Perioperative prediction of adnexal malignancy by an interpretable machine learning model for guiding ovarian preservation in premenopausal endometrial cancer

Data for this retrospective study were collected from the Third Affiliated Hospital of Sun Yat-sen University from January 2010 to August 2024.This study was conducted in accordance with Tripod guidelines and approved by the ethics committee of the Third Affiliated Hospital of Sun Yat-sen University (No. II2023-008–02). Informed consent was obtained from all patients. Inclusion Criteria: Patients diagnosed with EC confirmed by pathology. Pre-menopause. Histological type: Endometrioid adenocarcinoma. EC patients undergoing total hysterectomy with bilateral salpingo-oophorectomy with or without pelvic and para-aortic lymph node dissection. Surgery was the initial treatment, and no other treatment such as radiotherapy, chemotherapy, hormonal therapy, or targeted therapy was performed before surgery. Patients with complete clinical information and data. Patients diagnosed with EC confirmed by pathology. Pre-menopause. Histological type: Endometrioid adenocarcinoma. EC patients undergoing total hysterectomy with bilateral salpingo-oophorectomy with or without pelvic and para-aortic lymph node dissection. Surgery was the initial treatment, and no other treatment such as radiotherapy, chemotherapy, hormonal therapy, or targeted therapy was performed before surgery. Patients with complete clinical information and data. Exclusion Criteria: Patients with severe dysfunction of the heart, liver, kidney, or other major organs. Patients with other malignant tumors or conditions affecting serum tumor marker and inflammatory marker levels. Patients with severe dysfunction of the heart, liver, kidney, or other major organs. Patients with other malignant tumors or conditions affecting serum tumor marker and inflammatory marker levels. To ensure the integrity of the multimodal data, we adopted a complete case analysis strategy. Specifically, patients with missing primary predictors were excluded from the final cohort, as detailed in our inclusion criteria (f). Data were extracted from electronic medical records. Standardized variables included: Clinical information: including age, height, weight, reproductive status, history of hypertension and diabetes, personal and family history, etc.; Body mass index is calculated as the patient's weight (kg) divided by the square of the height (m). Preoperative serum examinations including white blood cell, neutrophil count, lymphocyte count, monocyte count, platelet count. Preoperative tumor markers including cancer antigen 125 and human epididymis protein 4. Imaging indicators: ovarian size, ovarian involvement, and endometrial thickness in ultrasound; myometrial invasion and ovarian involvement on magnetic resonance imaging. The pathological molecular markers that obtained from preoperative curettage: estrogen and progesterone receptors, P53, P16, Ki-67 and MMR proteins (dMMR; MLH1, MSH2, MSH6, PMS2). Intraoperative exploration indicators: whether ovarian appearance was abnormal, ascites cytology results. Clinical information: including age, height, weight, reproductive status, history of hypertension and diabetes, personal and family history, etc.; Body mass index is calculated as the patient's weight (kg) divided by the square of the height (m). Preoperative serum examinations including white blood cell, neutrophil count, lymphocyte count, monocyte count, platelet count. Preoperative tumor markers including cancer antigen 125 and human epididymis protein 4. Imaging indicators: ovarian size, ovarian involvement, and endometrial thickness in ultrasound; myometrial invasion and ovarian involvement on magnetic resonance imaging. The pathological molecular markers that obtained from preoperative curettage: estrogen and progesterone receptors, P53, P16, Ki-67 and MMR proteins (dMMR; MLH1, MSH2, MSH6, PMS2). Intraoperative exploration indicators: whether ovarian appearance was abnormal, ascites cytology results. CAM was defined as concurrent tubal or ovarian cancer change or metastasis in EC patients. Because concurrent ovarian cancer and fallopian tube or