Machine-learning-derived prediction models of recurrence of ovarian endometriosis after laparoscopic surgery

Endometriosis (EMT) described by the existence of endometrial glands and stroma outside the uterus is a chronic, gynecologic condition estimated to affect 190 million women worldwide. [ 1 , 2 ] Approximately 5% to 10% of women of reproductive age are affected by EMT, [ 3 ] which is associated with symptoms such as dysmenorrhea, pelvic pain, dyspareunia, and infertility. [ 2 ] This disease not only severely affects the physical health and quality of life of patients but also causes a huge economic burden on society. Among the various forms of EMT, ovarian endometriosis (OEM) is the most prevalent. [ 4 ] Ectopic endometrial tissue can infiltrate the ovarian cortex and proliferate, resulting in recurrent periodic hemorrhage and the formation of one or more characteristic cystic lesions, a condition known as OEM. Laparoscopic resection of ovarian lesions, also known as conservative laparoscopic surgery, is a widely used therapeutic strategy for OEM. However, owing to the complexity of EMT and limited surgical techniques, focal excision is often unable to completely remove the lesion; therefore, some patients with OEM may experience recurrence after conservative laparoscopic surgery. [ 5 ] The recurrence of the disease is noteworthy, as it may influence the postoperative pregnancy rate among women of childbearing age and lead to the reemergence of pelvic pain, thereby significantly impacting the physical and mental health of patients. Nowadays, artificial intelligence techniques have been implemented to resolve critical challenges in EMT diagnosis, prediction and study. [ 6 ] Machine learning, an emerging discipline of artificial intelligence, can be used to explore the relationships among complex data. Both Random Forest and extreme gradient boosting (XGBoost) are robust nonlinear models that offer the advantages of providing feature importance estimates and typically require less hyperparameter tuning. [ 7 ] The Multilayer Perceptron represents one of the earliest network types applied to bioinformatics, and owing to its simplicity and rapid training capabilities, it continues to be extensively utilized in numerous bio-modeling applications today. [ 7 – 9 ] The Gaussian Process is a probabilistic and nonparametric machine learning method. [ 10 ] In comparison to non-Bayesian machine learning methods, such as support vector machines and artificial neural networks, the Gaussian Process can provide uncertainty measures and confidence intervals. [ 11 ] Therefore, this study systematically reviewed and summarized the relevant data of patients with OEM who underwent laparoscopic conservative surgery in the Gynecology Department of Renmin Hospital of Wuhan University from January 2020 to January 2023. These patients were followed for 2 years to analyze the risk factors for postoperative recurrence in patients with OEM. A predictive model was then constructed using machine learning methods, which provided valuable insights for the clinical assessment and personalized management of the recurrence risk of OEM patients 2 years after laparoscopic conservative surgery.

Conceptualization: Qingzhen Xie. Data curation: Mengru Li, Shujun Wang. Formal analysis: Chaoyi Yang. Funding acquisition: Qingzhen Xie. Investigation: Chaoyi Yang. Methodology: Mengru Li, Shujun Wang. Project administration: Qingzhen Xie. Resources: Aili Tan. Software: Mengru Li, Shujun Wang. Supervision: Aili Tan. Validation: Mengru Li, Shujun Wang. Visualization: Aili Tan. Writing – original draft: Mengru Li. Writing – review & editing: Qingzhen Xie.

