Development and validation of a nomogram using interpretable machine learning to integrate CT radiomics and PET metabolic parameters for predicting benign-malignant differentiation of pulmonary space-occupying lesions

Objective: To construct a multimodal machine learning model integrating computed tomography (CT) radiomics, Positron Emission Tomography (PET) metabolic parameters, and clinical data for differentiating benign from malignant pulmonary space-occupying lesions (PSOLs), and develop an interpretable nomogram for clinical application. Methodology: = 116, 69 malignant, 47 benign) at a 7:3 ratio according to the chronological order of patient enrollment, to avoid data leakage and rigorously assess model generalizability. All malignant lesions were confirmed by pathological examination, while benign lesions were confirmed by pathology (82%) or clinical-imaging follow-up for at least 12 months (18%). CT radiomic features with Intraclass Correlation Coefficient (ICC) values >0.75 were selected, and a Radiomics-score (Rad-score) was generated using the Least Absolute Shrinkage and Selection Operator (LASSO) regression algorithm. Integrated models [Logistic regression, random forest (RF), support vector machine (SVM), eXtreme Gradient Boosting (XGBoost)] were developed by fusing the Rad-score, clinical variables, and PET metabolic parameters. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1-score, and Brier score. Model calibration was assessed via calibration curves, and clinical utility was validated by decision curve analysis (DCA). Model interpretability was achieved using SHapley Additive exPlanations (SHAP) values for the optimal XGBoost model, and a clinically applicable, interpretable nomogram was constructed based on the core predictive features identified by SHAP analysis to facilitate clinical translation. Results: SUVmax, Rad-score, and delayed phase total lesion glycolysis (TLG_d) as the top three key predictive features. Conclusions: The predictive logic of the optimal XGBoost model was decoded via SHAP analysis to identify core predictive features, and a clinically applicable, interpretable nomogram was further established based on a multivariate logistic regression model using these core features, to facilitate the clinical translation of our model.