Development And Validation Of Personalized Survival Prediction Model For Adenosarcoma Of Uterus Based On Deep Learning Method

Background:Adenosarcoma is a rare malignant tumor in female genital tract,incidence rate is about 5%of uterine sarcoma.Because of its extremely low incidence rate,only a few cohort studies have provided some risk factors that affect prognosis.The survival,epidemiology and end results(SEER)database is a population-based cancer dataset with a huge amount of cases.Using SEER database for large sample study can greatly improve the accuracy of survival prediction for adenosarcoma patients.In addition,the rapid development of artificial intelligence has brought new possibilities for clinical work.The deep learning method can process the input raw data by machine,and use the multi-layer neurons in the neural network to explore automatically.This study aims to develop a personalized survival prediction model for patients with adenosarcoma by using deep learning technology.Methods:1.Data collection and processing:Extract the clinical information of adenosarcoma patients from SEER database,and exclude the cases with multiple tumors and incomplete follow-up data.Finally,15 variables including age and race were selected for further analysis.2.Establishment and evaluation of deep learning model:Based on neural multi task logistic regression model(N-MTLR),deep learning neural network is established.The data set is randomly divided into training set,verification set and test set.The data of training set and verification set are used for model training,and the prediction ability is verified by independent test set.At the same time,the prediction effectiveness of survival prediction deep learning model and Cox proportional hazards model on the same data set was compared.The main indicators to evaluate the predictive effectiveness of the models are the consistency index(c-index)and the comprehensive Brier score.3.Drawing of individual survival prediction curve:K-M(Kaplan-Meier)curve was drawn according to traditional adenosarcoma staging system.The deep learning model grouped adenosarcoma patients according to the survival and prognosis risk score.The individual survival curve was drawn to verify the predictive effectiveness of risk score grouping.Results:1.Statistical characteristics of patients:A total of 797 patients with adenosarcoma were enrolled in the SEER database from 1973 to 2014.The data were divided into training set(n=519,65%),verification set(n=143,18%)and test set(n=143,18%).After correlation analysis,the 15 variables analysed in this study included age,year of diagnosis,race,Hispanic or not,marital status,stage,tumor diameter,differentiation grade,lymph node metastasis or not,number of resected lymph nodes,number of positive lymph nodes,distant metastasis or not,tumor invasion scope,operation or not and operation type.2.Survival prediction performance of Cox proportional hazards model:The c-index of Cox proportional hazards model in survival prediction is 0.726,and the comprehensive Brier score is 0.17.The median absolute error was 1.615 and the average absolute error was 2.223.When drawing the survival curve,some part of the predicted survival curve are drawn outside the confidence interval of the actual survival curve,and there is a large absolute error.3.Survival prediction performance of deep learning model:The c-index of survival prediction deep learning model in the external test set is 0.774,and the comprehensive Brier score is 0.14.The median absolute error was 2.621 and the average absolute error was 1.989.When drawing the survival curve,almost all the regions that predict the survival curve appear in the confidence interval of the actual survival curve.4.Individual survival prediction for adenosarcoma patients:The K-M curve of patients with traditional staging system showed that there was no significant difference in the survival rate of patients with stage Ⅱ,Ⅲ and Ⅳ.The patients were divided into three groups based on the prognosis by deep learning model.The individual survival curve showed that the prognosis of the three groups was significantly different.Conclusion1.Deep learning technology is an effective method to predict the survival of patients with adenosarcoma;2.Deep learning model has better prediction performance than Cox proportional hazards model;3.Personalized survival prediction can be realized through deep learning model.