Patients hospitalized in the Gynecology Department of Wuhan University Renmin Hospital between January 2020 and January 2023 who underwent laparoscopic conservative surgery and were confirmed to have ovarian EMT by postoperative pathology were selected. Detailed information on patients during hospitalization was collected according to the inclusion and exclusion criteria. Age, family history, ovarian cyst diameter, r-American Fertility Society (r-AFS) staging, and other clinical characteristics were also included. These patients were followed until January 2025, when 338 patients were eventually identified as a research cohort to build and validate the proposed model. The patients were divided into recurrence and non-recurrence groups according to postoperative recurrence. The overview of the workflow is presented in Figure 1 . The surgical procedures were performed by a minimum of 2 associate chief physicians or higher-level practitioners. The surgical process was performed with meticulous attention to detail. All patients included in this study underwent pelvic ultrasound examinations 1 month post-surgery, and the ultrasound findings did not indicate the presence of ovarian endometriotic cysts. The overall flowchart of the study. A patient was considered to have relapsed if they had new lesions with characteristic EMT echogenicity indicated by imaging findings (lesions with a diameter > 2 cm). [ 12 ] A patient was included in the research if she satisfied the following standards: nonpregnant, age ≥ 18 years old, a operative indication for EMT or other benign gynecologic diseases, laparoscopic ovarian cyst removal was performed, high-quality clinical data, including general data, medical history, physical examination, laboratory examination, surgical status, pathological report, the postoperative pathology confirmed that OEM, with a postoperative follow-up time of 2 years. Women who were pregnant or menopausal, women with a history of adnexectomy (including fallopian tube or ovary removal), hysterectomy, rectal excision, operative history of serious pelvic infectious diseases, women with autoimmune diseases, women with coagulation dysfunction, women with the administration of gonadotropin-releasing hormon analog ≥ 3 months before surgery, complicated with severe heart and brain diseases, pulmonary insufficiency, blood system and acute-stage infectious diseases, long-term follow-up was not possible, and serious complications occurred during and after surgery, such as ureter injury and bladder injury, combined with congenital reproductive tract malformation, such as congenital vaginal atresia. The following clinical data were collected for all patients diagnosed with OEM: age (years), preoperative history of dysmenorrhea (yes/no), previous history of EMT (yes/no), previous history of pelvic and uterine surgery (number of operations), preoperative history of pregnancy (number of pregnancies), family history (yes/no), diameter of ovarian cyst (cm), single/bilateral cyst, type of ovarian cyst (I/IIa/IIb/IIc), luteal phase surgery (yes/no), preoperative rupture (yes/no), combined with adenomyosis (yes/no), combined with peritoneal EMT (yes/no), combined with deep infiltrating EMT (DIE) (yes/no), r-AFS stage (I/II/III/IV), preoperative serum carbohydrate antigen 125 (CA125) level (U/mL), anti-Müllerian hormone (AMH) (ng/mL), platelet count, platelet/lymphocyte ratio, systemic immunoinflammatory index (systemic immunoinflammatory index = platelet count × neutrophil count/lymphocyte count), thrombin time (TT), D-dimer, natural pregnancy (yes/no), test-tube pregnancy (yes/no), postoperative abortion (yes/no), postoperative live birth, breastfeeding duration (months). The intraoperative staging method used was EMT staging (r-AFS staging), which was released by the American Society for Reproductive Medicine in 1996. The scores were divided into 4 stages as follows: stage I, score 1 to 5; stage II, score 6 to 15; stage III, score 16 to 40; stage IV, score > 40. [ 13 ] Before building the model, SPSS24 software (International Business Machines Corporation, New York) was used for statistical analysis. Chi-square test was applied to categorical (binary) variables, while Spearman correlation analysis was used for continuous variables. A P value < .05 was considered statistically significant. Statistically significant influence features were selected to construct correlation models. To lessen the bias caused by missing data, factors showing exceeded 20% missing data were excluded during the data collection stage. The multiple imputation technique was used to solve missing data within the remaining variables for the analysis. To improve the model stability, 10-fold cross-validation was executed. In the dataset, 70% of the data were designated for the training set, whereas 30% were allocated to the test set, thereby establishing the training and test sets. Four algorithms were used in model construction: Random Forest, Gaussian process, XGBoost and Multilayer Perceptron (MLP). The patients’ baseline data and postoperative recurrence situation were trained and learned, the recurrence prediction model was established and verified, and the accuracy was trained and calculated in experiments with different combinations of hyperparameters. The area under the receiver operating characteristic curve (AUROC) was the major evaluation indicator for model performance assessment. Various performance indicators, including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, F1 score, and Kappa value, were analyzed. Furthermore, the AUPRC was evaluated. The statistical analyses and modeling procedures were applied using Python (version 3.9) and the significance threshold was set to a bilateral P -value <.05.

