An interpretable machine learning model integrating 54 test indicators and age: a single-center study on precise discrimination between serous ovarian cancer and ovarian cysts

Le résumé fourni par la source

Abstract Background Serous ovarian cancer (SOC) is a highly malignant subtype of epithelial ovarian tumors, most patients are diagnosed at an advanced stage (III/IV). However, existing single tumor markers have limited diagnostic performance. Therefore, this study aims to develop an interpretable machine learning model by integrating 54 test indicators and age to accurately differentiate between SOC and ovarian cysts. Methods A retrospective analysis was conducted on 648 patients with pathologically confirmed diagnoses (SOC = 155 cases, benign cysts = 493 cases). The i-Research software developed by Roche, based on R language, was used to randomly divide the dataset into a training set and a validation set at a 7:3 ratio. An integrated feature selection approach combining LASSO, Random Forest (RF), and Decision Tree algorithms was applied to identify six key indicators for model construction. Eleven machine learning algorithms, including logistic regression (LR), RF, and XGBoost, were evaluated using eight performance metrics to identify the optimal model. SHAP values were employed to provide multi-level interpretability. Results The LR model was selected as the final model because of its better interpretability and clinical usability on the internal validation set ( n = 195): AUC = 0.979 (95% CI: 0.958–0.994), sensitivity = 89.4%, specificity = 93.9%, accuracy = 92.8%, F1 score = 92.9%, and Kappa coefficient = 80.9%. Age, RAR, CA125, CA153, and MLR were identified as risk factors, whereas RBC count was found to be a protective factor. Conclusion This study innovatively constructs an interpretable LR model offering a practical and reliable tool for the preoperative differential diagnosis of SOC and ovarian cysts. Clinical trial number Not applicable.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.