Diagnostic performance of stacking ensemble combined with SHAP interpretation for benign and malignant pulmonary space-occupying lesions

Le résumé fourni par la source

Early diagnosis and treatment of lung cancer are critical for improving patient survival rates and prognosis. The diagnosis of lung cancer relies on multimodal features, yet the performance of single models remains limited. Current studies have insufficiently explored multimodal feature fusion strategies and the intrinsic decision-making mechanisms of models, restricting their clinical application. To evaluate the performance improvement of a Stacking ensemble algorithm integrating radiological characteristics, clinical data, and laboratory indicators in diagnosing benign and malignant pulmonary occupying lesions, and to interpret the model's decision-making mechanism using the SHAP method, thereby providing a novel strategy for clinical intelligent auxiliary diagnosis. This study retrospectively enrolled 618 pulmonary space-occupying lesions from 595 patients with pathologically confirmed Pulmonary Space-Occupying Lesions. Participants were divided into a Training set and Test Set at a 7:3 ratio, with a separate independent External validation set (126 lesions from 118 patients) established. Employed an enhanced multi-algorithm voting integration strategy (Boruta, stability selection, and LASSO regression) to perform feature screening. Eleven fundamental classification models were constructed using selected stable features (including AdaBoost, ExtraTrees, KNN, Logistic Regression, CatBoost, RF, SVM, MLP, GBM, LightGBM, and XGBoost) and optimized through fusion with the Stacking algorithm. Model efficacy was assessed using the receiver operating characteristic curve (AUC), Brier score, and clinical decision curve (DCA), with the SHAP attribution method employed to analyze feature importance and model ensemble weights. In the test set, the Stacking-XGBoost model demonstrated optimal comprehensive diagnostic performance (AUC 0.83, 95% CI 0.76–0.89), with false-negative cases decreasing from 17 to 13 compared to its XGBoost base model (AUC = 0.80, 95% CI 0.73–0.87). The learning curve confirmed model convergence after the sample size reached 400 cases. The calibration curve demonstrated high concordance between select models predicted probabilities and actual incidence (XGBoost Brier = 0.1641; Stacking-RF Brier = 0.1504), though the Hosmer–Lemeshow test ( p < 0.01) indicated some degree of calibration deviation. Decision curve analysis indicated clinical net benefit for the core model across broad threshold ranges. External validation demonstrated variable performance across models, with the base ExtraTrees model achieving the highest AUC of 0.78 (95% CI 0.69–0.86), followed by the Stacking-GBM ensemble with an AUC of 0.76 (95% CI 0.67–0.84). Ablation analysis confirmed that multimodal fusion significantly outperformed the imaging-only model (AUC 0.802 vs. 0.739, p = 0.019). SHAP interpretation was performed at three levels: model-level contribution identified KNN and XGBoost as the highest-weighted base models in the meta-model.; ensemble input contribution summarized LUNG-RADS, Pleural Contact Area, Surface Area, and Age as common core features across models, with CEA showing in specific base models (e.g., XGBoost); raw feature contribution confirmed that LUNG-RADS ranked first in global feature importance, while imaging-based high-risk indicators (e.g., Pleural Contact Area) and laboratory tumor markers (CYFRA21-1) demonstrated consistent positive decision contributions across all base models. The Stacking ensemble framework based on multi-modal features (especially Stacking-XGBoost) demonstrated a positive trend in reducing missed diagnosis risks for malignant lesions, while improving diagnostic accuracy and generalization capability. Combined with SHAP's multi-level interpretation mechanisms, this synergistic pattern integrating imaging and serum indicators provides a referable clinical paradigm for developing interpretable AI-assisted diagnostic systems.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.