Abstract:

Objective: To develop a machine-learning-based predictive model for diagnosing tuberculous pleurisy (TBP) to improve clinical diagnostic accuracy. Methods: We retrospectively collected clinical data of 523 pleural effusion patients (375 with TBP and 148 with non-TBP) admitted in Xi’an Chest Hospital between January 2020 and December 2021. Fifteen indicators, including adenosine deaminase (ADA), tuberculosis infection T-cell spot test (T-SPOT.TB), and C-reactive protein (CRP), were incorporated. Seven machine learning algorithms, including random forest, support vector machine, and neural network, were employed to construct predictive models. Model performances were evaluated using 5-fold cross-validation. Feature importance was analyzed using SHapley Additive exPlanations (SHAP). Results: The model developed with Neural Network demonstrated optimal performance, achieving an area under the curve (AUC) of 0.932 on the test set, with an accuracy of 88.6%, precision of 94.4%, and recall rates of 89.3%. SHAP analysis identified ADA (SHAP value=0.12~0.18) and T-SPOT.TB (SHAP value=0.10~0.15) as two most significant predictors, with a notable synergistic effect (P<0.001). Conclusion: The Neural Network machine learning model developed in this study exhibited excellent diagnostic performance. Through interpretable analysis, key predictive factors and their interactions were elucidated, providing a novel tool for precise diagnosis of TBP. This model can assist clinical decision-making, particularly for cases in the “gray zone” under conventional diagnostic criteria.

Key words: Tuberculosis, Pleurisy, Diagnosis, computer-assisted, Models,statistical, Artificial intelligence algorithms