Identification and validation of a novel machine learning model for predicting severe pelvic endometriosis: A retrospective study

In this study, the stage of endometriosis was determined according to the scoring system of the revised American Society for Reproductive Medicine (rASRM) 7 . Patients with stage IV disease (score > 40) were categorized as the severe group. The diagnosis of severe endometriosis relies on visualization during surgery (typically involving laparoscopy) and pathology.

The data for this single-center retrospective study were obtained from 471 patients who sought consultation from the Department of Gynecology and Ultrasound at the First Hospital of China Medical University between December 2019 and December 2023. Patients underwent surgical treatment and were diagnosed with endometriosis by postoperative pathology. The inclusion criteria were as follows: (a) patients whose pathology confirmed a diagnosis of endometriosis, (b) patients who underwent surgical treatment, and (c) patients who were older than 18 years and who underwent transvaginal ultrasonography and urinary system ultrasonography before surgery (demographic characteristics and pathological, laboratory and sonographic imaging data were available). The exclusion criteria were as follows: (a) had endometriotic lesions located beyond the pelvic area, (b) had malignant tumors, (c) had a surgical history of endometriosis or other pelvic disorders, (d) received hormonal treatment within the last three months, and (e) missing data on primary outcomes. This research protocol followed the principles outlined in the Declaration of Helsinki and received approval from the Ethics Committee of the First Affiliated Hospital of China Medical University ([2023]577). All participant information was kept confidential. No names or other identifying information was included in the study. The flowchart of the study is presented in Fig.  1 . Fig. 1 Flowchart of the study design. Flowchart of the study design.

A total of 39 variables were evaluated, including basic information such as age, severe dysmenorrhea, and other menstrual cycle-related symptoms (MCRS) including constipation, diarrhea, urinary frequency, and urgency. The visual analog scale (VAS) was employed to assess the severity of dysmenorrhea in patients. Menstrual history included menopausal status, parity, gravidity, number of abortions, age at menarche, and period and cycle length. Laboratory examination results included prothrombin time (PT), prothrombin time activity (PTA), international normalized ratio (INR), and Cancer Antigen 125 (CA125) levels. Ultrasound data included ultrasonically suspected endometriotic nodules (USEN), adenomyosis, retroflexed uterus, negative sliding sign, pelvic fluid observed in the cul-de-sac (PFOC), bilateral OEs, number of OE cavities, papillary projections of OE, maximum OE diameter, and ovarian mobility. The diagnosis of adenomyosis was based on the criteria proposed by MUSA (Morphological Uterus Sonographic Assessment) 20 . Typical sonographic signs of the ultrasound data are shown in Fig.  2 . Fig. 2 Typical ultrasonography of related ultrasound data. Typical ultrasonography of related ultrasound data. A. obliterated Pouch of Douglas; B. bilateral OEs; C. unilateral OE; D. adenomyosis; obliterated Pouch of Douglas; B. bilateral OEs; C. unilateral OE; D. adenomyosis; E. papillary projections (arrow heads); F. OE cavities (arrow heads); G. ultrasonically suspected endometriotic nodules; H. anteversion of uterus; I. anteverted retroflexed uterus.