A total of 381 patients with OEM were included in this research according to the inclusion criteria. After applying the exclusion criteria, 53 patients were excluded and 338 patients were finally included. A total of 64 patients relapsed within 2 years after surgery, with a recurrence rate of 18.9%. The age range of the included patients was 20 to 51 years old, with a mean age of 33.9 ± 6.97 years and a median age of 34 years. The patient’s general information, medical history-related information, surgical data, preoperative blood results, and postoperative treatment are shown in Tables 1 – 4 . Baseline characteristics of all the patients included. Baseline characteristics of all the patients included. Baseline characteristics of all the patients included. Baseline characteristics of all the patients included. AMH = anti-Müllerian hormone, CA125 = carbohydrate antigen 125. In this study, the Chi-square test was applied when the independent variable was binary, and the Spearman analysis was used for the remaining variables. Through exhaustive analysis of all patients, 16 key variables were found to be significantly associated with postoperative recurrence. These variables included age, preoperative pregnancy, preoperative history of dysmenorrhea, family history, diameter of the ovarian cyst, single/bilateral cyst, type of ovarian cyst, r-AFS stage, combined with peritoneal EMT, combined with DIE, preoperative serum CA125 level, AMH, TT, total duration of medication, postoperative live birth, and breastfeeding duration. Finally, 16 features were determined to build the final model. The clinical features associated with postoperative recurrence are shown in Tables 1 – 4 . Four machine learning models were used to predict postoperative recurrence in all the patients with OEM. The indicators used to assess the predictive ability of the model are listed in Tables 5 and 6 . The ROC and precision–recall curves for each model are presented in Figures 2 and 3 , respectively. Owing to the small number of patients, the dataset was divided into training and testing sets. It can be observed from Table 5 that the training sets of the Random Forest and MLP models were overfitted. The Gaussian Process had the best predictive power and the AUROC of the test set was 0.90. The sensitivity, specificity, PPV, NPV, and accuracy rate of the test dataset for the Gaussian Process were 0.75, 0.90, 0.46, 0.97, and 0.88, respectively. The performance index of each model training set. AUPR = area under the precision–recall curve, AUROC = area under the receiver operating characteristic curve, MLP = Multilayer Perceptron, NPV = negative predictive value, PPV = positive predictive value, XGBoost = extreme gradient boosting. The performance index of each model test set. AUPR = area under the precision–recall curve, AUROC = area under the receiver operating characteristic curve, MLP = Multilayer Perceptron, NPV = negative predictive value, PPV = positive predictive value, XGBoost = extreme gradient boosting. Receiver operating characteristic curve of 4 models (A, B, C, D are the ROC curves of Random Forest, Gaussian Process, XGBoost, and Multilayer Perceptron, in turn). ROC = Receiver operating characteristic, XGBoost = extreme gradient boosting. Precision–recall curve of 4 models (A, B, C, D are the PR curves of Random Forest, Gaussian Process, XGBoost, and Multilayer Perceptron, in turn). PR = precision–recall curve, XGBoost = extreme gradient boosting.

