A Study On The Adjuvant Chemotherapy Decision In Intermediate Risk Breast Cancer Patients

Objective: The aim of this study was to analyze the factors related to adjuvant chemotherapy(ACT)decision in intermediate breast cancer patients and to evaluate the predictive accuracy of MNDACT clinical risk classification and PREDICT.We aimed to develop an initiative nomogram model integrated with clinico-pathologic parameters and recurrence score(RS)to predict the ACT recommendation in this subset of patients.Methods: A cohort of breast cancer patients undergoing 21-gene assay after surgery was retrospectively reviewed from January 2014 to December 2016 in Ruijin Hospital Shanghai Jiaotong University School of Medicine.The intermediate risk breast cancer was defined as RS of 11?30.The ACT decision of each patient was determined by multidisciplinary team(MDT)discussion.Co-relations of different ACT decision groups and routine clinic-pathologic factors were evaluated in the univariable analysis.Logistic regression was applied to determine independent predictive factors for ACT decision.Patients enrolled were divided in different clinical risk groups by MINDACT criteria and PREDICT.The predictive accuracy of MINDACT criteria and PREDICT was evaluated by area under curve(AUC)of Receiver Operating Characteristics(ROC).All the patients enrolled were included in the model construction,all of whom make up the training set.Clinico-pathological variables were screened for potential predictors of ACT decision,based on multivariate logistic model and a predictive nomogram was generated.The performance of the predictive model was evaluated with respect to discrimination(quantified by AUC)and calibration and internally validated within the training set to test the accuracy of the nomogram.The Bootstrap method was adopted in the internal validation.The model was evaluated in different node status subgroups and was compared with MINDACT criteria and PREDICT.Results(1)Among 504 patients included,255(50.6%)were recommended to receive ACT by MDT while 249(49.4%)were not.(2)Age,tumor grade,p T,p N,molecular subtype and RS were significantly correlated with ACT decision(p <0.05).(3)Age,grade,p T2,p N1,molecular subtype and RS were independent predictors of ACT decision(p <0.05)in mutivariable anaysis.(4)279(55.3%)patients were at low clinical risk(ACT not-recommended group)while 225(44.7%)were at high clinical risk(ACT recommended group)by MINDACT clinial risk classification.232(46.0%)patients were recommended to receive ACT while 272(54.0%)were not by PREDICT.The AUC of these 2 risk evaluation models were 0.687(95%CI:0.640?0.734,p<0.05)and 0.693(95%CI: 0.646?0.739,p<0.05),presenting a low discrimination.(5)Six clinicopathologic variables(age,tumor grade,p T,p N,molecular subtype and RS)were included in our model.In the overall population(n=504),the model showed a high dicrimination(AUC =0.905,95%CI: 0.879?0.931).Nomogram presented significantly greater discrimination beyond MINDACT criteria and PREDICT in all patients(p<0.01).We performed the internal validation with Bootstrap sampling and the corrected AUC remained agreeable(0.897).The calibration of the nomogram was good without apparent over or under-prediction compared with the observed rates(p=0.912).(6)In node-negative subgroup(n=398),AUC of nomogram(0.913,95%CI: 0.886?0.941)was significantly larger than that of MINDACT criteria and PREDICT.In 1?3 node-positive subgroup(n=106),AUC of nomogram was 0.828(95%CI: 0.726?0.929)while the other 2 prognostic models showed no predictive discrimination in this subset of patients.Conclusion(1)Age,grade,tumor burden(p T,p N),molecular subtype and RS were related to ACT decision in intermediate breast cancer patients.(2)MINDACT criteria and PREDICT showed certain predictive ability in this subset of patients,however,its accuracy required to be further improved.(3)A novel nomogram was developed based upon clinical and biological covariates to predict ACT decision in intermediate risk breast cancer patients.Compared with MINDACT criteria and PREDICT,the novel model showed a better predictive accuracy not only in the overall population but also in different node status subgroups.