To Explore The Predictive Factors Of Benign And Malignant Pulmonary Nodules And The Construction Of Prediction Model

Objective（s）:The clinical data,imaging features of pulmonary nodules,quantitative parameters of artificial intelligence and tumor markers were analyzed,and the predictive factors of benign and malignant pulmonary nodules were discussed and a visual line chart prediction model was constructed to provide the basis for personalized diagnosis and treatment of pulmonary nodules.Methods:A total of 271 patients with pulmonary nodule in the Ganmei Hospital of the first people’s Hospital of Kunming from December 1,2019 to November30,2022 were collected retrospectively.The clinical data,imaging features,quantitative parameters of artificial intelligence and tumor markers were sorted out.The variables with statistical significance were screened out by univariate analysis,and three kinds of feature engineering（Stepwise-Logistic regression from classical statistical perspective,Lasso-Logistic regression from statistical learning perspective,random forest and e Xtreme Gradient Boosting from machine learning perspective）method were selected for variable screening and model construction.The De Long test was used to determine whether the 95%confidence interval of the area under the receiver operating characteristic curve of different models overlapped,and the Hosmer-lemeshow test was used to evaluate the fitting degree of the model.The visual calibration curve was drawn by Bootstrap free sampling and repeated sampling for 500 times.The clinical generalization of the model was evaluated by decision curve analysis.Finally,a nomogram was drawn.Results:1.General information:Among the 271 patients,there were 116 males（42.8%）and 155 females（57.2%）.The ages were between 16 and 83,the median age was 56（47,66）years old.87（32.1%）had a history of smoking and 184（67.9%）did not smoke.There were 61 cases（22.5%）with lung diseases,including 27 cases of chronic obstructive pulmonary disease（10%）,22 cases of pulmonary tuberculosis（8.1%）,11cases of bronchiectasis（4.1%）and 1 case of bronchial asthma（0.4%）.17 cases（6.3%）had a family history of lung cancer,of which 6 cases（2.2%）were first-degree relatives.Benign lesions were found in 132 cases（48.7%）,malignant lesions in 139cases（51.3%）.Including ground glass nodules in 107 cases（39.5%）,partial solid nodules in 56 cases（20.7%）,pure nodules in 108 cases（39.8%）,and The diameter of nodules was between 6.0 and 29.8 mm.The median diameter was 12.4（10.3,18.2）mm.2.In the training group,the clinical features,artificial intelligence quantitative parameters,imaging features and serum tumor markers of patients with pulmonary nodules were analyzed by univariate analysis.The results showed that there were significant differences in sex,age,history of chronic disease,artificial intelligence parameters,type of pulmonary nodule,vacuole sign,vascular convergence sign,burr sign,pleural traction sign,boundary of pulmonary nodule,SCC,CYFRA21-1 and Pro GRP between benign and malignant groups.However,there was no significant difference in family lung cancer history,lung disease history,smoking history,nodule site,air bronchial sign,CA15-3,CA19-9,CA125,CEA and NSE.3.The Lasso-Logistic model is chosen according to the simplicity and interpretability of the model.Multivariate regression analysis showed that vascular convergence sign,volume,percentage,HU value,AI predictive risk probability,CYFRA21-1 and Pro GRP were independent risk factors for malignant pulmonary nodules.The Odds Ratio of AI predictive risk probability,vascular convergence sign,volume and Pro GRP were>3.4.The area under the ROC curve of the Lasso-Logistic model is 0.939（95%CI:0.892～0.986）.The specificity and sensitivity for the prediction and evaluation of malignant pulmonary nodules are 71.8%and 99.8%,respectively.HL test results show that the model has a high degree of fitting to the predicted results,and DCA indicates that the model has good clinical practicability.According to the constructed prediction model of benign and malignant pulmonary nodules,the risk prediction value of malignant pulmonary nodules corresponding to the line map is more than95%.Conclusion（s）:The prediction model established in this study is composed of vascular convergence sign,volume,HU value,real ratio,AI predictive risk probability,CYFRA21-1 and Pro GRP as follows:malignant probability of pulmonary nodule=e ^Y/（1+e ^Y）,Y=-1.97+1.28×X _{vascular convergence sign}+1.36×X _volume-1.42×X_HUvalue-1.44×Xactual ratio+1.93×XAI predictive risk probability+1.10×XPro GRP+0.94×XCYFRA21-1.The model has high accuracy in predicting benign and malignant pulmonary nodules and clinical practical value,so it provides a practical tool for individual diagnosis and treatment of pulmonary nodules.