Application Of Machine Learning And Cox Model In Survival Prediction Of Osteosarcoma

Objective:To investigate the prognostic factors affecting the survival of patients with osteosarcoma,construct machine learning and Cox proportional risk regression survival prediction models,and compare the predictive efficacy between the two types of models.Methods:1.Select the clinical case data of osteosarcoma from 2010 to 2015 from the SEER database,determine the relevant variables,clean the data,and include 910 cases for retrospective analysis,including basic information and tumor characteristics and prognostic information.Using X-tile software to convert continuous variables(age,tumor size)into categorical variables,the Random number table method was divided into training set and validation set according to the ratio of 7:3.Kaplan-Meier method and Log-rank test were used to analyze the influence of various variables on the survival rate of patients with osteosarcoma,and the survival curve was drawn.2.The training set used univariate and multivariate Cox regression analysis to determine independent prognostic factors(P<0.05)and constructed Cox proportional hazard regression model(Cox-PH model)for the 1-year,3-year,and 5-year survival of patients with osteosarcoma.The Nomogram visualizes the model,and the area under the curve(AUC)and calibration curve(Calibration curve)evaluate the prediction performance of the model.Result verification in the validation set.3.The training set constructed two machine learning prediction models of Neural Network(NN)and Random Forest(RF)for the survival of patients with osteosarcoma at 1,3,and 5 years and the AUC value evaluation model Predict the performance and the validation set will verify the results.Compare the AUC value between Cox proportional hazards regression model and two machine learning prediction models.Results:Univariate Cox regression analysis showed age,gender,primary tumor site,surgery,radiotherapy,chemotherapy,marital status,AJCC7 clinical-stage,T stage,N stage,M stage,bone metastasis,brain metastasis,liver metastasis,lung metastasis 15 factors of tumor size are significantly related to the survival and prognosis of patients(P<0.05).Multivariate Cox regression analysis showed that age,gender,primary tumor site,type of surgery,AJCC7 clinical stage,and tumor size were independent risk factors affecting the survival and prognosis of patients with osteosarcoma(P<0.05).In the training set and the validation set,the Cox model predicted 1,3,and5-year survival AUC values were 0.907,0.821,0.79 and 0.864,0.812 and 0.743,respectively;the best cut-off values were 0.065,0.371,0.038 and 0.095,0.18,0.039,respectively.The calibration curves of the training set and the validation set are the same.The Cox-PH model overestimates the risk in predicting the 1-year survival of patients with osteosarcoma,and the calibration curve fit is located at the upper left of the reference line.The prognostic prediction of 3,5 years survival and the actual outcome have a better fit.In the training set and the validation set,the Neural Network model predicts the AUC values for the 1,3,and 5-year survival of osteosarcoma patients are 0.896,0.830,0.80 and 0.809,0.794,0.743,respectively;In the training set,Random Forest predicts osteosarcoma patients the OOB error rates for 1,3,and 5 years of survival are: 8.28%,19.38%,and 19.53%;In the validation set,the Random Forest model predicts AUC for 1,3,and 5-year survival of 0.884 and 0.767,0.731,respectively.Comparing the AUC values of the three models,the Cox-PH model,NN model,and RF model all predicted the survival of patients with osteosarcoma better,and the former two models were better than the random forest model in predicting the survival of patients with osteosarcoma at 3 and 5years.Conclusions:1.Age,gender,AJCC7 th clinical-stage,primary tumor location,surgery,tumor size are significantly related to the survival and prognosis of patients with osteosarcoma.2.Cox-PH model,NN model,and RF model all better predicted the survival of patients with osteosarcoma,and the former two models were better than the RF model in predicting medium and long term survival.