Metformin promotes smear conversion in tuberculosis‐diabetes comorbidity and construction of prediction models

Background The comorbidity of tuberculosis (TB) and diabetes mellitus (DM) is a global health concern. Metformin is commonly used in DM but the potential effectiveness in comorbid patients is uncertain. This retrospective study aims to investigate the effect of metformin on TB‐DM comorbidity and construct prediction models. Methods Patients diagnosed with TB‐DM in West China Hospital were retrospectively enrolled from Nov 2013 to Sep 2019. Electronic health records of patients were extracted. Two‐month smear conversion (2SC) was considered an outcome indicator of TB. Univariate and multivariate logistic regression (LR) were used to assess the role of metformin and other independent predictors. Meanwhile, prediction models were built by LR, elastic net regression, support vector machine, k‐nearest neighbors, and random forest. Results A total of 927 individuals were recruited, among which 408 (44.01%) were metformin‐exposed patients. A higher 2SC rate was observed in the metformin users. Other impact factors such as smoking, glucose, and creatinine levels were also identified. Multivariable models were then constructed using filtered variables. The support vector machine model yields the highest AUC (0.808, 95% CI: 0.767–0.849) and specificity (83.24%). LR model outperformed others in terms of sensitivity (69.71%). Conclusion This retrospective study of a large population from southwestern China provides strong clinical evidence for the positive effects of metformin in TB‐DM. Metformin is associated with a better therapeutic outcome and promising for the adjuvant therapy of TB‐DM. Furthermore, a combination of support vector machine and LR models is recommended to discriminate the patients with poor treatment outcomes. Patients with tuberculosis‐diabetes comorbidity were retrospectively enrolled to explore metformin's role in anti‐tuberculosis treatment. Metformin usage was recognized as an independent protective factor for 2‐month smear conversion by univariate and multivariate logistic regression (LR). Furthermore, prediction models based on electronic health records, including random forest, support vector machine, k‐nearest neighbors, elastic net regression and LR, are expected to identify the patients with poor prognosis. (The figure was created with BioRender.com.).