A LASSO regression analysis was conducted using 39 independent variables with vertical lines plotted at values selected using tenfold cross-validation, resulting in 18 features with nonzero coefficients (Fig.  3 A, B), including negative sliding signs, bilateral OEs, PFOC, severe dysmenorrhea, eosinophil%, MCRS, thrombin time (TT), cycle period, retroflexed uterus, ovarian mobility, prothrombin activity (PTA), D-dimer levels, CA125, maximum OE diameter, and menopausal status. Fig. 3 Demographic and clinical feature selection using LASSO regression. (A) Feature Selection Process Using LASSO Regression. (B) Coefficient Profiles and L1 Norm Across Different Log(λ) Values. Demographic and clinical feature selection using LASSO regression. (A) Feature Selection Process Using LASSO Regression. (B) Coefficient Profiles and L1 Norm Across Different Log(λ) Values. The data of 18 features selected by LASSO regression analyses were used to generate seven ML models to predict severe endometriosis. The AUROC curves and accuracy of the seven ML models are presented in Fig.  4 . The performances of these seven models are listed in Supplementary Table S2. Comparative analysis revealed that RF models exhibited the most robust performance in predicting severe endometriosis. Supplementary Fig. S1 A visually illustrates the relationship between the actual feature values and their corresponding SHAP values in the RF model, showing how each feature contributes to the predicted probability. The “negative sliding sign” emerged as the most influential feature in this model. Variable importance in projection (VIP) scores were used to rank the variables in terms of their importance to the predictive model. The variables are presented in Supplementary Fig. S1 B. Fig. 4 Performance of ML models to predict severe endometriosis. (A) Comparison of Accuracy and ROC-AUC for Machine Learning Models in Predicting Severe Pelvic Endometriosis. (B) ROC Curves of Seven Machine Learning Models for Predicting Severe Pelvic Endometriosis. Performance of ML models to predict severe endometriosis. (A) Comparison of Accuracy and ROC-AUC for Machine Learning Models in Predicting Severe Pelvic Endometriosis. (B) ROC Curves of Seven Machine Learning Models for Predicting Severe Pelvic Endometriosis. The final model was identified by reducing features of the RF model. Among the 18 features, six were significant (Supplementary Fig. S2). The six features were put in the predictive model to reduce the number of features, and the AUC for the ROC curve is shown in Fig.  5 A. The Hence, we focused on the six-feature RF model, which included negative sliding signs, PFOC, bilateral OEs, severe dysmenorrhea, eosinophil% and MCRS, as the final model for further analysis. Table  2 summarizes the performance of the final model. The final model achieved AUROCs of 0.865 and 0.720 in the training and testing datasets, respectively. We examined the variance inflation factor (VIF) to detect multicollinearity (ordered by importance) (Supplementary Table S3). All predictor variables had a variance inflation factor (VIF) values of less than 10, indicating no multicollinearity. Table 2 The performances of the final model. Model Set Metric Value Six-feature prediction model Train Accuracy 0.776 Train AUROC 0.865 Test Accuracy 0.718 Test AUROC 0.720 The performances of the final model. Fig. 5 (A) SHAP Values for Feature Importance in the final model. (B) Mean SHAP Values of Key Features Influencing the final model. (C) ROC Curves for the final model’s Performance on Training and Testing Sets. (A) SHAP Values for Feature Importance in the final model. (B) Mean SHAP Values of Key Features Influencing the final model. (C) ROC Curves for the final model’s Performance on Training and Testing Sets. To ensure the clinical utility and transparency of the model, we employed the SHAP method for interpretability, which quantifies the contribution of each variable to the prediction. This allowed us to understand the model’s decision-making process at both global and individual levels. The SHAP summary plot (Fig.  5 B) visualize the average contribution of each feature, ranked in descending order, providing a global understanding of their influence. Additionally, individual Shapley values (Fig.  5 C), known as “Local Shapley values”, reveal how specific features contribute to the prediction for each patient, allowing clinicians to understand the reasoning behind the model’s output on an individual level. By leveraging the SHAP method, we provided a clear and interpretable framework for clinicians to understand the model’s predictions and foster trust and confidence in its application. Furthermore, local interpretations examine the process by which individualized predictions are generated for specific individuals through the integration of personalized input data. Fig.  6 A-D illustrates a patient diagnosed with severe endometriosis, as confirmed by laparoscopic surgery and postoperative pathological analysis. The actual measured values of relevant characteristics are presented in the waterfall diagram in Fig.  6 D. Conversely, Fig.  6 E-F depicts a patient with non-severe endometriosis, also confirmed through laparoscopic surgery and postoperative pathological findings. The characteristics and their corresponding measurements, which influence the classification towards either the severe or non-severe endometriosis category, are detailed in Fig.  6 F. Fig. 6 Local Shapley values for interpretation of the final model. (A) left OE; (B) right OE; (C) obliterated Pouch of Douglas; (D) waterfall plot of patients at high risk of severe endometriosis; (E) right OE; (F) Color Doppler Flow Imaging of right OE; (G) patients at low risk of severe endometriosis. Local Shapley values for interpretation of the final model. (A) left OE; (B) right OE; (C) obliterated Pouch of Douglas; (D) waterfall plot of patients at high risk of severe endometriosis; (E) right OE; (F) Color Doppler Flow Imaging of right OE; (G) patients at low risk of severe endometriosis. The horizontal ordinate represents the decision score. f is the function of the predictive model, f(x) is the final decision probability for input x, and E[f(x)] is the expected value (mean value) of the final model’s decision probabilities for all training samples (which is 0.396). For sample (D), the negative