Development Of Machine Learning Models To Predict Platinum Resistance Recurrence Of Epithelial Ovarian Cancer

Objectives: To develop models for predicting platinum resistance recurrence of epithelial ovarian cancer based on routine clinical and laboratory data by machine learning method.Methods: This was a retrospective cohort study.Firstly,we assembled 1392 epithelial ovarian cancer patients who received platinum-based chemotherapy in Yunnan Cancer Hospital from January 1,2012,to June 30,2022.Patients’ clinico-pathologic charaeristics,routine laboratory data,operation related information,details of chemotherapy,and survival outcomes were collected.According to the relapse status,patients were divided into platinum-sensitive arm and platinum-resistant arm.Secondly,to select suitable variables,29 factors that might affect the recurrence of platinum resistance were screened by two different variable screening methods,lasso regression and univariate and multivariate logistic regression,respectively.Thirdly,based on the variables screened in the above two ways,five machine learning algorithms were used to construct prediction models,including KNN,SVM,RF,DTA,XGBoost,and compared with the traditional model fitting method LR.A10-fold cross-validation method was used for internal validation and to compare the performance metrics among the models.The model’s performance indicators included AUC,sensitivity,specificity,and average accuracy.Finally,the best models were visualized by web calculator,nomogram,decision tree diagram,and variable importance,etc.Results:1.Variable Selection1.1 Lasso regressionIn the lasso regression,seven variables were selected,including neoadjuvant chemotherapy cycle,FIGO stage,involvement of appendix,involvement of omentum,diaphragmatic dome invasion,residual tumor size,and first-line chemotherapy cycle.1.2 Logistic regressionThe logistic regression analysis showed that LDH,FIGO stage,platelet count,supraclavicular lymph node metastasis,initial treatment strategy,residual tumor size,platinum(carboplatin/others),and chemotherapy cycles were the independent influencing factors of platinum-resistant recurrence.2.Model Development and Validation7 factors selected by Lasso and 8 factors selected by logistic were used respectively to devolope models by five machine learning algorithms(KNN,SVM,RF,DTA and XGBoost)and LR traditional methods.The models were verified by ten-fold cross-validation internally.The LR model which based on the lasso regression performed best,with an AUC of 0.738,a sensitivity of 0.541,a specificity of 0.836,an average accuracy of 0.796,and the prefered cut-off value is 0.154.Meantime the XGBoost model based on the logistic regression showed the best performance with an AUC of 0.784,the sensitivity of 0.735,a specificity of 0.713,an average accuracy of 0.804,and the prefered cut-off value is 0.240.At last,we visualized these two models by nomogram,web calculator and the importance of each variable,respectively.Conclusions:We successfully developed prediction models for platinum-resistant epithelial ovarian cancer recurrence based on routine clinical and laboratory data.Among the models constructed in these six algorithms,the LR model performs the best if the model is built based on 7 variables selected by lasso regression;if the model is built based on 8 variables screened by logistic regression,the XGBoost model performs the best,and both of them have high AUC value and average accuracy in internal validation,which can be used in clinical practice.However,due to the difference in time and space,the influencing factors may change,and the model needs to evolve continuously.We have provided an idea of model evolution.