Risk Prediction Models Of Central Lymph Node Metastasis In Cn0 Papillary Thyroid Microcarcinoma

BackgroundAt present,the value of prophylactic central lymph node dissection(PCLND)in papillary thyroid carcinoma(PTC)patients who have no preoperative or intraoperative evidence of nodal metastasis(cN0)remains unclear.Moreover,currently available preoperative or intraoperative features cannot accurately identify cNO PTC patients who have central lymph node metastasis(CLNM),so it is difficult to decide whether to implement PCLND.It appears to be a good choice to implement selective PCLND for some cNO PTC patients,which can not only avoid a lot of unnecessary PCLND and prevent over treatment,but also reduces the risk of a second surgery.However,the results of post-operative pathological examination suggested that current guidelines do not perform very well.Therefore,it is always very necessary to perform individual patient-level predictive analytics of CLNM in cNO PTC patients,which can not only guide the decision of PCLND and choose suitable candidates for PCLND,but also optimize outcomes at the patient and population levels.A review of the studies that have been carried out on the prediction model of CLNM of PTC patients,all of which have different limitations,including:(1)There are many research design limitations,such as lack of verification and poor extrapolation ability;(2)The predictive indicators used by the various research institutes vary greatly,and many of the indicators are not routinely available before PCLND;(3)There is heterogeneity in the study population,including the patients who performed prophylactic and therapeutic dissection of central and lateral compartment neck;(4)Unclear clinical application value,lack of evaluation of clinical usefulness,etc.In addition,most of the predictive models for CLNM of PTC patients are constructed using logistic regression methods,and few studies involve other methods.ObjectiveBy analyzing preoperatively and intraoperatively identifiable clinical,ultrasonography and pathological features of cN0 patient,explore the risk factors of central zone lymph node metastasis.Three statistical methods of logistic regression(LR),random forest(RF)and support vector machine(SVM)were used to construct a CLNM prediction model that meets the characteristics of cNO PTC patients.And evaluate the performance of each model and comparative study,and on this basis to choose good models using bayesian model averaging(BMA)method to build a combined forecasting model,in order to build an optimal statistical model to provide a basis for the decision-making of PCLND.MethodsThis study retrospectively collected 753 eligible cNO PTC patients and divided them into training set(n=606)and test set(n=147)according to the operation time.Combined with clinical risk factors,preoperative ultrasonography features and intraoperative pathological information,LR,RF and SVM were used to construct the prediction model of CLNM in accordance with the characteristics of patients with cNO PTC.On this basis,models with good performance were selected and the combination forecasting model was constructed by BMA method.The discriminative ability of each model was compared and evaluated by the receiver operating characteristic curve(ROC)and the area under the curve(AUC),and the AUC of the training set and test set was compared by the bootstrap method.Evaluate the calibration ability of each model through the calibration curve.The clinical usefulness of each model was evaluated by decision curve analysis and compared with four popular PCLND strategies.ResultsSeven predictors were included as candidates for model construction,including age,sex,taller than wide shape,microcalcification,tumor size,multifocality,and extrathyroid extensions.Among the three individual models,the AUC of the RF model performed best in the training set,but the AUC of the LR model performed best in the test set.Moreover,both in the training set and the test set,the AUC of each model was superior to the 2015 American thyroid association guidelines and the 2018 Chinese thyroid cancer diagnosis and treatment guidelines.The calibration curve shows that the LR model,the RF model and the BMA model performed well in the training set and the test set,while the SVM model performed relatively poorly.The decision curve shows that the RF model performs best in the training set,while the LR model performs better in the test set.In contrast,the combination forecasting model of LR model and RF model has a good net benefit in both the training set and the test set.In addition,when the threshold probability of clinical decision is between 30%and 60%in both the training and test sets,the net benefit of PCLND decision using each model is higher than that of the four popular PCLND strategies.ConclusionsThe four risk assessment models of CLNM constructed in this study in accordance with the characteristics of cNO PTC patients have a good distinguishing ability and are all better than the four currently popular PCLND strategies.Doctors or patients can choose the appropriate threshold probability between 30%and 60%according to their preference for PCLND,and make PCLND decisions by using each model,which can achieve higher net benefits than the four currently popular PCLND strategies.In view of the good performance of the combination forecasting model of LR model and RF model in the training set and test set,we believes that these two models can be used in combination to jointly evaluate the risk of CLNM in cNO PTC patients,so as to screen suitable candidates for PCLND.