The Application Of Dynamic Bayesian Network In Differential Diagnosis Of Benign And Malignant Pulmonary Nodules

Objectives:Lung cancer is the main cause of cancer deaths worldwide.The differential diagnosis of benign and malignant pulmonary nodules is the focus and difficulty of current clinical work.In our research,the theory of Bayesian Network(BN)is applied to predict the malignant probability of pulmonary nodules in the differential diagnosis of benign and malignant pulmonary nodules.Our research developeds a novel static prediction model for the malignant probability of pulmonary nodules by combining the patient’s medical history,demographic characteristics,CT imaging characteristics,serum tumor marker indicators and some follow-up information.Compared with the traditional prediction model,the model developed in our research further improves the ability to differentiate between benign and malignant pulmonary nodules due to the inclusion of more available information.Our research further applies the theory of Dynamic Bayesian Network(DBN)to the differential diagnosis of benign and malignant pulmonary nodules,and develops a dynamic prediction model.The dynamic model can predict the malignant probability of pulmonary nodules based on the patient’s previous follow-up examination results and clinical data,so as to provide more objective evidence before the patient undergoes clinical biopsy or surgical resection and other invasive operations,avoiding unnecessary invasive operations.Methods:(1)According to the inclusion criteria,a total of 981 lung nodule patients are enrolled in the research in Zhong Da Hospital,Nanjing.Among the 981 patients,there are 647 cases(65.95%)of malignant nodules and 334 cases(34.05%)of benign nodules.Chi-square test and Fisher’s exact probability method are conducted to screen the factors that may be related to the benign and malignant outcomes of pulmonary nodules.Our research uses 10-fold cross-validation to validate the generalization abilities of models.The performance of models is evaluated by sensitivity,specificity,accuracy,receiver operating characteristic curve(ROC)and area under the curve(AUC).Finally,the comprehensive performance of BN model and other traditional prediction model(including MAYO model,Brock model and PKUPH model)are compared.(2)Our research collects case data from January 2014 to December 2020 who went to Zhong Da Hospital,Nanjing for pulmonary nodule follow-up.According to the case selection criteria,a total of396 patients with pulmonary nodules are enrolled in the research.Among the 396 patients,there are214 cases(54.04%)of malignant nodules,182 cases(45.96%)of benign nodules.All of the 396 patients have participated in more than one follow-up of pulmonary nodules,217 patients have participated in more than two follow-ups and 129 patients participated in more than three follow-ups.Our research develops a DBN model based on appropriate modifications to the structure of BN model’s considering the characteristics of the follow-up data.Finally,the comprehensive performance of DBN model and BN model are compared.All the analyses are performed using SAS 9.4 and R3.6.0.Results:According to the purpose of our research,the results are divided into the following two parts:(1)Diagnosis model of pulmonary nodules based on BNAfter the screening of factors,it can be found that age group,pulmonary tuberculosis history,number of pulmonary nodules,nodule location,maximum nodule length,nodule type,lobular sign,spicule sign,pleural depression sign,vacuole sign,vessel convergence sign,calcified nodules,CEA level,CYFRA21-1 level,first and last follow-up time interval,and first and last follow-up nodule change variables are statistically significant between the benign and malignant pulmonary nodule groups,so the above variables are included in the model development.By comparing the 10-fold cross-validation results of each model,it can be found that among the BN models developed using three different methods,the BN-C model developed by the hybrid method has the most excellent overall prediction performance: sensitivity = 0.811,specificity = 0.808,accuracy = 0.810,AUC =0.854.In the comparison between BN and other machine learning models,it can be found that the artificial neural network model has the highest sensitivity,0.916,but the lowest specificity,0.620.The random forest model has the highest overall accuracy rate of 0.828,but its specificity is also low,of 0.692.The BN-C model developed by the hybrid method has the highest specificity(0.808)and AUC(0.854),and the sensitivity(0.811)is also maintained at an acceptable level.It also can be found in the comparison with the MAYO model,the Brock model and the PKUPH model,the BN-C model is better than the MAYO model,the Brock model and the PKUPH model in all performance indicators.(2)Diagnosis model of pulmonary nodules based on DBNAfter verifying the ability of the DBN model to identify the deterioration trend of pulmonary nodules,it can be found that the sensitivity of the model remains at a high level(0.880 to 1.000)at each follow-up point,while the specificity,the overall accuracy and AUC of the model are lower at the T1 follow-up point.However,with the increase in the times of follow-up,the specificity,overall accuracy and AUC of the model steadily increase,and reaching a higher level(0.886-0.959)from the T3 follow-up point.After verifying the early diagnosis ability of the DBN model,it can be found that89.3%(75/84)of the patients who participated in at least two follow-ups of pulmonary nodules and are finally diagnosed with lung cancer can be diagnosed in advance through the model.After verifying the ability of the DBN model to predict the final outcome of pulmonary nodules,it can be found that the sensitivity,overall accuracy,and AUC of the model remains at a high level(0.798～1.000)at each follow-up point,while the model specificity is lower at T0 and T1 follow-up point,but with the increase in the times of follow-up,the specificity improves rapidly.In the comparison between DBN and BN,it can be found that the predictive performance of the two models is steadily improved with the increase in the times of follow-up.For patients who participated in multiple followups,the prediction accuracy of the two models reach a very high level.At most follow-up points,the performance of the DBN model are better than the BN-C model.Conclusion:(1)Compared with the BN model developed by other methods and other commonly used machine learning models,the BN-C model developed by the hybrid method has the highest specificity(0.808)and AUC(0.854),and the sensitivity(0.811)is also maintained at an acceptable level,which shows that the overall prediction performance of the BN model is excellent.It can accurately detect lung cancer patients while reducing the misdiagnosis rate to the greatest extent,thereby avoiding unnecessary invasive operations.In addition,compared to other “Black Box” models,the graphical nature of the BN enables it to reflect in a more intuitive form the complex relationships between various factors and the pulmonary nodules.(2)The DBN model developed in our research has excellent ability to identify the deterioration trend of pulmonary nodules and predict the final outcome of lung nodules,and the accuracy of the model steadily improves with the increase in the times of follow-up.In addition,our research further explores the model’s ability to diagnose lung cancer patients in advance,and found that 89.3%(75/84)of the patients who participated in at least two follow-ups of pulmonary nodules and are finally diagnosed with lung cancer can be diagnosed in advance through the model.This result shows that the DBN model can identify the deterioration trend of lung nodules before clinicians,which is conducive to the early detection,diagnosis and treatment of lung cancer patients,thereby improving the survival rate of lung cancer patients.(3)The DBN model developed in our research has better performance at most follow-up points than the BN model.The DBN model can use more follow-up information when predicting the malignant probability of pulmonary nodules,increasing the total amount of information that can be used for prediction,so that more accurate predictions can be made.It can be foreseen that the DBN model will have a good application value in the follow-up of patients with pulmonary nodules.