ovarian metastasis are both ovarian adverse events, it is beneficial to evaluate these two outcomes together when examining the safety of ovarian preservation. Categorical variables were encoded using ordinal transformation, while continuous variables underwent z-score normalization via the Standard-Scaler algorithm to address scale-dependent heterogeneity. From an initial pool of 30 candidate predictors, we implemented a two-stage feature selection protocol: 1. Multicollinearity Elimination; 2. Least Absolute Shrinkage and Selection Operator regression. To mitigate the issue of class imbalance between two groups, we implemented a comprehensive data preprocessing strategy combining the Synthetic Minority Oversampling Technique. The dataset was strategically partitioned through random stratified sampling into a training set (80%, n  = 237) and a validation set (20%, n  = 59). The training set was used for model development, while the validation set served as an independent cohort for performance evaluation. To ensure a comprehensive evaluation of predictive performance across diverse algorithmic approaches, we implemented eight distinct machine learning classifiers, each representing a unique computational framework for pattern recognition in clinical data. The selected algorithms encompassed both traditional statistical learning models and advanced ensemble techniques, including Random Forest (RF), e-Xtreme Gradient Boosting (XG-Boost), Support Vector Machine (SVM), Logistic Regression (LR), Gradient Boosting Decision Tree (GBDT), Adaptive Boosting (AdaBoost), Multilayer Perceptron (MLP), and Naïve Bayes (NB). This selection facilitated a robust comparison of models with varying capacities to capture linear and non-linear relationships within multimodal clinical data. Each algorithm was configured with optimized hyperparameters through a systematic grid search coupled with 5-fold cross-validation to mitigate overfitting and enhance model generalizability. The model development process adhered to fundamental signal processing principles by treating multimodal clinical features as input signals subjected to feature scaling and dimensionality reduction prior to model training. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC-ROC) as the primary metric for discriminative ability. The Brier score was employed to assess calibration, quantifying the degree of deviation between predicted probabilities and actual outcomes, with lower values indicating better alignment. To enhance clinical interpretability, SHAP values were computed for the optimal model, enabling quantitative assessment of feature contributions and providing insights into the relationship between input variables and prediction outcomes. This rigorous methodological framework ensures that the resulting predictive tool not only achieves high analytical accuracy but also maintains computational efficiency and clinical interpretability, thereby supporting its potential integration into perioperative decision-support systems. Continuous variables are reported as medians with interquartile ranges. Categorical variables are reported as numbers and proportions (n [%]). According to the normality test results, Mann Whitney U test was used to compare the groups. Categorical variables were compared by Chi square test and Fisher’s exact test. All statistical analyses were performed using IBM SPSS Statistics 22 (SPSS Inc., Chicago, IL, USA). The predictive model construction and graphical representations were implemented using Python V3.7 (Python Software Foundation) and Prism 10.0 (GraphPad Software, San Diego, CA, USA), respectively.

