Random Survival Forest Model for Predicting Recurrence of Postoperative Ovarian Endometrioma

Objective: To develop and compare a random survival forest (RSF) model and a Cox regression model—for predicting postoperative recurrence in patients with ovarian endometrioma. Methods: The primary outcomes of the study were the recurrence of ovarian endometrioma and the time to recurrence. Recurrence-related clinical variables were assessed using two methods: Cox regression analysis and the random survival forest algorithm. The performance of the prediction models was evaluated using time dependent Receiver Operating Characteristic (time-dependent ROC) curve, calibration curve, and Decision Curve Analysis (DCA) curve. Results: 109 patients experienced recurrence of postoperative ovarian endometrioma, while 469 patients did not experience recurrence. The multivariate Cox regression model identified four independent variables. The RSF model was constructed using 12 important clinical variables. In the training set, the Cox model presented a lower C-index compared to the RSF model (0.699 vs. 0.987). However, in the testing set, the Cox model had a slightly higher C-index compared to the RSF model (0.771 vs. 0.763). The classification accuracy of the RSF model and the traditional Cox regression model was evaluated using two performance metrics: NRI and IDI. The NRI at 3 years and 5 years were 0.728 and 0.775 respectively. The IDI at 3 years and 5 years were 0.599 and 0.603 respectively. Conclusion: The predictive performance of the RSF model was slightly better than that of the traditional Cox model for predicting postoperative recurrence in patients with ovarian endometrioma. Keywords: ovarian endometrioma, random survival forest, cox model, prognosis