A Deep Learning Model For Individualized Survival Prediction Of Vulvar Cancer Was Established Based On SEER Database

Objective:Vulvar cancer is a rare malignant tumor in the female reproductive system.Due to the mild and non-specific symptoms in the early stage,some patients have missed the best treatment opportunity when diagnosed.It was found in clinical work that the survival and prognosis outcomes of patients with vulvar cancer were not the same during the same period,and more risk factors affecting the prognosis needed further exploration and research.Due to the low incidence of vulvar cancer,the SEER database of the United States is introduced for retrospective analysis,and the deep learning survival prediction model of vulvar cancer is attempted to be constructed to provide theoretical basis for making accurate diagnosis and treatment decisions.Method:Based on the SEER database,patients with vulvar cancer who met the inclusion and exclusion criteria from 2004 to 2015 were selected as study subjects.According to clinical experience and variable matrix analysis,16 variables were selected as risk factors for analysis.Based on the N-MTLR model,a deep learning survival prediction model was developed,and four control models were established for comparison.Based on the risk scoring system created by the deep learning model,patients with vulvar cancer were re-grouped,and compared with the traditional FIGO staging in terms of survival prediction performance,and a new individualized survival prediction method was created based on the risk score.In order to verify the accuracy of the model,the clinical data of 85 patients withvulvarcancer in Qilu Hospital from 2005 to 2018 were collected as an external validation set.Results:Based on the SEER database,the clinical and prognostic information of vulvar cancer patients from 2004 to 2015 were selected,and the risk factors affecting the prognosis were screened.After data cleaning,a sample database containing 2240 patients’ clinical data was formed.They were randomly divided into training set and test set,and the ratio was 0.5:0.5.After 2000 iterations of the deep survival learning model,the loss value decreased from 30000 to 16710.The consistency index of the survival prediction model was 0.793,and the comprehensive Brier score was 0.13,which was better than that of the control model.The deep learning model created a risk scoring system,and the patients were re-divided into three groups of low,medium and high risk.K-M curves were drawn for the three groups of patients,and there were significant differences in survival prognosis among the three groups(P<0.001).As a comparison,according to the traditional FIGO staging,the K-M curves of the four groups of patients were drawn,and the results showed that there was no significant difference in the prognosis of patients with stage Ⅲ and Ⅳ,so it was impossible to formulate individualized diagnosis and treatment plans for stage Ⅲ and Ⅳ.In order to verify the accuracy of risk grouping,one patient was randomly selected from different risk score groups,and the survival curve was drawn at the same time.After repeated verification for 4 times,the survival prognosis was significantly different(P<0.05).A total of 85 patients with vulvar cancer in Qilu Hospital were collected.The KM curve and ROC curve were drawn according to the traditional FIGO stage and the risk score of the deep learning model.The KM curve of traditional FIGO staging showed that the prognosis of patients in stage Ⅰ was significantly different from that in the other three stages,but there was no significant difference between patients in II,Ⅲ and Ⅳ groups,and the area under the ROC curve was only 0.5749.The KM curve of the risk score grouping of the deep learning model showed that the prognosisofpatients in the low-risk group was significantly different from the other two groups,and the follow-up time of the medium and high-risk groups was prolonged,and the prognosis difference between the two groups gradually appeared.The area of the ROC curve was 0.6379,which was slightly superior to the traditional FIGO staging.Conclusion:1.The prediction performance of the deep learning model for survival prediction of vulvar cancer is better than that of the traditional prediction model.2.The deep learning model for survival prediction of vulvar cancer can realize the personalized staging survival prediction of vulvar cancer,and the prediction results are reliable.