Development of an early auxiliary diagnostic model for disseminated tuberculosis in HIV-negative tuberculosis patients.

Background: Disseminated tuberculosis (DI-TB) can induce multiple organ dysfunction syndrome and carries a high mortality rate. This study aimed to establish an early-warning diagnostic model to facilitate the identification of DI-TB high-risk patients and improve clinical prognosis. Methods: We retrospectively enrolled 1,122 HIV-negative hospitalized patients with disseminated tuberculosis (DI-TB) or pulmonary tuberculosis (PTB) admitted to Nanjing Second Hospital. Using conventional laboratory tests and multidimensional lymphocyte subset analysis, a diagnostic model was constructed employing machine learning algorithms. Feature importance analysis was performed using the Shapley Additive Explanations (SHAP) framework. Results: We screened nine significant features for model development. The XGB model demonstrated favorable diagnostic performance in the internal test set, achieving an AUC of 0.983, a Brier score of 0.025, an accuracy of 0.949, sensitivity of 0.914, specificity of 0.961, F1-score of 0.897, MCC of 0.864, NPV of 0.972, and PPV of 0.881. The SHAP interpretability system ranked features by importance, illustrating their positive or negative contributions to XGB predictions and explaining how individual features influenced model outputs. Conclusion: This study developed and internally evaluated a machine learning-enabled auxiliary risk-stratification model for identifying dissemination risk among HIV-negative hospitalized TB patients. The XGB model showed favorable internal performance and may provide a potential auxiliary tool for early DI-TB risk assessment, but prospective external validation is still required before routine clinical application.