Construction and application of a diagnostic model for tuberculosis in patients with pneumoconiosis

Objective To develop and validate a machine learning-based model for diagnosing the presence of tuberculosis (TB) in patients with pneumoconiosis, utilizing complex clinical data to support early identification and clinical decision-making. Methods This retrospective case-control study analyzed the risk of TB among patients with pneumoconiosis using data extracted from electronic medical records. A total of 325 patients with occupational pneumoconiosis who were admitted to Beijing Chest Hospital, Shilong Hospital, and Fuxing Hospital between January 2019 and June 2024 were included. Participants were classified into a case group (pneumoconiosis with TB) and a control group (pneumoconiosis without TB). Diagnostic variables included demographic characteristics, occupational exposure history, medical history, laboratory parameters, etiological indicators, and imaging features. The study outcome was the occurrence of TB. Multivariable regression analysis was performed to identify independent risk factors for concomitant pulmonary TB. Based on the selected variables, eight machine learning models were developed to construct diagnostic algorithms. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis. The best-performing model was further interpreted using the Shapley Additive Explanations (SHAP) framework. Results Regression analysis identified ten key diagnostic factors associated with TB in pneumoconiosis patients: etiological type, serum albumin, PaO 2 /FiO 2 ratio, HbA1c, calcification, bilateral lung lesions, cavitation, prior antibiotic use, pleural effusion, and duration of dust exposure. These diagnostic factors were further classified into three categories: occupational factors (etiological type, duration of dust exposure), nutritional and metabolic indicators (serum albumin, PaO 2 /FiO 2 ratio, HbA1c), and radiological features (calcification, bilateral lung lesions, cavitation, pleural effusion). Eight machine learning models were developed using these variables. After comprehensive evaluation of discrimination, calibration, clinical utility, and model stability, the random forest (RF) model demonstrated the best overall performance. The RF model achieved an AUC of 0.997 (95% CI: 0.994–1.000) in the training set and 0.905 (95% CI: 0.833–0.977) in the validation set, which outperformed other models. Conclusion Patients with pneumoconiosis are at a high risk of developing TB. Machine learning models can effectively aid in identifying TB in this population, with the RF model showing superior diagnostic performance. This approach may facilitate early auxiliary diagnosis and timely clinical intervention.