Artificial intelligence for prediction of endometrial intraepithelial neoplasia and endometrial cancer risks in pre- and postmenopausal women

Abstract

The current approach to endometrial cancer screening requires that all patients be able to recognize symptoms, report them, and carry out appropriate interventions. The current approach to endometrial cancer screening could become a problem in the future, especially for Black women and women from minority groups, and could lead to disparities in receiving proper care. Moreover, there is a lack of literature on artificial intelligence in the prediction and diagnosis of endometrial intraepithelial neoplasia and endometrial cancer. This study analyzed different artificial intelligence methods to help in clinical decision-making and the prediction of endometrial intraepithelial neoplasia and endometrial cancer risks in pre- and postmenopausal women. This study aimed to investigate whether artificial intelligence may help to overcome the challenges that statistical and diagnostic tests could not. This study included 564 patients. The features that were collected included age, menopause status, premenopausal abnormal bleeding and postmenopausal bleeding, obesity, hypertension, diabetes mellitus, smoking, endometrial thickness, and history of breast cancer. Endometrial sampling was performed on all women with postmenopausal bleeding and asymptomatic postmenopausal women with an endometrial thickness of at least 3 mm. Endometrial biopsy was performed on premenopausal women with abnormal uterine bleeding and asymptomatic premenopausal women with suspected endometrial lesions. Python was used to model machine learning algorithms. Random forest, logistic regression, multilayer perceptron, Catboost, Xgboost, and Naive Bayes methods were used for classification. The synthetic minority oversampling technique was used to correct the class imbalance in the training sets. In addition, tuning and boosting were used to increase the performance of the models with a 5-fold cross-validation approach using a training set. Accuracy, sensitivity, specificity, positive predictive value, and F1 score were calculated. The prevalence of endometrial or preuterine cancer was 7.9%. Data from 451 patients were randomly assigned to the training group, and data from another 113 patients were used for internal validation. Of note, 3 of 9 features were selected by the Boruta algorithm for use in the final modeling. Age, body mass index, and endometrial thickness were all associated with a high risk of developing precancerous and cancerous diseases, after fine-tuning for the multilayer computer to have the highest area below the receiver operating characteristic curve (area under the curve, 0.938) to predict a precancerous disease. The accuracy was 0.94 for predicting a precancerous disease. Precision, recall, and F1 scores for the test group were 0.71, 0.50, and 0.59, respectively. Our study found that artificial intelligence can be used to identify women at risk of endometrial intraepithelial neoplasia and endometrial cancer. The model is not contingent on menopausal status or symptoms. This may be an advantage over the traditional methodology because many women, especially Black women and women from minority groups, could not recognize them. We have proposed to include patients to provide age and body mass index, and measurement of endometrial thickness by either sonography or artificial intelligence may help improve healthcare for women in rural or minority communities.