A retrospective cohort of premenopausal patients diagnosed with EC was identified between January 1, 2010, and August 31, 2024. 296 eligible cases (9.2% of initial cohort) were ultimately included for final analysis (Fig.  1 ). Fig. 1 Flow chart for model development and validation Flow chart for model development and validation The study cohort included 296 premenopausal patients with endometrial endometrioid carcinoma. The median age was 48 years [IQR:45–51], the youngest patient in the cohort was 29 years old; overall, 91 patients (30.7%) were aged ≤ 45 years, and 36 (12.2%) were aged < 40 years. The median BMI was 23.6 kg/m 2 [IQR: 21.6–26.0]. Preoperative imaging identified CAM by ultrasonography and 6 by MRI, with tumor sizes ranging from 2 to 8 cm. Notably, 20 patients (69.0%) with normal or benign-appearing adnexal lesions on preoperative ultrasonography and MRI were confirmed to have malignant adnexal involvement on postoperative pathology. The imbalance in tumor staging between the two groups is because the presence of CAM would classify the case as stage III or above. Consequently, among patients with pathologically confirmed CAM, none were staged as I or II (Table  1 ). Table 1 Baseline characteristics of the patients Variable CAM p -value Positive ( n  = 29) Negative ( n  = 267) Total Age at diagnosis (years) 48(43–51) 49(45–51) 0.40 BMI (kg/m2) 23.8(20.8–26) 23.5(21.7–26.3) 0.45 Gravidity 2(1–3) 2(2–3) 0.11 Parity 1(1–2) 2(1–2) 0.14 Family history of cancer (n, %) 0.72  Positive 3(10.3) 23(8.6)  Negative 26(89.7) 244(91.4) Diabetes (n, %) 0.21  Positive 5(17.2) 26(9.7)  Negative 24(82.8) 241(90.3) Hypertension (n, %) 0.19  Positive 2(6.9) 47(17.6)  Negative 27(93.1) 220(82.4) Endometrial thickness in US (mm) 15(11.5–25.5) 16(11.1–22) 0.91 Adnexal involvement in US < 0.001  Positive 4(13.8) 1(0.4)  Negative 25(86.2) 266(99.6) Adnexal involvement in MRI  50% 3(10.3) 22(8.2) Stage < 0.001  I 0 243  II 0 16  III 16 18  IV 1 2 Abbreviations: CAM Coexisting adnexa malignancy, BMI body mass index, US Ultrasound, MRI Nuclear Magnetic Resonance Imaging Among the 29-coexisting adnexal malignancy -positive patients (9.8%), 10 (34.5%) were aged < 45 years. Ovarian involvement was observed in 19 cases (65.5%), isolated fallopian tube involvement in 8 cases (27.6%), and concurrent ovarian and fallopian tube involvement in 2 cases (6.9%) Baseline characteristics of the patients Abbreviations: CAM Coexisting adnexa malignancy, BMI body mass index, US Ultrasound, MRI Nuclear Magnetic Resonance Imaging Among the 29-coexisting adnexal malignancy -positive patients (9.8%), 10 (34.5%) were aged < 45 years. Ovarian involvement was observed in 19 cases (65.5%), isolated fallopian tube involvement in 8 cases (27.6%), and concurrent ovarian and fallopian tube involvement in 2 cases (6.9%) Elevated serum Cancer Antigen 125 and Human Epididymis Protein 4 levels were detected in 74 (25%) and 108 (36.5%) patients, respectively (Table  2 ). Table 2 Serum markers and pathological features of the patients Variable Coexisting adnexal malignancy p -value Positive ( n  = 29) Negative ( n  = 267) Histological grade 0.18  1 14(48%) 174(65.2%)  2 11(37.9%) 68(25.5%)  3 4(13.8%) 25(9.3%) CA 125  35U/ml 18(62.1%) 56(21%) HE-4 0.041  ≤ 70 pmol/L 13(44.8%) 175(65.5%)  > 70 pmol/L 16(55.2%) 92(34.5%) WBC 7.12(6.17,9.97) 6.32(5.32,7.98) 0.009 NLR 2.74(2.14,3.98) 2.21(1.69,2.91) 0.003 PLR 180.7(141.8,236.2) 161.5(130.4,222.1) 0.25 MLR 0.24(0.19,0.35) 0.23(0.17,0.31) 0.15 SII 941.4(642.7,1406) 644.9(461.2,954.7) 0.002 SIRI 1.36(0.74,1.94) 0.86(0.57,1.4) 0.005 Appearance of the ovaries (n, %) < 0.001  Positive 13(44.8%) 21(7.9%)  Negatives 16(55.2%) 246(92.1%) Peritoneal washings (n, %) 0.013  Positive 5(17.2%) 11(4.1%)  Negative 24(82.8%) 256(95.9%) ER expression  < 0.001  - 0(0%) 4(2.5%)  Low 10(45.5%) 17(10.7%)  High 12(54.5%) 138(86.8%) PR expression 0.013  - 1(4.5%) 6(3.8%)  Low 8(36.4%) 20(12.7%)  High 13(59.1%) 132(83.5%) P53 0.86  - 1(4.5%) 8(5.4%)  + 21(95.5%) 139(94.6%) Ki-67 0.78  - 0(0%) 2(1.2%)  Low 3(13%) 17(10.6%)  High 20(87%) 142(88.2%) dMMR phenotype*  Suspected 2(22.2%) 19(25.3%) 0.84  Negative 7(77.8%) 56(74.7%) Abbreviations: CA-125 cancer antigen 125, HE-4 Human Epididymis Protein 4, ER estrogen receptor, PR progesterone receptor, WBC white blood cell, NLR neutrophil count/lymphocyte count, MLR monocyte count/lymphocyte count, PLR platelet count/lymphocyte count, SII platelet count × neutrophil count/lymphocyte count, SIRI neutrophil count × monocyte count/lymphocyte count * dMMR phenotype includes all mismatch repair system components expression, that is, MLH1: MutL homolog 1; MSH2: MutS homolog 2; MSH6: MutS homolog 6 gene; and PMS2: PMS1 homolog 2 Serum markers and pathological