Construction and validation of a preoperative malignancy risk prediction model for ovarian-adnexal masses based on clinical and ultrasonographic features.

Authors

Abstract

To develop and validate a simple yet effective nomogram based on clinical and ultrasound factors for predicting the malignant risk of ovarian-adnexal masses and to evaluate its diagnostic performance and clinical utility. This retrospective study enrolled 930 patients with pathologically confirmed ovarian-adnexal masses (involving a total of 1,380 masses) from January 2020 to December 2025. Clinical data and ultrasound images of the patients were extracted from the hospital information system. Two gynecologic sonographers with more than 10 years of clinical experience, who were blinded to the pathological results, independently reviewed all ultrasound images, and all discrepancies in image interpretation were resolved through consensus discussion. To ensure the rigor of validation, all masses were randomly assigned to a training set (966 masses, numbered 1 to 966) and a validation set (414 masses) at an approximate ratio of 7:3 using a random function. In the training set, univariate analysis was first performed to screen for variables with statistical significance (<i>p</i> < 0.05). Multivariate logistic regression analyses was applied to explore the associations between each variable and the benign or malignant nature of ovarian-adnexal masses. Lasso regression analysis was conducted for the standardized processing and feature selection of the statistically significant variables. A nomogram for predicting the malignant risk of ovarian-adnexal masses was constructed based on the feature screening results of Lasso regression. To further validate the diagnostic efficacy of the model, the predictive indicators screened above were adopted to build classification models in the training set using four machine learning algorithms, namely Logistic Regression, Random Forest, XGBoost and LightGBM, and the diagnostic efficacies of these models were compared in the validation set. The area under the receiver operating characteristic curve, accuracy, sensitivity and specificity were used as the primary indicators to evaluate model performance. Internal validation of the models was completed via the bootstrap method with 1,000 resamples. Calibration curves were plotted to verify the consistency between the predicted probabilities of the nomogram and the optimal machine learning model and the actual observed results. Clinical decision curves were drawn to analyze the net benefit rate at different probability thresholds, thereby assessing the clinical practical value of the models. Finally, risk score-predicted probability calibration plots and decision plots were used for further evaluation of the prediction models to clarify their stability and clinical application value. LASSO regression analysis identified 8 statistically significant independent risk factors for the malignant risk of ovarian-adnexal masses, which were categorized into four groups: clinical characteristics (menopausal status), ultrasound features (internal echogenicity, internal septations, blood flow signals), serum tumor markers (CA125, HE4), and routine blood parameters (platelet count [PLT], platelet-to-hemoglobin ratio [PHR]). A visual nomogram prediction model was successfully constructed based on these factors. Four machine learning algorithms were separately applied to the training and validation sets for comparative analysis, and the results demonstrated that all models exhibited excellent discriminatory performance: the area under the receiver operating characteristic (ROC) curve of LASSO regression analysis reached 0.97; the Logistic Regression model achieved an AUC of 0.95 in the validation set; the Random Forest model yielded AUCs of 0.95 and 0.94 in the training set and validation set, respectively; both the XGBoost and LightGBM models attained an AUC of 0.96 in both the training and validation sets. Calibration curve analysis showed a high degree of coincidence between the model-predicted curves and the ideal curve, indicating good consistency between the predicted probabilities and actual observations. Decision curve analysis revealed that the model generated significant clinical net benefits over a wide range of probability thresholds, confirming its important value in assisting clinical decision-making. Ultimately, through risk score-predicted probability calibration and decision visualization analysis, the model constructed in this study can provide a quantitative basis for formulating individualized treatment plans for patients with ovarian-adnexal masses. A nomogram developed to predict the malignant risk of ovarian-adnexal masses using clinical and ultrasound factors showed promising diagnostic performance. This model is framed as a hypothesis-generating tool and is intended to assist in the clinical differentiation of benign and malignant ovarian-adnexal masses, with its utility yet to be confirmed by further external validation studies.

Tags

Topics