Development And Validation Of Ultrasound Characteristics-based Thyroid Cancer Risk Prediction Model In Patients With Thyroid Nodules

Part1.Development and validation of ultrasound characteristics-based thyroid cancer risk prediction model in patients with thyroid nodulesObjective:To develop and validate an ultrasound characteristic based risk prediction model for thyroid malignancy from a thyroid nodule cohort.Methods:A consecutive thyroid nodules cohort was established,and the demographic,clinical,biochemical,ultrasound characteristics of thyroid nodules were prospectively recorded before fine needle aspiration(FNA).FNA cytology for all nodules,pathology for those received surgical procedure were collected.The whole cohort was randomly assigned to training set and validation set.Univariate Logistic regression was used to determine the association of nodules’ characteristics with thyroid malignancy.Multivariate Logistic regression was used to establish a thyroid cancer prediction model.Model presentation was fascinated by a risk score for thyroid cancer.The measures of discrimination were the AUC with receiver operating-characteristic(ROC)curves,sensitivity,specificity of model.Calibration refers to the degree of agreement between model-derived probabilities and observed probabilities.The calibration of model was determined by Hosmer-Lemeshow goodness of fit test.Results:(1)Our cohort consisted of 2333 thyroid nodules.The training cohort included 1183 thyroid nodules.The average size for thyroid nodules was 1.96± 1.05cm.Among thyroid nodules randomly assigned to training cohort,156 nodules(13.19%)were diagnosed as thyroid cancer.Internal validation cohort consists of 1027 thyroid nodules with 142 thyroid cancers.And external validation cohort was derived from Peking Union Medical College Hospital and Beijing Tongren hospital thyroid nodules ultrasound database,which included 68 benign thyroid nodules and 324 malignant thyroid nodules.(2)Univariate Logistic regression identified sex,age,multifocal,duration,suspicious cervical lymph nodes,external thyroid extension(ETE)and composition,echogenicity,shape,margin,calcification of nodules as variates significantly associated with thyroid cancer.Further multivariable Logistic regression showed sex,age,multifocal and composition,echogenicity,aspect ratio(A/T),margin,calcification of nodules independently associated with thyroid cancer.A prediction model for thyroid cancer was developed using these 8 variables.And TNScore was developed as presentation of our new models.(3)Internal and external validation both showed TNScore had good predictive performance,with AUC of 0.845[95%CI(0.823-0.866),p<0.001]in internal validation cohort and 0.948[95%CI(0.920-0.968),p<0.001]in external validation cohort.The calibration of TNScore was also excellent with x2 value of 7.188(p=0.516)and 3.641(p=0.820)in internal and external validation respectively.Conclusion:A risk-assessment tool,which we developed to integrate demographic and ultrasound characteristic of thyroid nodules and was validated in internal and external cohort,can be used to estimate the probability of thyroid cancer in clinical practice.Part2.Comparison of TNScore with models from 2015 ATA guideline and 2017 ACR TIADSObjetives:To compare performance of our new thyroid cancer clinical prediction model TNScore with that of models from 2015 ATA guideline and 2017ACR TIRADS.Methods:Comparison was performed using Xiamen cohort and Beijing cohort respectively.The diagnostic performance of our model was compared with that of models from 2015 ATA guideline and 2017ACR TIRADS.Model discrimination was evaluated with ROC curve,sensitivity and specificity of models.Model calibration was evaluated with reclassification index,while decision curve analysis(DCA)was used to compare the models’clinical net benefit.Results:(1)A total of 2332 thyroid nodules were included in Xiamen Cohort with 2034 benign thyroid nodules and 298 malignant thyroid nodules.The average age for was 46.0±13.1 years,and 1908 patients were female.The average thyroid nodule size was 1.95±1.02cm cm.The Beijing cohort include 379 thyroid nodules with 62 benign thyroid nodules and 317 malignant thyroid nodules.(2)For Xiamen cohort,the sensitivity,specificity and AUC of TNScore for diagnosing thyroid cancer were 66.90%,87.83%,0.845[95%CI(0.823,0.866),p<0.001]respectively,while the sensitivity,specificity and AUC of ATA model being 67.13%,76.88%,0.751[95%CI(0.725,0.776),p<0.001].And for ACR TIRADS model,the sensitivity,specificity and AUC were 53.15%,93.48%,0.815[95%CI(0.791,0.837),p<0.001].The difference of AUC between TNScore and ATA model was 0.0941[95%CI(0.0513,0.137),p<0.001],while that being 0.0313[95%CI(0.0024,0.0602),p=0.034]for TNScore and ACR TIRADS model.TNScore also demonstrated good calibration as compared with ATA,ACR TIRADS models,with absolute NRI of 3.2%,4.7%,and additive NRI of 15,8,respectively.DCA showed better net clinical benefit by TNScore as compared with that by ATA,ACR TIRADS models.(3)For Beijing cohort,the sensitivity,specificity and AUC of TNScore for diagnosing thyroid cancer were 89.59%,92.06%,0.950[95%CI(0.922,0.969),p<0.001]respectively.The difference of AUC between TNScore and ATA model was 0.0350[(0.0134,0.0566),p=0.001)],and 0.0303[(0.0040,0.0566),p=0.0238]for TNScore and ACR TIRADS model.For calibration,TNScore was better than ATA and ACR TIRADS models,with absolute NRI of 5.1%,3.6%,and additive NRI of 13.2,7.6 respectively.Conclusion:The diagnostic performance of TNScore is better than that of ATA,TIRADS models in Chinese thyroid nodules cohort.