EMT is classified according to the anatomical location of the lesions and is categorized as peritoneal, ovarian, or DIE. [ 14 , 15 ] Ovarian endometriosis is the most prevalent form of EMT among women, with a reported prevalence ranging from 17% to 44%. [ 4 ] Surgical intervention is frequently employed as a therapeutic option in patients with intolerable pain and infertility associated with EMT. [ 16 ] Although conservative surgery can improve patients’ clinical symptoms and fertility, it has a high recurrence rate, which may be due to the fact that hidden EMT lesions cannot be completely removed, and the residual lesions continue to infiltrate under the action of hormones, immunity and other factors, leading to the emergence of new lesions, etc. However, the specific cause remains unknown. The recurrence rate after conservative EMT surgery varies widely. This may be related to the varying inclusion and exclusion criteria, follow-up time, sample size, and research methods used in different studies. Choi et al reported that the recurrence rate of OEM patients 2 years after surgery was 5.8%. [ 17 ] The results of a retrospective cross-sectional study on the postoperative recurrence of patients with OEM indicated that the recurrence rate of OEM patients with 1 year after conservative surgery was 22.6%. [ 18 ] The research results of Li et al showed that the recurrence rate of OEM patients 5 to 10 years after surgery was 15.4% to 22.5%. [ 12 ] In a study by Su et al, the recurrence rate of OEM after laparoscopic conservative surgery was 16.98%. [ 19 ] In this study, the recurrence rate of OEM patients within 2 years after the operation was 18.9%, which is consistent with previous studies. In this study, significant statistical differences were observed between the non-recurrence and recurrence groups in terms of preoperative pregnancy, age, and family history, consistent with previous research findings. [ 12 , 20 , 21 ] The risk of recurrence in patients with EMT decreases with age, potentially due to the hormone-dependent nature of EMT and the gradual decline in ovarian function, a key organ for estrogen secretion, with advancing age. [ 22 , 23 ] Furthermore, EMT is a polygenic disease, with genetic factors contributing to 47% of its development. [ 24 ] Our findings indicate that ovarian cyst diameter, r-AFS stage, presence of DIE, and preoperative CA125 levels are associated with postoperative recurrence, consistent with numerous studies. [ 5 , 25 , 26 ] Additionally, this study identified that bilateral ovarian involvement, presence of peritoneal EMT, and type of ovarian cysts were related to postoperative recurrence. The presence of peritoneal type EMT and bilateral ovarian involvement are components of the r-AFS score, which correlates with the postoperative recurrence. For patients with larger lesions, multilocular cysts, and capsular walls that are not easy to peel off, the more difficult the operation, the higher the risk of residual lesions, which continue to grow under the action of postoperative sex hormones secreted by the ovary, and the higher the possibility of recurrence found on postoperative examination. Patients with EMT typically exhibit elevated coagulation and inflammatory states. [ 27 ] Coagulation-related indicators (activated partial thromboplastin time, TT, fibrinogen, and D-dimer) have been shown to be valuable in the noninvasive diagnosis of EMT. [ 28 ] However, a consensus on the significance of inflammation- and coagulation-related indicators in predicting recurrence remains elusive. Among the inflammation- and coagulation-related indicators examined in this study, only TT level was associated with postoperative recurrence in patients with EMT. This study also found that postoperative medication use and its duration were related to recurrence, corroborating the findings of previous studies. [ 17 , 19 ] In our study, no significant correlation was found between postoperative pregnancy and recurrence, however, postoperative live birth and lactation duration were associated with recurrence. This may be attributed to early pregnancy miscarriages in patients who did not deliver, whereas those who successfully delivered experienced longer pregnancies. During pregnancy, patients do not ovulate or menstruate, eliminating the menstrual reflux phenomenon, which is a key hypothesis for the etiology of EMT. [ 29 ] Additionally, hormonal changes during pregnancy, particularly the high progesterone state, may inhibit the growth of ectopic lesions through various mechanisms, [ 30 , 31 ] potentially reducing the recurrence rate of EMT. The selection of these 4 models for this study was attributed to their suitability for small data samples and their robust fitting capabilities. [ 32 – 36 ] Additionally, Gaussian processes offer the advantage of uncertainty measurement and provision of confidence intervals. In instances