Endometriosis (EMs) is a common gynecological and hormone-dependent disease [1, 2]. It is estimated that 25%–50% of infertility patients have EMs, and 30% to 50% of EMs patients have difficulty in conception [3, 4]. In addition, EMs is also one of the most important causes of female infertility, which has brought a burden to many families who are eager to give birth [5, 6]. The typical features of EMs is the abnormal growth of endometrial tissue outside the uterus, such as pelvic organs, peritoneum, and even distant organs. This abnormal growth can cause a series of symptoms that seriously affect the quality of life of patients, among which chronic pelvic pain is the most common, and some patients have severe pain, which seriously interferes with daily activities [7, 8].

With the progress of modern medical technology, laparoscopic surgery has become the mainstream method for the treatment of EMs due to its significant advantages such as small trauma, rapid postoperative recovery and short hospitalization time. Under laparoscopic direct vision, doctors can accurately locate and effectively remove or destroy ectopic endometrial lesions, significantly alleviating the symptoms of patients [9]. At the same time, gonadotropin-releasing hormone agonist (GnRHa) plays a key role in postoperative adjuvant therapy [10]. GnRHa inhibits the secretion of gonadotropins by binding to the receptors on the surface of pituitary gonadotropin cells, thereby reducing the level of estrogen in the body, shrinking the ectopic endometrial tissue due to the lack of estrogen support, and achieving the purpose of inhibiting its growth [11]. However, although laparoscopic combined with GnRHa treatment can effectively improve the symptoms and quality of life of patients in the short term, the long-term recurrence problem has always been a major problem in clinical treatment. According to research statistics, there is a significant difference in the recurrence rate of endometriosis. The average recurrence rate of 2 years is between 20% (0%~89%), while the average recurrence rate of 5 years is in the range of 50% (15%~56%) [12, 13]. Disease recurrence not only means that patients must bear physical problems again, such as increased pain, infertility, and even deterioration, but also brings a heavy economic burden to patients and their families. In addition, the difficulty of treatment after multiple relapses increases significantly, and more aggressive treatment may be needed, which can further affect the patient’s physical and mental health. Therefore, it is of great importance to explore the influencing factors of EMs recurrence and construct a risk prediction model, which is helpful to identify high-risk patients of EMs recurrence as soon as possible and take targeted preventive interventions to improve the clinical prognosis of EMs patients.

Previous prediction of long-term recurrence of EMs has mainly depended on the experience of clinicians and some simple clinical indicators, such as patient age, disease stage, and surgical method [14, 15]. However, these methods are greatly affected by subjective factors, and cannot comprehensively consider the patient’s complex condition, so it is difficult to comprehensively and accurately predict the recurrence risk of individual patients. However, most of the existing prediction models are based on traditional statistical methods, which have problems such as insufficient data volume, insufficient model complexity, poor generalization ability, and are difficult to cope with complex clinical scenarios. In recent years, the rapid development of big data and machine learning technology has brought new opportunities for research and practice in the medical field. Machine learning algorithms can thoroughly mine and analyze massive and complex data, find potential associations and patterns, and build more accurate prediction models. The random forest (RF) model and gradient boosting machine (GBM) model in machine learning have unique advantages in dealing with complex data and constructing prediction models. The RF model has good generalization ability and anti-overfitting ability by constructing multiple decision trees and synthesizing their results for prediction. The GBM model is based on the gradient descent algorithm, and gradually iterates to improve the performance of the weak learner to build a more powerful prediction model. In other disease areas, such as breast cancer, hip fracture and cardiovascular disease, etc. [16, 17, 18, 19], these two models have been successfully applied to early screening and prognosis prediction of diseases. In the field of EMs, the application of RF and GBM models in the prediction of long-term recurrence after laparoscopy combined with GnRHa treatment is still scarce. In view of the high recurrence rate of EMs and the limitations of traditional prediction methods, this study aims to use RF and GBM models to comprehensively analyze the multiple factors affecting the long-term recurrence of EMs, and to construct an accurate recurrence prediction model. It provides a scientific basis for clinicians to formulate personalized treatment plans and implement effective recurrence prevention measures, in order to reduce the long-term recurrence rate of EMs in patients and improve the overall prognosis of EMs.