features of the patients Abbreviations: CA-125 cancer antigen 125, HE-4 Human Epididymis Protein 4, ER estrogen receptor, PR progesterone receptor, WBC white blood cell, NLR neutrophil count/lymphocyte count, MLR monocyte count/lymphocyte count, PLR platelet count/lymphocyte count, SII platelet count × neutrophil count/lymphocyte count, SIRI neutrophil count × monocyte count/lymphocyte count * dMMR phenotype includes all mismatch repair system components expression, that is, MLH1: MutL homolog 1; MSH2: MutS homolog 2; MSH6: MutS homolog 6 gene; and PMS2: PMS1 homolog 2 Immunohistochemical analysis revealed ovarian metastasis in 45.5% of estrogen receptor-negative/low-expression cases and 40.9% of progesterone receptor-negative/low-expression cases. Comparison of the pathological features between the group with or without coexisting adnexal malignancy showed significant differences in the appearance of the ovaries( P  < 0.001), peritoneal washings( P  = 0.013), estrogen receptor ( P  = 0.013) and progesterone receptor ( P  < 0.001). The following inflammatory markers were also significantly elevated in the CAM group: White Blood Cell count( P  = 0.009), Neutrophil to Lymphocyte Ratio ( P  = 0.003), Systemic Immune-inflammation Index ( P  = 0.002) and Systemic Inflammatory Response Index ( P  = 0.005). From the initial pool of 30 candidate predictors, 4 features were excluded due to multicollinearity (Fig.  2 ). Then after Least Absolute Shrinkage and Selection Operator regression (Fig.  3 ), 16 features were selected for inclusion in the machine learning model: four clinical characteristics (BMI, Gravidity, Parity, Family history of cancer), two imaging characteristics (Endometrial thickness measured on ultrasound, Adnexal involvement in MRI), four serum biomarkers (Cancer Antigen 125, Human Epididymis Protein 4, White Blood Cell count, Platelet-to-Lymphocyte Ratio) and six histopathological features (Appearance of the ovaries, Peritoneal washings, P16, P53, estrogen receptor/progesterone receptor status). Fig. 2 Multicollinearity analysis of predictors Fig. 3 LASSO regression analysis of predictors Multicollinearity analysis of predictors LASSO regression analysis of predictors The predictive performance of these selected features was evaluated using eight machine learning model. To provide a comprehensive assessment, we reported multiple complementary metrics: AUC for overall discriminative ability, accuracy for overall correctness, recall (sensitivity) for the ability to detect true positive cases, specificity for identifying true negatives, F1 score as a balanced measure accounting for class imbalance, and Brier score for calibration. As detailed in Table  3 , the Naive Bayes model demonstrated superior discriminative performance. Subsequent use validation set confirmed the model's performance. The ROC curves for all models are shown in Fig.  4 . Table 3 Predictive performances of the eight ML models for predicting Model AUC (95%CI) Accuracy Recall Specificity F1 Score Brier RF 0.80(0.71,0.88) 0.68 0.55 0.9 0.67 0.22 XGBoost 0.81(0.73,0.90) 0.82 0.72 0.88 0.78 0.27 SVM 0.80(0.72,0.89) 0.85 0.79 0.83 0.8 0.2 LR 0.77(0.67,0.86) 0.67 0.64 0.88 0.72 0.29 GBDT 0.78(0.70,0.87) 0.66 0.74 0.83 0.77 0.28 AdaBoost 0.72(0.62,0.82) 0.74 0.69 0.79 0.72 0.23 MLP 0.78(0.70,0.87) 0.75 0.53 0.9 0.64 0.36 Naive Bayes 0.92(0.86,0.97) 0.91 0.83 0.89 0.85 0.11 Fig. 4 ROC curves for eight ML models Predictive performances of the eight ML models for predicting ROC curves for eight ML models To elucidate feature contributions and interpret model predictions, we implemented SHAP. SHAP summary plots (Figs.  5 ; 6 ) illustrate feature contribution magnitudes and directions. In the optimal performing Naive Bayes model, the top five predictive features for ovarian preservation were that Cancer Antigen 125 (2.08), Estrogen Receptor (1.58), Human Epididymis Protein 4 (0.8), White Blood Cell count (0.8) and Platelet-to-Lymphocyte Ratio (0.72) as the predominant predictive features. Fig. 5 Importance score ranking of the characteristics Fig. 6 SHAP summary plots Importance score ranking of the characteristics SHAP summary plots