sliding sign and bilateral OEs make + 0.181 and + 0.105 contributions, respectively, supporting the decision to apply severe endometriosis. For sample (F), the contributions of the sliding sign, negative bilateral OEs, were − 0.268 and − 0.214, respectively, supporting the decision to treat non-severe endometriosis. Compared with non-severe endometriosis, severe endometriosis is characterized by more altered pelvic anatomy and extensive adhesions 25 . Laparoscopy is costly and undoubtedly risky for patients with severe endometriosis. The diagnostic value of US images and clinical parameters for endometriosis have been explored previously. However, little is known about the diagnostic capacity of endometriosis staging. Therefore, we proposed a predictive model based on the ML model to easily and reliably predict severe endometriosis. In this study, we identified and evaluated a novel ML model for the prediction of severe pelvic endometriosis that comprised routinely available clinical sonographic and laboratory data. The model achieved AUROCs of 0.865 and 0.720 in the training and testing datasets, respectively. Given the moderate diagnostic performance, it is considered acceptable when taking into account the availability and cost of predictive factors. As illustrated in Fig.  6 , our study identified that a negative sliding sign, in conjunction with a PFOC, proved advantageous for diagnosing severe endometriosis. This finding underscores the positive impact of these ultrasonic markers on the diagnostic accuracy for severe endometriosis, as they are indicative of an obliterated cul-de-sac in affected patients. Previous studies have shown that the TVS sliding sign has good diagnostic performance for predicting an obliterated cul-de-sac. In a more recent meta-analysis published in 2022, the pooled estimated sensitivity and specificity of TVS for detecting obliterated cul-de-sac were 88% (95% CI: 81–93%) and 94% (95% CI: 91–96%), respectively 26 . PFOC which means the presence of freely accumulating fluid in the cul-de-sac may suggest the absence of significant adhesions in the area, as adhesions would otherwise restrict fluid movement and accumulation. A 2023 study conducted by Mick proposed that the presence of a triangular-shaped accumulation of pelvic fluid in patients with a retroverted uterus may serve as a novel ultrasonic indicator for an unobliterated cul-de-sac. In an abnormal setting in which there is obliteration, we cannot elicit the triangle sign because the fluid cannot fill the densely adherent cul-de-sac 27 . According to the American Society for Reproductive Medicine (ASRM) classification, if an obliterated cul-de-sac is observed, the disease is classified as severe endometriosis 7 , 8 . In this study, we also assessed the contribution of certain scoring items in the ASRM system, including bilateral OEs, to the predictive model. Similarly, Zhao et al. 28 attempted to identify predictors of preoperative endometriosis severity and proposed a prediction formula. This retrospective study included patients with severe dysmenorrhea, which is consistent with our findings. Nevertheless, this study offers a more comprehensive analysis than the previous one, as it refines the content of pelvic examinations, laboratory tests, and transvaginal sonography (TVS) tests, encompassing a total of 39 variables. The strength of this study is that we developed a predictive model for severe endometriosis. These predictors are readily available and provide new insights into the pathogenesis of endometriosis. The gynecologist is capable of conducting a thorough and comprehensive preoperative assessment to formulate an optimal treatment strategy, determine the necessity of surgical intervention, and devise an appropriate surgical approach. In terms of laboratory indices, patients with severe endometriosis had higher D-dimer levels, indicating hypercoagulability, than those with non-severe endometriosis. This may be attributed to activation of the intrinsic coagulation system by recurrent bleeding in patients. Coagulation-related factors have excellent diagnostic value in endometriosis and are correlated with disease prognosis and endometriotic cyst diameter 29 , 30 . A multicenter prospective study reported lower EO patients with endometriosis than in controls ( p = 0.045), which differs from the results of this study 31 . Although promising, this study has some limitations that warrant consideration. This retrospective, single-center design and relatively small sample size limit the generalizability of the findings. External validation of the prediction model is essential to assess its performance in different populations and settings. In this study, the staging method we adopted is the most widely used rASRM staging system, which has certain limitations in DE staging. Considering this staging system does not take into account the presence of deep endometriosis, it is possible that some women that indeed had severe endometriosis were wrongly classified. The model had an AUROC of 0.72 in the test set, which may be related to the reasons mentioned above. Furthermore, while repeatability analysis within the training and testing sets demonstrated high consistency, some degree of error might be present owing to the inherent uncertainties in segmentation. In this study, we developed a predictive model for stage IV endometriosis using machine learning, with a random forest model demonstrating superior performance to predict the severity of this disease before surgery. The SHAP method provides personalized risk assessments that enhance the interpretability and clinical utility of the model. This computer-aided approach holds potential for assisting clinicians in managing severe endometriosis, potentially improving noninvasive diagnostics, reducing delays, and optimizing surgical outcomes. Future research should focus on the external validation of the model with larger, multi-center datasets to explore the impact of different data sources and feature engineering strategies. Incorporating additional data sources, such as MRI results or genetic markers, could further enhance the model’s predictive accuracy. Additionally, investigating the application of the model in various clinical scenarios and assessing its cost-effectiveness will further enhance its impact on patient care.

