Integration of serological markers with O-RADS ultrasound risk stratification achieves high diagnostic accuracy for early ovarian cancer: development and validation of an interpretable model.

This study aimed to enhance preoperative diagnostic accuracy for early ovarian cancer (EOC) by constructing and validating an interpretable predictive model integrating serological markers with O-RADS ultrasound risk stratification. A total of 674 patients with adnexal masses (270 EOC, 404 benign) were enrolled and randomly assigned to training (n=438) and validation (n=236) sets. Multivariate logistic regression identified cancer antigen 125 (CA125; P=0.002), human epididymal protein 4 (HE4; P=0.001), carcinoembryonic antigen (CEA; P=0.005), maximum tumor diameter (P=0.005), and Ovarian-Adnexal Reporting and Data System (O-RADS) classification (P<0.001) as independent predictors. Among individual indices, O-RADS performed best (AUC=0.854), followed by CA125 (AUC=0.783), HE4 (AUC=0.744), CEA (AUC=0.717), and maximum tumor diameter (MTD; AUC=0.696). The comprehensive model achieved an AUC of 0.969 (95% CI: 0.953-0.985) in the training set and 0.973 (95% CI: 0.952-0.994) in the validation set, with sensitivity of 0.907/0.897, specificity of 0.914/0.980, and accuracy of 0.911/0.949, significantly outperforming both the marker combination model (AUC=0.918) and individual indices (DeLong test, all P<0.05). Robust calibration was confirmed (training set: Brier score=0.062, Hosmer-Lemeshow P=0.126; validation set: Brier score=0.056, Hosmer-Lemeshow P=0.578), and decision curve analysis demonstrated positive net benefits across a wide threshold probability range. SHAP analysis identified O-RADS classification, CA125, and HE4 as the most influential predictors, with notable interactions among them. Bootstrap resampling and 5-fold cross-validation confirmed model stability. The comprehensive model demonstrated excellent diagnostic performance, robust calibration, and clinical utility, providing a reliable interpretable tool for preoperative EOC risk stratification.