Laparoscopic surgery is an effective method for the treatment of EMs, but the recurrence rate is high [23]. Incomplete surgical resection, residual lesions and new lesions are closely related to EMs recurrence [24, 25]. Patients with recurrent EMs have an increased risk of organ damage and ovarian function damage during reoperation, and the postoperative pregnancy rate is significantly decreased [26, 27]. GnRHa can inhibit pituitary activity, reduce estrogen levels, and block ectopic endometrial growth [28]. A number of studies have reported that the effective rate of laparoscopy combined with GnRHa is as high as more than 90%, and the effect on severe infiltration or occult lesions is significant [24, 29, 30]. However, recent studies have shown that the treatment still has the risk of recurrence [31], and the side effects of the drug will affect the reproductive function [32, 33]. This study found that the recurrence rate of EMs patients after laparoscopy combined with GnRHa treatment for 1-year was 19.68%, suggesting that laparoscopy combined with GnRHa treatment for EMs patients still cannot avoid the risk of recurrence. The analysis may be related to the different inclusion and exclusion criteria, follow-up time and research methods of the subjects. How to effectively reduce the risk of postoperative recurrence of EMs patients has attracted more and more attention from the medical community. Traditional prediction methods mainly rely on doctors’ clinical experience and simple clinical indicators, which are limited in accuracy and difficult to meet clinical needs. Therefore, it is particularly necessary to explore and apply more advanced prediction models.

The RF and GBM algorithms have high diagnostic ability and accuracy in disease risk prediction. They can provide the important evaluation of each variable for disease risk prediction, and provide reference information for subsequent clinical decision-making through data dimensionality reduction, pre-screening candidate features, and variable importance ranking [34, 35]. In this study, RF and GBM models were applied to the prediction of long-term recurrence of EMs treated by laparoscopy combined with GnRHa. Based on univariate analysis, the order of importance of the predictors of laparoscopic combined with GnRHa in the treatment of EMs based on the two groups of models was r-AFS staging, course of disease, tender posterior fornix, history of intrauterine operations and history of preoperative endometriosis. The reason for the analysis may be: patients with higher r-AFS staging are more likely to relapse after surgery, which may be due to the deeper infiltration of lesions with higher staging, easy adhesion, and the difficulty of thorough dissection [36]. Tobiume et al. [37] showed that r-AFS score was independently associated with recurrence after EMT, and patients with high r-AFS score were more likely to relapse. The palpable nodules in the posterior fornix were invaded by endometriosis lesions into the rectovaginal septum and uterosacral ligament. The longer the course of the disease, the higher the stage, the deeper the infiltration of the lesions, and the greater the difficulty of complete clearance of the treatment, which increases the risk of residual lesions and is more likely to lead to relapse. In severe cases, they may lead to the closure of the uterine rectal fossa, oppress the sensory nerve at the lesion site, and cause pain in the posterior fornix. The research has shown that posterior fornix tender nodules are risk factors for EMs after surgery [22]. Most of the lesions of the posterior fornix tenderness nodules are hidden, and the infiltration is deep. It is difficult to completely remove them during surgery, resulting in a greater risk of postoperative recurrence [38]. Patients with a history of uterine cavity operation and dysmenorrhea have a higher recurrence rate, which may be related to the late stage of dysmenorrhea patients, usually accompanied by pelvic adhesions, and pelvic lesions are difficult to completely peel off, and excessive uterine cavity surgery can easily lead to lesions being planted in the ovary and pelvic peritoneum, leading to repeated occurrence of endometriosis [39, 40]. Study has found that estrogen and its receptors are highly expressed in patients with endometriosis dysmenorrhea. Lessey and Young [41] proposed that the growth of ectopic lesions depends on estrogen and its receptors, which may also be a reason for improving the postoperative recurrence rate of such patients. Previous anti-endometriosis drug treatment can cause endometrial lesions atrophy, edema and other morphological changes, but to cover up the formation of the lesion, meaning it is difficult to completely remove the lesion during surgery, which then leaves the hidden danger of postoperative recurrence [42].