Ultrasonic examination was performed with a GE Voluson E8, a GE Voluson E10 (GE Healthcare, Chicago, IL, USA), or a Philips EPIQ7 (Philips EPIQ7, Bothell Everett Highway, Washington, USA) with a 1.0–5.0 MHz transabdominal probe and a 5.0–9.0 MHz transvaginal probe. All patients underwent preoperative transabdominal and transvaginal ultrasound based on four four-step screening approaches proposed by the International Deep Endometriosis Analysis (IDEA) consensus 19 . All examinations were conducted by independent, experienced sonographers who were blinded to other imaging, pathological findings and stage of rASRM.

We performed feature selection using the least absolute shrinkage and selection operator (LASSO), which compresses variable coefficients to prevent overfitting and address severe covariance problems. Subsequently, seven ML algorithms, including logistic regression (LR), recursive partitioning and regression trees (rpart), random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM), k-nearest neighbors (KNN), and neural network (NNET), were used to develop the predictive model. These ML algorithms, known for accurately capturing non-linear relationships between variables, have been effectively used in clinical prediction 21 – 23 . The performance of the model in predicting severe endometriosis was evaluated using the area under the receiver operating characteristic curve (AUROC) and accuracy analysis. The final model was determined after hyperparameter tuning via grid search and 10-fold cross-validation for each algorithm. Characteristic importance was assessed using Shapley additive interpretation (SHAP), which provides the degree of influence of each feature in the sample on the model and the positive and negative effects on the model. Higher absolute SHAP values were associated with features that had the greatest impact on the model’s predictive scores. All predictive models were implemented using R (version 4.3.1) software with the mlr3 package. The study participants were divided into a severe group ( n  = 183) and a non-severe group ( n = 125) according to the rASRM system 7 . Patients were randomly divided into training and testing groups at a ratio of 8:2. The testing dataset was used to evaluate and compare the performance of each model. Continuous variables are described as the mean ± standard deviation (‾ x ± s), and comparations between groups were performed using the t-test for continuous variables. For continuous variables that did not follow a normal distribution, medians and quartiles are presented, and Wilcoxon tests were used to compare groups. Categorical variables are expressed as frequencies (percentages), and comparisons between groups of unordered categorical variables were made using the χ 2 test or Fisher’s exact probability method. For missing data, as shown in Supplementary Table S1 , the random forest algorithm was used for interpolation 24 . Sensitivity analyses will be employed to investigate the impact of missing data. A P  <  0.05 indicated a significant difference. A total of 308 patients were included in this study. A comparison of the preoperative variables and detailed information on the features of the demographics between the severe and non-severe groups is presented in Table  1 , which indicates that there was no statistically significant difference between the two groups ( p  > 0.05). Among the women screened, the rate of severe endometriosis was approximately 59.2% (183/308). The study population ( n  = 308) was randomly divided into a training set (246) and a testing set (62), which were used to further validate the predictive models. No statistically significant difference existed