To investigate the independent predictive factors for diagnosis of tuberculous meningitis (TBM) based on regional characteristics in Qinghai, China, and to construct and validate a diagnostic prediction model.

The clinical data of 931 patients with infectious meningitis, including 739 cases with TBM and 192 cases with non-tuberculous infectious meningitis (non-TBM) treated at Qinghai Provincial Fourth People’s Hospital from January 2019 to January 2024 were collected, retrospectively. They were divided into a training set (651 cases, including 522 cases of TBM and 129 cases of non-TBM) and a validation set (280 cases, including 217 cases of TBM and 63 cases of non-TBM) with a ratio of 7∶3. The influencing factors of TBM diagnosis were screened by Lasso regression and multivariate Logistic regression, and a nomogram prediction model was constructed. The diagnostic efficiency of five machine learning models (decision tree, Logistic regression, random forest, support vector machine and extreme gradient boosting) were evaluated and compared by the receiver operating characteristic (ROC) curve.

The general demographic data and biochemical parameters between patients in training set and validation set were without significant differences (all P>0.05); the baseline characteristics between patients in TBM group and non-TBM group across both the training set and validation set were also without significant differences (all P>0.05). By Lasso dimensionality reduction screening of predictive variables and multivariate regression analysis, a total of 6 independent predictors for TBM were identified: consciousness disorder (OR=2.988, 95%CI: 1.242-7.555, P=0.017), gamma interferon release assay (IGRA) (OR=4.279, 95%CI: 2.074-9.120, P<0.001), CSF monocyte ratio≥50% (OR=4.061, 95%CI: 1.88-9.174, P<0.001), magnetic resonance imaging (MRI) suggesting tuberculoma (OR=7.080, 95%CI: 2.019-34.77, P=0.006), MRI suggesting cranial base meningeal enhancement (OR=3.022, 95%CI: 1.429-6.519, P=0.004) and computerized tomography (CT) showing pulmonary lesions (blood disseminated tuberculosis: OR=46.140, 95%CI: 8.765-856.9, P<0.001). AUC of the Logistic regression model in training set and validation set were both 0.948 (95%CI: 0.927-0.969, 0.920-0.975), with sensitivities of 87.16% and 85.71%, and specificities of 89.92% and 90.48%. The Extreme Gradient Boosting (XGB) model performed the best in diagnosing TBM (AUC=0.954).

The Logistic regression model based on clinical and imaging features exhibits excellent diagnostic accuracy for TBM and is suitable for clinical application.