Establishment And Diagnostic Performance Evaluation Of Risk Prediction Models For Papillary Thyroid Carcinoma In Patients With Thyroid Nodules

Purpose:We aimed to analyze the clinical information such as routine preoperative examinations results of patients accepted thyroid surgery for thyroid nodule（TN）and explore the influencing factors of papillary thyroid carcinoma（PTC）.Simple models of the risk of papillary cancer in patients with thyroid nodules were then established based on the general condition of the patient,preoperative laboratory examination indicators and ultrasound results.Furthermore the diagnostic performance were evaluated.Materials and Methods:It was a retrospective study.Clinical data of patients undergoing thyroid surgery due to thyroid nodules in Department of general surgery,Nanjing drum tower hospital from January 2018 to January 2020 were collected.Which included general conditions such as gender,age,height,weight,blood pressure,calculated body mass index（BMI）,etc;and laboratory tests such as thyroid function related hormones and antibodies,liver and kidney function,lipid indexes,electrolytes,blood routine,and coagulation function;as well as imaging tests,ie ultrasound TI-RADS（Thyroid imaging,reporting and data system）classification（Kwak 2011 version）.Patients were divided into two groups as PTC and benign TN according to the postoperative pathological results.The independent sample t test,Wilcoxon rank sum test and chi-square test were used for difference analysis according to the characteristics of the data.Spearman correlation analysis was used to analyze the clinical characteristics of PTC patients.Based on Logistic regression analysis and Back propagation neural network（BPNN）,we constructed prediction models for the nature of patients with thyroid nodules,plotted the receiver operation characteristic（ROC）curves according to the predicted and true values of each model,and calculated the sensitivity, specificity,negative predictive value（NPV）and positive predictive value（PPV）of the predictive model diagnosis.P<0.05 was considered statistically significant.Result:1.A total of 2028 patients were included in this study,with an average age of 45.26±12.98 years old.There were 546 benign TN patients and 1482 PTC patients diagnosed by postoperative routine pathology.We got 525 male and 1503 female and the male to female ratio was 1/2.86.2.Through difference analysis,we found that PTC patients have higher TSH, Tg Ab,TPOAb,Alb,UA,HDL-C and Apo A-I levels,while lower age,systolic blood pressure（SBP）,thyroglobulin,DBIL,Fasting Glucose,BUN,blood sodium and blood chlorine levels than benign TN patients.Multivariate Logistic regression analysis indicated that TSH（Risk ratio,OR=1.112）and positive thyroid autoantibodies（OR=1.633）were independent risk factors for papillary carcinoma in patients with thyroid nodules,while patient age（OR=0.966）,thyroglobulin（OR=0.996）and HDL-C（OR=0.502）levels were independent protective factors.（P values are all<0.05）3.Multivariate Logistic regression analysis showed that the male gender（OR=1.270,95%CI 0.994～1.622,P=0.058）,younger than 45 years old（OR=1.270,95%CI 1.025～1.574,P=0.029）,maximum tumor diameter<1 cm（OR=1.439,95%CI1.158～1.787,P=0.001）,and multifocality（OR=1.576,95%CI 1.266～1.960, P<0.001）were risk factors for cervical lymph node metastasis.4.We took X₁（Age）,X₂（TSH）,X₃（thyroid autoantibodies,Tab）,X₄（Thyroglobulin）,X₅（HDL-C）,X₆（TI-RADS classicfication）as independent variables and pathological results（PTC and benign TN）as the dependent variable and established the regression equation as follow:Logit（P）=-1.885-0.035X₁+0.106X₂+0.491X₃-0.004X₄-0.690X₅+2.423X₆.The AUC obtained by plotting the ROC curve was 0.940.With patient age,gender,SBP TSH,Tab,thyroglobulin,TI-RADS classification,Apo A-I,and Albumin as input variables and pathological results as output variables,we constructed a BP neural network model structure.The MATLAB software randomly selected 80%of the input cases as the training set,the remaining20%as the prediction set.After learning the 1622 training set,the network model output the prediction value of 406 patients in the prediction set.Finally the AUC for its ROC curve was 0.943.Both models have a high degree of fit to the data.Besides,the AUC of single TI-RADS diagnosis ROC was 0.867.Conclusion:The value of serum TSH,positive thyroid autoantibody（Tg Ab and TPOAb）in patients with thyroid nodules were independent risk factors for PTC,while age and HDL-C were probable protective factors.In PTC patients,the male gender,younger than 45 years old,maximum tumor diameter<1 cm,and multifocality were risk factors for cervical lymph node metastasis.Our study based on Logistic regression analysis as well as BP neural network built models of PTC risk in patients with thyroid nodules.The diagnostic performance of our BP neural network model was not inferior to either Logistic regression model or single TI-RADS,which can be used as an optional auxiliary tool to improve the clinical diagnosis accuracy of thyroid papillary carcinoma.