This study developed and validated a machine learning-based predictive model for CAM in premenopausal endometrial cancer patients, integrating 16 clinically accessible predictors from multimodal data sources. Through systematic comparison of eight algorithms, the Naïve Bayes classifier demonstrated superior performance (AUC: 0.92, Brier score: 0.11), achieving high discriminative accuracy while maintaining robust calibration. SHAP interpretability analysis further quantified feature contributions, identifying Cancer Antigen 125, Estrogen Receptor status, Human Epididymis Protein 4, White Blood Cell count, and Platelet-to-Lymphocyte Ratio as key determinants of CAM risk. This model represents a clinically translatable decision-support tool that can enhance perioperative risk stratification for ovarian preservation candidates. The management of young patients with EC is challenging. The NCCN guidelines [ 4 ] allow ovarian preservation to be considered in premenopausal women with early stage, ascitic cell-negative and normal intraoperative ovarian appearance. The European ESGO-ESTRO-ESP Guidelines for EC [ 7 , 8 ] recommend ovarian preservation in premenopausal women with ⅠA, G1, or EC. However, this indication still has a blind spot: In our study, 7 of 29 patients with CAM were fully met the indications of ovarian preservation in the guidelines. Walsh et al. [ 9 ] similarly reported 25% (26/102) CAM in young EC patients-15% had normal preoperative imaging, and 15% had normal appearance of the ovaries appendages during surgery. Critically, the 5-year overall survival rate was significantly lower in patients with adnexal involvement (69.6% vs 94.4%, p  < 0.001) [ 1 ], there is an urgent need for objective, data-driven tools to improve preoperative risk assessment. Current preoperative imaging exhibits inadequate sensitivity for CAM detection [ 3 ] and are therefore insufficient to establish an indication for ovarian preservation. Existing predictive models [ 3 , 10 ] relying on Logistic Regression and nomograms suffer from limited nonlinear modeling capacity, feature selection bias, and overfitting vulnerabilities. We tried to find a more accurate, scientific and convenient way to making CAM risk assessment and surgical planning [ 11 ]. In this study, we explored 30 variables that were assessed by perioperative examination and intraoperative findings. Easy access to data facilitates clinical applications. We applied eight machine learning algorithms to fit each of the 16 predictor variables individually and evaluated the discriminative features of each model by benchmarking to determine the best performance model for predicting CAM. Compared with traditional statistical methods, machine learning algorithms can directly capture nonlinear and complex interactions without the need for multi-step statistical analysis, reducing the subjectivity of artificial feature selection. In our study, the Naïve Bayes algorithm had the best prediction performance (AUC, 0.92; Brier score, 0.11). This algorithm does not require any ex-ante assumptions on data allocation, which makes the algorithm applicable to data at any point in time, enabling real-time clinical implementation [ 6 ]. In parallel, Macis et al. developed a hybrid convolutional neural network and machine learning pipeline using contrast‑enhanced CT to differentiate endometriosis‑associated ovarian cancer. While their work focuses on ovarian cancer subtyping and ours on predicting concurrent adnexal malignancy in premenopausal endometrial cancer, both studies underscore the growing role of interpretable, image‑based AI models in guiding personalized surgical decisions in gynecologic oncology [ 12 ]. SHAP analysis identified several key predictors, including Cancer Antigen 125, Estrogen Receptor status, Human Epididymis Protein 4, White Blood Cell count, and Platelet-to-Lymphocyte Ratio, that significantly influence concurrent adnexal malignancy risk. Notably, these biomarkers are not systematically incorporated into current ovarian preservation guidelines, highlighting a critical gap in existing clinical decision frameworks. Beyond traditional markers such as white blood cells and platelets, emerging indices including the