XGBoost-SHAP肺结节早期识别可解释性框架构建OA

Objective:To achieve early identification of pulmonary nodules and visual interpretation of key variables through interpretable machine learning,and to facilitate precise prevention,control,early diagnosis and treatment of lung cancer.Methods:This study enrolled individuals at high risk of lung cancer and completed clinical screening.Their high-risk assessment data and imaging results were extracted.Participants were divided into high-risk and low-risk groups for pulmonary nodules based on China's Lung Cancer Screening Standard(T/CPMA 013-2020).Variables with differences identified by univariate analysis were used as predictors,with pulmonary nodule grouping as the dependent variable,to construct an interpretable XGBoost-SHAP identification framework for early nodule detection and visual result interpretation.Results:A total of 644 high-risk individuals were included,with 199(30.9%)in the high-risk pulmonary nodule group.The XGBoost model achieved an accuracy of 0.914 6,sensitivity of 0.758 7,specificity of 0.984 3,F1-score of 0.845 8,and AUC of 0.974 1 for nodule grouping.SHAP analysis revealed that higher SHAP values-and thus increased risk of nodule enlargement-were associated with greater smoking intensity,exposure to secondhand smoke from colleagues/family,infrequent kitchen ventilation during cooking,excessive intake of processed foods,occupational exposure to asbestos/radon,insufficient intake of protein,fruits and vegetables,and manual labor occupation.Conclusion:The constructed interpretable framework performs well in early pulmonary nodule identification.Changes in nodule size are associated not only with traditional risk factors(e.g.,smoking habits,secondhand smoke exposure,cooking fume exposure,occupational asbestos/radon exposure)but also with the participants'dietary habits.