where the predicted uncertainty is elevated, it is prudent to exercise caution in verification, potentially prompting patients to undergo further diagnostic procedures, such as pelvic MRI. Although these 4 machine learning models have not been previously employed to predict the recurrence of EMT, they have been extensively utilized to forecast the prognosis of analogous diseases and disease gene mutations. Specifically, XGBoost and random forest have been applied to predict the recurrence of ovarian tumors, [ 37 – 39 ] whereas Gaussian processes have been employed to assess TP53 mutations in endometrial cancer. [ 40 ] Furthermore, the Multilayer Perceptron demonstrates high sensitivity in predicting recurrence following oral cancer surgery. [ 41 ] Su et al [ 13 ] established a nomogram model with an AUC of 0.895 to predict the risk of postoperative recurrence in OEM patients. Compared to the research conducted by Su et al, our study incorporated a greater number of clinical analysis indicators and a larger patient cohort. Additionally, we employed 4 distinct machine-learning techniques to develop recurrence prediction models. In this study, the training sets for the random forest and MLP models exhibited an overfitting. The overfitting phenomenon may be attributed to an imbalance between positive and negative samples. Specifically, the model is predisposed to favor negative samples because of their greater abundance than positive samples. However, the prediction model established using the Gaussian process demonstrated robust predictive capability, with the AUROC of the test set reaching 0.90. The sensitivity, specificity, PPV, NPV, and accuracy rate of the test dataset for the Gaussian Process were 0.75, 0.90, 0.46, 0.97, and 0.88, respectively. Based on the results of this study, we believe that patients after OEM surgery need long-term drug therapy and long-term management to delay recurrence, improve patient prognosis, and avoid repeated and multiple operations. The Gaussian process-based prediction model developed in this study serves as a valuable tool for medical professionals to evaluate the risk of postoperative recurrence in patients with OEM. It also aids in devising prevention strategies to enhance the prognosis of patients. In particular, for women with fertility considerations, personalized treatment and fertility guidance plans can be formulated based on the Endometriosis Fertility Index score and prediction model presented in this study. For instance, patients with a high risk of recurrence and infertility, even with a high Endometriosis Fertility Index score, are advised to consider reducing the time spent attempting conception and pursuing medical intervention at an earlier stage. For patients with long-term postoperative fertility needs, it is recommended that they continue medication until they plan for pregnancy and are informed of the potential for recurrence within 2 years. This information allows patients to strategically plan their conception attempts, thereby avoiding a decline in pregnancy rates due to a lack of understanding of the disease. For patients who successfully conceive post-surgery, extending the breastfeeding period is advised to mitigate disease recurrence. Consequently, we advocate the integration of this model into standard diagnostic and treatment protocols for patients with EMT. However, our research has some limitations. Firstly, this study was a single-center retrospective analysis. Besides, the sample size was limited. Secondly, the prediction model constructed in this research has not yet undergone prospective verification, and the reliability of the model awaits further verification. Thirdly, patients undergoing medical or nonconservative treatment were overlooked or not investigated. Therefore, future research should involve multicenter and large-sample studies to further verify the reliability of the model and provide more powerful evidence for clinical applications.

In summary, age, preoperative pregnancies, preoperative history of dysmenorrhea, family history, diameter of ovarian cyst, single/bilateral cyst, type of ovarian cyst, r-AFS stage, combined with peritoneal EMT, combined with DIE, preoperative serum CA125 level, AMH, TT, total duration of medication, postoperative live birth, and breastfeeding duration were associated with the prognosis of patients with OEM undergoing laparoscopic conservative surgery. The predictive model established using these metrics can guide medical professionals to assess the risk of postoperative recurrence in patients with OEM and help develop prevention strategies, especially to help women with fertility requirements to develop personalized treatment and fertility guidance programs.

We would like to acknowledge the https://paperpal.editage.cn for their outstanding scientific proofreading and editing services that was provided for this manuscript.