关键词: 结核, 肾脏病学, Logistic模型, 决策树, 神经网络(计算机), 预测

Abstract:

Objective: To evaluate the application value of logistic regression model, decision tree model, and neural network model in predicting active tuberculosis among patients with chronic kidney disease (CKD). Methods: A retrospective analysis was conducted on 392 CKD patients admitted to the Third People’s Hospital of Kunming between January 2021 and January 2024. Among them, 266 patients with active tuberculosis were included in the observation group, and 126 patients without active tuberculosis were included in the control group. Clinical data and laboratory indicators were collected for both groups. Logistic regression, decision tree and neural network models were used to identify influencing factors and construct the risk prediction models. The model performance was compared using receiver operating characteristic (ROC) curve. Results: The area under the curve (AUC) of the logistic regression model was 0.726 (95%CI: 0.766-0.777), with a sensitivity of 45.1% and a specificity of 92.1%. Alcohol consumption, hemoptysis, low lymphocyte count, low hematocrit, high uric acid, high fibrinogen degradation products, and low IL-10 level were identified as independent predictors (all P values <0.05). The AUC of the decision tree model was 0.825 (95%CI: 0.783-0.868), with a sensitivity of 62.0% and a specificity of 82.5%. Fibrinogen degradation products served as the primary stratification variable, and further included variables such as CD4⁺ T lymphocytes, lymphocyte count, hematocrit, loss of appetite, and uric acid to construct a decision path. The neural network model demonstrated the best predictive performance, with an AUC of 0.876 (95%CI: 0.843-0.909), a sensitivity of 60.9%, and a specificity of 98.4%. Feature importance analysis indicated that IL-10, hematocrit, fibrinogen degradation products, and CD4⁺ T lymphocytes were the top four predictors. Conclusion: The neural network model exhibited the best predictive ability for detecting CKD patients with active tuberculosis. Key predictors identified in this study—such as low lymphocyte count, low CD4⁺ T lymphocyte count, and high fibrinogen degradation product—may help define the key population for screening. The three models provide complementary strenghths and may be applied synergistically applied in clinical practice. Enhanced tuberculosis screening is warrented for CKD patients with immunosuppression (e.g., low lymphocyte count, CD4⁺ T lymphocyte count), activation of coagulation and fibrinolysis (e.g., increased fibrinogen degradation products), or hemoptysis.

Key words: Tuberculosis, Nephrology, Logistic models, Decision trees, Neural networks (computer), Forecasting