Development of a prediction model for recurrence in patients with malignant ovarian germ cell tumors undergoing fertility-sparing surgery: A multicenter retrospective study.

OBJECTIVE: To investigate the clinical characteristics associated with recurrence in malignant ovarian germ cell tumors (MOGCT) patients who received fertility-sparing surgery (FSS) and to develop a machine learning-based risk prediction model to support treatment decision-making for these specific MOGCT patients. METHODS: This study enrolled MOGCT patients who received FSS at four university-teaching hospitals in China between 2005 and 2020. The primary outcome was disease-free survival (DFS). Four machine learning algorithms-Cox Proportional Hazards Model (Cox), Random Survival Forest (RSF), Cox Proportional Hazards Model Boosting (CoxBoost), and eXtreme Gradient Boosting (XGBoost)-were used to develop the risk prediction models for DFS. The Cox Proportional Hazards Model was selected to construct the final risk prediction model, and visualized by a nomogram. The prediction model was internally validated with bootstrapping, and its performance was estimated by the receiver operating characteristic (ROC) curve, calibration curves, and decision curve analysis (DCA). RESULTS: A total of 264 MOGCT patients receiving FSS were enrolled in this study. Tumor size (hazard ratio [HR] ≥15 cm vs. <15 cm = 6.15, 95% confidence interval [CI]: 1.29-29.34; P = 0.023), tumor laterality (HR Right-sided vs. Left-sided = 0.13, 95% CI: 0.03-0.57; P = 0.007), and International Federation of Gynecology and Obstetrics (FIGO) stage (2014) (HR Stage II-III vs. Stage I = 7.57, 95% CI: 2.50-22.96; P < 0.001) were associated with DFS and subsequently selected for the establishment of prediction models. Among the four machine-learning models, the Cox algorithm-based model showed relatively high predictive accuracy and was thus chosen for the establishment of the final model. The areas under the ROC curves of the prediction model were 0.838 for 3-year DFS and 0.851 for 5-year DFS. Meanwhile, the model demonstrated excellent discriminative ability, calibration, and favorable clinical utility. CONCLUSION: This multicenter study developed a Cox algorithm-based prediction model that may help assess the individual recurrence risk in MOGCT patients who received FSS.