Development and validation of prediction models for coexisting adnexa malignancy in patients with endometrial cancer using machine learning algorithm: A population-based cohort study.

Abstract

e17617 Background: As the rate of ovarian conservation in endometrial cancer patients declined over the course of a decade, the study was to build and validate prediction models based on machine learning algorithms using preoperative predictors to evaluate the risk of coexisting adnexa malignancy, in order to guide decision-making for patients and clinicians before the surgery. Methods: Patients surgically treated for endometrial cancer from 30 medical center which forming the Chinese Endometrial Carcinoma Consortium (CECC) between 2010 and 2019 were enrolled. These patients were randomly divided at ratio of 7:3 into development cohort and validation cohort. Preoperative clinical and histopathological features were included as predictors, and least absolute shrinkage and selection operator (LASSO) was used for selecting candidate predictors. Machine learning (ML) algorithms including logistic regression (LR) analysis, support vector machine (SVM), random forest (RF), AdaBoost, decision tree (DT), naïve Bayes (NB), and multilayer perceptron (MP) were applied to develop the models. The best final model was selected by integrating area under the receiver operating characteristic curve (AUC) value as well as the relative sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and overall accuracy into account. Results: Among 3108 patients in the study cohort, 212 (6.7%) coexisted with adnexal malignancy. Data from 2226 female patients were divided into the cohort for model development. Ultimately the model included 10 predictors selected by LASSO: personal cancer history, age, CA-125, neutrophil lymphocyte ratio (NLR), tumor grade, myometrial involvement, cervix involvement, lymph node involvement, adnexal involvement, and extrauterine involvement. Among all ML algorithms, LR model outperformed the other algorithms, with an AUC of 80% and an accuracy of 85% in the validation cohort. Further evaluation proved that the model was well-calibrated and showed satisfied clinical utility. Based on LR, an online calculator was created for enhancing the clinical interpretability. Conclusions: This model provides individualized estimates of risk of adnexa malignancy in patients with endometrial cancer, with all model inputs available at the time before the surgery. A prospective feasibility study will be needed prior to implementation in the clinic.