In this study, the performance of the two models was compared. The results showed that on the training set, the RF model showed a higher AUC value of 0.902, and the sensitivity was 100%. This shows that the RF model has excellent discrimination ability on training data and can accurately identify positive and negative samples. In contrast, the AUC of the GBM model was 0.851, and the sensitivity and specificity were 89.20% and 68.20%, respectively. The Delong test results further confirmed that the AUC of the RF model on the training set was significantly better than that of the GBM model (Z = 2.838, p = 0.005). The advantages of the RF model on the training set may be due to its algorithmic characteristics. RF makes its predictions by constructing multiple decision trees and combining their results. This integrated learning method allows the model to have a stronger generalization ability when dealing with complex data, and can effectively avoid overfitting. In addition, the RF model uses a random sampling method for the construction of each decision tree during the training process, which increases the diversity of the model and improves its learning ability for different feature combinations [33]. The GBM model uses a gradient boosting strategy to gradually optimize model performance by iteratively training multiple weak learners. However, due to its fine fitting of training data, it may lead to over-learning certain specific patterns on the training set, affecting the generalization ability of the model, and then lagging behind the RF model on the AUC of the training set.

On the validation set, the AUC of the RF model was 0.859, and the GBM model was 0.852. The Delong test showed that there was no significant difference between the two models (Z = –0.239, p = 0.811). The sensitivity and specificity of the RF model were 69.2% and 92.9%, respectively, and the sensitivity and specificity of the GBM model were 76.9% and 87.5%, respectively, which indicated that the RF model and the GBM model performed well in predicting EMs recurrence samples, and the discrimination ability and prediction effect of the two models were comparable. The calibration curve shows that the RF and GBM models have high consistency between the prediction probability and the actual results in the training set and the validation set. This indicates that the two models are more accurate in predicting the probability of event occurrence, and jointly provide a quantitative decision-making basis for patients to formulate individualized intervention programs (such as GnRHa treatment duration, follow-up density) and optimize medical resource allocation (reducing excessive examination and treatment). In practical applications, if the fitting effect of training data is required to be high, the RF model is a better choice; If more attention is paid to the stability of the model on the data, both RF and GBM models can be considered. In addition, when the model is applied to the actual scene, it is also necessary to consider the factors such as the interpretability and computational cost of the model, and make decisions after comprehensive trade-offs. Therefore, RF and GBM models have their own advantages, and the choice of which model or whether to use it in combination should be determined according to the specific situation and needs.

However, this study also has some limitations: (1) First of all, this study is a single-center retrospective study, the sample size is relatively small, there may be selection bias, and the extrapolation of the model needs to be verified further. (2) Secondly, this study only included limited clinical features as predictors. In the future, biomarkers such as genes and proteins can be further combined to improve the predictive efficacy of the model. (3) In addition, the interpretability of machine learning models is relatively poor. Although some methods (such as variable importance analysis) can be used to explain the decision-making process of the model, it is still difficult to explain the relationship between variables and outcomes as intuitively as traditional statistical models. Therefore, in the future, doctors can better understand the decision-making process of the model and improve the application value of the model in clinical practice by developing visualization tools or explanatory algorithms.

In summary, r-AFS staging, course of disease, tender posterior fornix, history of intrauterine operations and history of preoperative laparoscopy are the influencing factors of long-term recurrence of EMs treated by laparoscopy combined with GnRHa. The recurrence of such patients after treatment can be effectively predicted by constructing RF and GBM models. The RF model was preferentially selected for the initial screening of recurrence risk to avoid missed diagnosis, and the dynamic risk update was combined with the GBM model to optimize the intervention strategy. This study provides a quantitative tool for individualized management of EMs patients, which can accurately formulate GnRHa treatment duration and follow-up plan through model prediction probability, and provide a paradigm for machine learning prediction model transformation of benign diseases.

Data is available from the corresponding author on reasonable request.

KZ, RZ and YJ designed this study. KZ conducted research. KZ provided assistance in the design of the research plan and suggestions for the overall writing approach of the manuscript. RZ and YJ analyzed these data. All authors contributed to editorial changes in the manuscript. All authors read and approved the final manuscript. All authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work.

This study was conducted in accordance with The guiding principles of the Declaration of Helsinki, and the protocol was approved by the Ethics Committee of The First Affiliated Hospital, College of Medicine, Zhejiang University (Approval Number: 2025B0730). Informed consent was obtained from all patients or their families/legal guardians prior to participation.

This research received no external funding.

The authors declare no conflict of interest.