between the patient characteristics in the training and testing datasets. Table 1 Demographics characteristics between the severe and non-severe endometriosis groups. Variable Overall, Non-severe group, Severe group, P 2 N  = 308 1 N  = 125 1 N  = 183 1 Age(years) 0.03 18–35 113 (37%) 56 (45%) 57 (31%) 36–45 142 (46%) 47 (38%) 95 (52%) ≥ 46 53 (17%) 22 (18%) 31 (17%) Menopausal status 0.075 Premenopausal 296(96%) 117(94%) 179(98%) Postmenopausal 12 (4%) 8 (6%) 4 (2%) Age at menarche(years) 0.171 ≤ 14 41 (13%) 21 (17%) 20 (11%) > 14 267(87%) 104 (83%) 163(89%) Cycle-period(days) 0.595 ≤ 4 58 (19%) 22 (18%) 36 (20%) 5–6 123 (40%) 47 (38%) 76 (42%) ≥ 7 127 (41%) 56 (45%) 71 (39%) Cycle-length(days) 0.371 <25 94 (31%) 39 (31%) 55 (30%) 25–29 74 (24%) 25 (20%) 49 (27%) 30–34 140 (45%) 61 (49%) 79 (43%) Number of parities 0.833 0 84 (27%) 36 (29%) 48 (26%) 1 89 (29%) 33 (26%) 56 (31%) 2 67 (22%) 30 (24%) 37 (20%) 3 51 (17%) 18 (14%) 33 (18%) 4 13 (4.2%) 6 (4.8%) 7 (3.8%) ≥ 5 4 (1.3%) 2 (1.6%) 2 (1.1%) Number of gravidities 0.6 0 102 (33%) 45 (36%) 57 (31%) 1 185 (60%) 73 (58%) 112 (61%) ≥ 2 21 (6.8%) 7 (5.6%) 14 (7.7%) Number of abortions 0.741 0 170 (55%) 67 (54%) 103 (56%) 1 76 (25%) 33 (26%) 43 (23%) 2 45 (15%) 16 (13%) 29 (16%) 3 13 (4.2%) 7 (5.6%) 6 (3.3%) ≥ 4 4 (1.3%) 2 (1.6%) 2 (1.1%) Severe dysmenorrhea 0.026 Yes 130 (42%) 43 (34%) 87 (48%) No 178(58%) 82(66%) 96(52%) MCRS 0.068 Yes 35 (11%) 9 (7%) 26 (14%) No 273(89%) 114(93%) 157(86%) PT(s) 13.16 (0.85) 13.18 (0.96) 13.14 (0.77) 0.334 PTA 100 (11) 99 (11) 101 (11) 0.027 INR 1.00 (0.06) 1.01 (0.07) 1.00 (0.06) 0.101 APTT(s) 36.3 (4.0) 36.7 (4.4) 36.0 (3.7) 0.066 Fibrinogen(g/L) 3.05 (0.76) 3.01 (0.70) 3.08 (0.80) 0.611 Thrombin Time(s) 16.55 (1.01) 16.54 (1.03) 16.56 (1.00) 0.673 D-dimer(ln) −1.31 (−1.51, −0.94) −1.41 (−1.51, −1.02) −1.2 (−1.5, −0.88) 0.02 White blood cells(x10 9 /L) 5.89 (4.89, 7.20) 6.06 (4.97, 7.25) 5.81 (4.83, 7.10) 0.566 Neutrophil(x10 9 /L) 60 (52, 65) 59 (51, 65) 60 (52, 66) 0.327 Lymphocyte(x10 9 /L) 31 (25, 38) 32 (26, 39) 31 (25, 37) 0.288 Monocyte(x10 9 /L) 7.08 (1.68) 6.98 (1.72) 7.14 (1.66) 0.55 Eosinophil% 1.60 (0.80, 2.50) 1.50 (0.70, 2.30) 1.60 (0.90, 2.70) 0.134 Red blood cells(x10 12 /L) 4.17 (0.47) 4.23 (0.47) 4.14 (0.47) 0.202 Hemoglobin(g/L) 117 (18) 117 (18) 117 (19) 0.878 Plate(x10 9 /L) 259 (213, 301) 261 (210, 296) 259 (218, 301) 0.864 PLR 146 (113, 190) 135 (112, 189) 152 (116, 195) 0.336 NLR 1.91 (1.35, 2.54) 1.84 (1.31, 2.48) 1.93 (1.42, 2.69) 0.278 CA125((U/mL)) 48 (30, 80) 37 (25, 61) 58 (34, 91) < 0.001 USEN 0.012 Yes 43 (14%) 10 (8%) 33 (18%) No 265(86%) 115(92%) 150(82%) Adenomyosis 0.039 Yes 48 (16%) 13 (10%) 35 (19%) No 268(84%) 112(90%) 148(81%) Retroflexed uterus 0.046 Yes 78 (25%) 24 (19%) 54 (30%) No 230(75%) 101(81%) 129(70%) Sliding sign(negative) < 0.001 Yes 107 (35%) 12 (10%) 95 (52%) No 201(65%) 113(90%) 88(48%) PFOC < 0.001 No 113 (37%) 62 (50%) 51 (28%) Yes 195(63%) 63(50%) 132(72%) Bilateral OEs < 0.001 No 214 (69%) 108 (86%) 106 (58%) Yes 94 (31%) 17 (14%) 77 (42%) Number of OE cavities 0.019 0 23 (7.5%) 7 (5.6%) 16 (8.7%) 1 167 (54%) 80 (64%) 87 (48%) ≥2 118 (38%) 38 (30%) 80 (44%) Papillary projections of OE 0.357 Yes 21 (7%) 6 (5%) 15 (8%) No 287(93%) 119(95%) 168(92%) OE diameter(cm) 6.6(4.7, 8.3) 6.3(4.4, 8.0) 6.7 (4.9, 8.3) 0.45 Ovarian mobility < 0.001 No 182 (59%) 92 (74%) 90 (49%) Yes(one ovary) 98 (32%) 29 (23%) 69 (38%) Yes(two ovaries) 28 (9.1%) 4 (3.2%) 24 (13%) 1 : Mean (SD); Median (Q1, Q3); n (%); 2 : Fisher’s exact test; Wilcoxon rank sum test; MCRS: Menstrual-cycle-related-symptoms; PT: Prothrombin time; PTA: Prothrombin activity; INR: International normalized ratio; APTT: Activated partial thromboplastin time; D-dimer is ln-transformed; PLR: Platelet-to-lymphocyte ratio; NLR: Neutrophil-to-lymphocyte ratio; USEN: Ultrasonically suspected endometriotic nodules; PFOC: pelvic fluid observed in the cul-de-sac. Demographics characteristics between the severe and non-severe endometriosis groups. 