Development And Validation Of Prediction Models For Coexisting Adnexal Malignancy In Patients With Endometrial Cancer Using Machine Learning Algorithms

Background:The incidence of endometrial cancer(EC)is the sixth most common malignancy among women globally and tops the list of gynecological malignancies among women in the United States.Cancer data showed that the incidence of EC has increased over the past 20 years.Another noteworthy phenomenon is that the incidence of early-stage low-grade EC has been increasing in young patients in recent years.Although more and more retrospective studies have confirmed the safety of ovarian preservation in EC patients,and the guidelines have successively proposed corresponding indications for the needs of ovarian preservation in EC patients,the ovarian preservation rate of these young patients is still low.Due to ethical constraints,it is difficult to design and conduct prospective studies to verify and explore the safety and indications of ovarian preservation in EC patients.Currently,retrospective studies with small samples have been conducted to establish clinical prediction models to assess the risk of patients of EC with adnexal malignancies.However,due to limited preoperative available clinical characteristics of patients and small sample size,the prediction efficacy and clinical applicability of the models are not ideal.Objective:Based on machine learning algorithm and combined with the preoperative clinicopathological characteristics to develop the adnexal malignancy risk prediction model of EC patients,and the model was validated and visualized to provide references for the preoperative diagnosis and treatment decisions during clinical management of EC.Methods:Patients surgically treated for endometrial cancer from medical centers including the Qilu Hospital of Shandong University between 2000 and 2019 were enrolled.Preoperative clinical and histopathological features of patients included the age at diagnosis,previous history of bilateral oophorectomy for benign ovarian or fallopian tube disease(such as ovarian torsion,ovarian borderline tumor,etc),personal history of cancer(including one or more malignant tumors such as stomach cancer,colorectal cancer,pancreatic cancer,breast cancer,lung cancer,etc),and family history of cancer(immediate family members with endometrial cancer,ovarian cancer,stomach cancer,colorectal cancer,pancreatic cancer,breast cancer,etc),underlying diseases such as hypertension and diabetes,pathological type(endometrioid adenocarcinoma,non-endometrioid adenocarcinoma)of preoperative endometrial biopsy,histological grade(grade 1,grade 2/3),and carbohydrate antigen-125(CA-125)level(≤35 IU/L,>35 IU/L),neutrophil/lymphocyte ratio(NLR)(≤2.4,>2.4).Preoperative imaging or intraoperative evaluation information included myometrial invasion depth(<1/2,≥1/2),susceptible adnexal abnormalities(such as cyst-solid changes and abnormal blood flow signals under ultrasound,abnormal density or signal changes suggested by CT or MRI,abnormal changes detected during surgery),susceptible lymph node involvement and extrauterine metastasis.Firstly,the age distribution of EC patients with early stage,low grade as well as with adnexal malignancy was analysed.The EC cohort was dichotomized according to whether there was coexisting adnexal malignancy and the difference in preoperative clinicopathologic features between the two groups was compared by chi-square test.Then,the total EC patient cohort was randomly divided into the development cohort and the internal validation cohort in a ratio of 7:3,and the preoperative clinicopathological characteristics of patients in the development cohort and the validation cohort were compared.In the development cohort,least absolute shrinkage and selection operator(LASSO)was used for selecting candidate predictors,and machine learning algorithms including logistic regression analysis(LR),support vector machine,decision tree,random forest,AdaBoost,decision tree,naive Bayes,and multilayer perceptron were applied to develop the model.Model evaluation included area under the receiver operating characteristic curve value(AUC),the relative sensitivity,specificity,positive predictive value,negative predictive value,and overall accuracy.Then the models were tested in the validation cohort.The calibration curve and decision curve were also evaluated.And the best performed one was chosen for further visualization to improve clinical interpretability.Results:A total of 3180 patients with EC were included in this study.The proportion of young patients with EC aged 45 years and below accounted for 14%of patients with early stage and low grade.6.7%(212/3180)of EC patients were pathologically confirmed with coexisting adnexal malignancy,and one-fifth of them were younger than 45 years of age.The data of 2226 female patients were divided into development cohort for model training.Data from 954 patients were divided into validation cohort for internal test and evaluation of the model.Among the data in the development group,10 predictors were selected for model construction using LASSO feature screening,including personal cancer history of cancer,age,CA-125,NLR,histological grade,cervical involvement,depth of myometrium invasion in preoperative imaging or intraoperative exploration,susceptible adnexal malignancy,lymph node involvement and extrauterine involvement in preoperative imaging or intraoperative exploration.After the successful construction of the model,the model was comprehensively evaluated and the validation group data was used for internal tested.LR model outperformed the other algorithms,with an AUC of 80%,sensitivity of 61%,negative predictive value of 97%,accuracy of 85%in the validation cohort.Further evaluation showed that the model was well-calibrated and showed satisfied clinical utility.Based on LR model,dynamic nomogram were conducted and the corresponding webpage calculator was generated for clinical interpretability.Patients were divided into low,intermediate and high risk groups based on the model scores,which was convenient for clinical use.Conclusions:1.Based on multi-center data of EC patients,our model provides individualized estimates of risk of adnexal malignancy in patients with EC using machine learning algorithms,with all model inputs available at the time before the surgery.2.LR model is superior to other machine learning algorithms.Visualization of LR model can be realized by nomogram and online calculator,which can assist clinicians to quickly assess the risk of EC patients with adnexal malignant tumors,classify the risk levels,and contribute to clinical diagnosis and treatment.3.A prospective feasibility study will be needed prior to implementation in the clinic.