An explainable multi-label Diagnostic Prediction Model for lung diseases based on proteomic biomarkers.

OBJECTIVES: Differential diagnosis of pneumonia, tuberculosis, and lung cancer is highly challenging due to overlapping clinical presentations and high comorbidity rates. To overcome the limitations of traditional diagnostic methods, this study developed and validated an explainable machine learning model using proteomic data for multi-label classification of these three diseases. METHODS: Bronchoalveolar lavage fluid proteomic data were collected from 358 patients with confirmed lung diseases at the Shandong Provincial Public Health Clinical Center. We constructed a voting ensemble model integrating XGBoost, Random Forest, and Gradient Boosting algorithms based on clinical features, global proteomic statistical features, differentially expressed derived features, and disease-specific biomarker scores. Considering insufficient cancer samples and risk of missed diagnoses, a 1.5-fold weighting strategy was applied to the cancer class to enhance sensitivity. The model was evaluated using 5-fold stratified cross-validation and an independent external cohort of 110 cases. Feature contributions were interpreted using the Shapley Additive exPlanations (SHAP) method and biological significance of key proteins was determined using Gene Ontology functional enrichment analysis. RESULTS: The AUC values were 0.912±0.031 for tuberculosis detection and 0.813±0.037 for cancer detection. Thus, the ensemble model demonstrated excellent performance, significantly outperforming seven baseline models including logistic regression and support vector machines. SHAP analysis identified key protein biomarkers (Cancer: P02775, P61626; Tuberculosis: P0DOX2, P55259). Furthermore, the model achieved an overall accuracy of 86% with the independent external validation cohort. CONCLUSION: This study established an explainable, multi-label classification model based on proteomics that can provide a valuable reference for the differential diagnosis of complex lung diseases, especially those with comorbidities. The model shows good performance and interpretability, suggesting potential for precise diagnosis of lung diseases.