosis in itself. A key design choice was transparency.

The team applied SHapley Additive exPlanations, or SHAP, to decompose the optimal model's predictions into individual feature contributions, and this analysis identified male sex, low body mass index, advanced age, unintentional weight loss within the previous six months, and low annual household income as the most influential features. The researchers also deployed the model as a free web calculator, lowering the barrier to use in resource-constrained primary care and public health settings where laboratory infrastructure is limited. The authors are careful about the limits of what they have shown.

This was a case-control design with newly diagnosed, treatment-naive pulmonary tuberculosis patients as cases, so the model distinguishes prevalent active disease from latent infection rather than predicting which latently infected individuals will progress to disease over time. The web calculator is described explicitly as a research prototype, and the paper states that its clinical utility requires external validation and prospective evaluation before it can guide preventive-treatment decisions. The training and test data came from a single province, and performance in other populations, transmission settings, and health systems remains untested.

Even so, the combination of strong discrimination, good calibration, SHAP-based interpretability, and reliance on low-cost epidemiological indicators makes the model a candidate tool for risk stratification among people with latent tuberculosis infection, and the open-access publication means other research groups can attempt replication with their own data.

Subject of Research: Development and validation of an interpretable machine learning model for risk stratification among individuals with latent tuberculosis infection

Article Title: Development and validation of an interpretable machine learning model for risk stratification among individuals with latent tuberculosis infection: a multicenter case-control study

Article References: Wang, W., Sun, D., Zhang, Y., Xu, X., Chang, W., Zheng, D., He, M., Xie, C., Qi, Y., Xu, Y., Zhu, J., & Diao, L. (2026). Development and validation of an interpretable machine learning model for risk stratification among individuals with latent tuberculosis infection: a multicenter case-control study. BMC Public Health. https://doi.org/10.1186/s12889-026-30066-9

Image Credits: AI Generated

DOI: 10.1186/s12889-026-30066-9

Keywords: latent tuberculosis infection, pulmonary tuberculosis, machine learning, XGBoost, SHAP, risk stratification, LASSO regression, case-control study, epidemiological predictors, web calculator, BMC Public Health, Development