1 : Mean (SD); Median (Q1, Q3); n (%); 2 : Fisher’s exact test; Wilcoxon rank sum test; MCRS: Menstrual-cycle-related-symptoms; PT: Prothrombin time; PTA: Prothrombin activity; INR: International normalized ratio; APTT: Activated partial thromboplastin time; D-dimer is ln-transformed; PLR: Platelet-to-lymphocyte ratio; NLR: Neutrophil-to-lymphocyte ratio; USEN: Ultrasonically suspected endometriotic nodules; PFOC: pelvic fluid observed in the cul-de-sac.

As a common chronic inflammatory disease, endometriosis is defined by the presence of endometrial-like tissue outside the uterus and is closely associated with dysmenorrhea, pelvic pain, infertility, dysuria, and deep dyspareunia 1 – 3 . It poses a significant threat to quality of life, especially in women of reproductive age, with an increasing prevalence rate of 10% 4 . Endometriosis can be categorized into different phenotypes, such as ovarian endometriosis (OE), peritoneal endometriosis (PE), deep endometriosis (DE), and other types 5 . To adequately describe this complex, multifaceted disease, several classification systems for endometriosis have been developed, with the revised American Society for Reproductive Medicine (rASRM) being the most extensively used 6 . The rASRM classification system takes into account mainly the location, size, and depth of the implants and the degree of adhesions that are visualized during laparoscopic surgery. Based on these morphological descriptions, the rASRM classified the extent of endometriosis into four stages: minimal, mild, moderate, and severe (from Stage I to Stage IV) 7 , 8 . Treatment options, including medication, surgery, or a combination of both, largely depend on the extent and location of the disease, severity of symptoms, and desire of the patient to become pregnant 9 . Pharmacological therapies are effective for endometriosis-related pain and are often considered the first-line treatment 10 . However, laparoscopic surgery remains a crucial therapeutic approach for endometriosis. Significantly, abdominal surgery for severe endometriosis is far more complicated due to altered pelvic anatomy and extensive adhesions. Furthermore, laparoscopic surgery for severe endometriosis is associated with a high risk of complications 11 – 14 . In 2022, the new European Society of Human Reproduction and Embryology (ESHRE) guidelines challenged laparoscopy as the gold standard diagnostic test and recommended that clinicians use imaging during the diagnostic work-up for endometriosis 13 . Thus, if severe endometriosis can be identified with preoperative clinical data before surgery, clinicians should be able to fully understand the patient’s condition and make justifiable perioperative surgery protocols and treatment decisions, such as preoperatively utilizing gonadotropin-releasing hormone (GnRH) agonists before surgery. As of now, no reliable noninvasive predictive model exists for forecasting severe endometriosis prior to surgical intervention. Despite numerous efforts, predictive models based on machine learning related to endometriosis are predominantly utilized for the diagnosis of the condition, and have not yet been applied to stages 15 – 18 . Therefore, we developed severe endometriosis risk prediction models using existing clinical and imaging data through machine learning and artificial intelligence.

Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1