Comparison of Machine Learning Models for Long-Term Recurrence of Endometriosis Treated by Laparoscopy Combined With GnRHa

Background: To construct and compare the predictive efficacy of the random forest (RF) model and gradient boosting machine (GBM) model for long-term recurrence of endometriosis (EMs) treated by laparoscopy combined with gonadotropin releasing hormone agonist (GnRHa). Methods: A total of 254 patients with EMs who underwent laparoscopy combined with GnRHa in The First Affiliated Hospital, College of Medicine, Zhejiang University from July 2022 to December 2023 were retrospectively collected. All patients were followed up for 1 year, and the long-term number of recurrences was recorded. The corresponding influencing factors were obtained by single factor analysis, and the risk prediction model of the long-term recurrence of sub-EMs was constructed based on RF and GBM models. At the same time, the receiver operating characteristic (ROC) curve and calibration were used to compare the predictive value of the model constructed by the two algorithm models for long-term recurrence of EMs. Results: Univariate analysis showed that the course of disease, preoperative dysmenorrhea history, preoperative uterine cavity operation history, tender posterior fornix and revised American Fertility Society (r-AFS) stage were the influencing factors of postoperative recurrence in patients with EMs treated by laparoscopy combined with GnRHa (p < 0.05). Based on univariate analysis, RF and GBM models were constructed. The order of importance of the predictors of laparoscopy combined with GnRHa in the treatment of EMs was r-AFS staging, course of disease, tender posterior fornix, history of intrauterine operations and history of preoperative contraception. The ROC curve results of the RF model showed that the area under curve (AUC) of the model in the training set was 0.902 (95% CI: 0.857–0.947), and the sensitivity and specificity were 100.00% and 63.50%, respectively. The AUC in the validation set was 0.859 (95% CI: 0.741–0.976), and the sensitivity and specificity were 69.20% and 92.90%, respectively. The results of the ROC curve of the GBM model showed that the AUC of the GBM model in the training set was 0.851 (95% CI: 0.781–0.920), and the sensitivity and specificity were 89.20% and 68.20%, respectively. The AUC in the validation set was 0.852 (95% CI: 0.713–0.990), and the sensitivity and specificity were 76.90% and 87.50%, respectively. The calibration curve shows that the prediction probabilities of the RF model and the GBM model are highly consistent with the actual prediction in both the training set and the validation set. The results of Delong test showed that the training set AUC of RF model was better than that of GBM model, and the difference was statistically significant (Z = 2.838, p = 0.005). There was no significant difference in the validation set AUC between the RF model and the GBM model (Z = –0.239, p = 0.811). Conclusion: r-AFS staging, course of disease, tender posterior fornix, history of intrauterine operations and history of preoperative laparoscopy are the influencing factors in the long-term recurrence of EMs treated by laparoscopy combined with GnRHa. RF and GBM models can effectively predict the recurrence of such patients after treatment.