An assisted diagnostic and prognostic model for endometrial cancer using 36 serological markers and clinical variables from 562 patients.

Endometrial carcinoma (EC) has demonstrated a concerning epidemiological trajectory. Current evaluation systems for EC are limited to postoperative analysis, necessitating the development of a preoperative risk stratification model. Researchers aimed to create a machine learning-based approach using multiple serological markers and clinical variables to predict EC characteristics. A dataset of 508 EC patients and 54 patients with endometrial atypical hyperplasia (EAH) was utilized. Seven supervised machine learning classifiers were evaluated, with the Random Forest classifier being the most effective. The use of advanced computational models that integrate diverse biological markers and clinical factors has demonstrated superior predictive accuracy for various outcomes in endometrial cancer compared to traditional single-variable regression methods. Among these models, the Random Forest classifier achieved high area under the curve (AUC) values (0.81-0.94) and exhibited strong predictive performance. Other models, including Gradient Boosting Regression, Support Vector Machine, Logistic Regression, Gaussian Naive Bayes, Neural Networks, and Elastic Net, were also evaluated. The key predictors varied depending on the specific outcomes. The machine learning approach utilizing combined serological markers and clinical variables allows effective non-invasive preoperative prediction of key EC characteristics, including diagnosis, staging, metastasis risk, and prognosis, with the Random Forest classifier proving most effective. Specifically, human epididymis protein 4 (HE4) is pivotal for risk stratification (including stages and Mayo criteria), while carbohydrate antigen 125 (CA125) excels in detecting lymph node invasion. This serum-based strategy complements invasive methods, enhancing early EC management and enabling personalized treatment.