Development And Validation Of A Predicting Model For Prognosis In Early-stage Cervical Cancer:A Support Vector Machine-based Approach

ObjectiveThis study retrospectively analyzed the clinical data of patients with stage ? B-? A cervical cancer in our hospital,aiming to valuates the impact of clinical staging in the treatment and outcome of patients and the importance of the risk factors in the Sedlis criteria for early-stage cervical cancer.And established a new model based on support vector machine(SVM)analysis to better predict the clinical outcome of patients with early cervical cancer.MethodsWe performed a retrospective study of all patients diagnosed as cervical cancer with FIGO stage ? B1-?A2,and underwent radical hysterectomy(RH)and pelvic lymphadenectomy,at Qilu Hospital of Shandong University from January 2005 to December 2016.Clinical and pathologic data were collected in detail.Patients were followed-up by telephone for the clinical outcome(including postoperative recurrence and survival).The follow-up was ended on September 2017.The first part:SPSS19.0 software was used to statistical analysis.The chi-square test was performed for univariate analysis of the factors predicting lymph node metastasis(LNM),the binary Logistic regression model was used for multivariate analysis.Kaplan-Meier method was used for survival analysis and the survival curve was drawn.Log-rank test was used to compare the difference in survival.We used Cox proportional risk model for multivariate analysis of factors affecting the prognosis of patients.In the second part,the prediction model was established according to the existing medical knowledge,univariate analysis and multivariate analysis.Establishment of a model using support vector machine to verify the importance of clinical staging of cervical cancer and the importance of medium-risk factors in Sedlis criteria.The leave-out cross validation method was used to evaluate the model.ResultsThe first partTotally,647 patients were included in our study.Median age at diagnosis was 45 years(range 21-79).One hundred and seventy-nine patients(27.7%)were postmenopausal.Only 20(3.1%)patients had no childbirth.More than half of patients had FIGO stage IB1.Squamous cell carcinoma was the dominating histological type with 532 patients(82.2%).Few patients had surgical margins invasion or parametrium involvement(PI)·All patients underwent radical hysterectomy(RH)with pelvic lymphadenectomy,and 50 underwent para-aortic lymph node sampling or dissection.One hundred and fifty-three patients(23.7%)had positive pelvic lymph nodes.Among those patients,2.0%of them had both pelvic and para-aortic lymph nodes(LNs)metastasis,98.0%had only pelvic LNs metastasis,and no one had isolated para-aortic positive lymph nodes.The incidence of pelvic lymph node metastasis(PLNM)in patients with stage? B1 stage ? B2,stage II Al and stage II A2 were 18.6%,35.3%,29.7%and 31.0%,respectively.An univariate analysis suggested that clinical stage,PI,LVSI,tumor size and depth of cervical stromal invasion(DSI)were significantly related to PLNM.However,age,menstruation,delivery,grade of tumor,histological type,status of resection margin and the number of lymph nodes removed had no significant impact on lymph node metastasis(p>0.05).Multivariate analysis revealed that LVSI(p<0.001),tumor size(p=0.044)and DSI(p<0.001)were significantly and independently associated with LNM.The median duration of the follow-up was 29 months(range,6-145 months).During our follow-up,68 recurrences and 44 deaths were identified.Clinical stage,the status of lymph node,PI,tumor size and DSI played an important role on recurrence and survival.Meanwhile,menopause,parity,histology,grading of the tumor,surgical margins,LVSI,the number of lymph node removed and adjuvant therapy after surgery were not statistically significant with disease free survival(DFS)and overall survival(OS).The status of pelvic lymph node(p<0.001),PI(p=0.032)and DSI(p=0.039)were found to be independent indicator for recurrence.And the status of pelvic lymph node(p<0.001),PI(p=0.010)were also found to be independent indicator for survival.The second partBased on the prognostic risk factors(clinical stage,lymph node,parametrium,interstitial infiltration depth and tumor size)obtained in the first part,the SVM-based prediction model showed accuracy of combination including clinical staging was higher than the combination without considering clinical staging(P<0.05).The good prediction accuracies can be achieved,when the two independent factors of parametrium and lymph nodes composed a combination.Replacing the factors in the Sedlis Criteria showed incremental prognostic value.Among the thirteen possible combinations of the five variables,Eight predictive combinations for survival were better than Sedlis standard,and ten predictive combinations for recurrence were better than Sedlis standard.Our results indicated that the model of tumor size,histology,and degree of differentiation was a better predictor of recurrence and survival than the model of intermediate-risk factors in Sedlis criteria(69.5%± 11.5%vs.50.5%±6.4%;68.9±8.9%vs.51.1±4.0%).Conclusion1.LVSI,tumor size and DSI are independent risk factors for LNM of cervical cancer.Furthermore,extensive lymphadenectomy could not improve a significant influence on survival in patients regardless of nodal metastasis status.2.The status of lymph node,PI and DSI were related to the DFS.Positive lymph node and PI were significantly and independently associated with OS.3.More and more evidences support the surgical-pathologic staging of cervical cancer,and there is a better combination of risk factors(tumor size,histology,and degree of differentiation)to predict the clinical outcome of cervical cancer,4.The prognosis model of early cervical cancer was successfully established based on SVM.