Neutrophil to Lymphocyte Ratio, Platelet-to-Lymphocyte Ratio, as well as the more recently developed Systemic Immune-inflammation Index [ 13 ] and [ 14 ] have been recognized as valuable indicators of systemic inflammatory response [ 15 ]. Previous studies have confirmed that these inflammatory markers and lymphatic metastasis of endometrial cancer are associated with poor prognosis [ 16 ]. The inclusion of inflammatory markers enhances the biological plausibility of our model, as these signals reflect systemic immune activation and tumor microenvironment activity [ 17 ]. Cancer antigen 125 is an epitope on a high molecular weight glycoprotein recognized by monoclonal antibodies, and it has been used as a tumor marker for ovarian cancer for several decades since its discovery. A significant proportion (80%) of women with primary epithelial ovarian cancer and secondary ovarian tumors (70%) are diagnosed based on elevated cancer antigen 125 levels [ 18 ]. The widespread use of cancer antigen 125 as a tumor marker for screening and monitoring in the treatment of ovarian cancer suggests the possibility of a specific correlation between cancer antigen 125 and ovarian metastasis. Despite its clinical relevance, current NCCN and ESMO guidelines do not fully integrate these biomarkers into ovarian preservation criteria. From a biomedical signal processing perspective, these easily accessible blood biomarkers can be used as effective biosignal inputs to provide a complementary dimension to risk assessment. From a clinical signal processing perspective, this study demonstrates how heterogeneous biomedical data including serum biomarkers, histopathological features, and inflammatory indices can be fused into a unified predictive framework. The model functions as an intelligent diagnostic support system, converting multidimensional preoperative signals into a quantifiable risk probability. This approach aligns with emerging trends in signal fusion for personalized oncology, advancing beyond single-modality assessment. The validated Naive Bayes model achieves high discriminative accuracy (AUC 0.92) in identifying patients at risk of concurrent adnexal malignancy. Integration of this tool into perioperative workflows may significantly enhance surgical decision-making, enabling personalized treatment plans that balance oncological safety with ovarian preservation. To our knowledge, this represents the first machine learning-based predictive model specifically developed for quantifying the risk of concurrent adnexal malignancy in premenopausal endometrial cancer patients. A key strength lies in the integration of multimodal, preoperative clinical signals that spanning serum biomarkers, imaging parameters, and histopathological features into an interpretable computational tool. The application of SHAP enhances model transparency by identifying the relative contribution of each feature, thereby facilitating clinical trust and adoption. Furthermore, the use of the Naïve Bayes algorithm ensures computational efficiency and suitability for real-time clinical implementation. Several limitations should be acknowledged. This was a single-center retrospective study, which may limit the generalizability of the model. In addition, the analysis did not include long-term monitoring for metachronous ovarian malignancies following ovarian preservation. Future prospective studies should focus on external validation and longitudinal follow-up to evaluate the model’s performance in predicting long-term oncological outcomes.

We present a machine learning-driven tool that effectively translates preoperative biomedical signals into a personalized CAM risk estimate. By combining predictive accuracy with model interpretability, this system supports safer, more individualized surgical planning and represents a meaningful advance in the convergence of clinical medicine and intelligent signal processing.

Endometrial cancer(EC) is one of the most common malignant tumors of the female reproductive tract, with a steadily increasing incidence worldwide [ 1 ]. Particularly concerning is the incidence rise among young women, escalating from 2.9% to 18.7% between 1981 and 2020 [ 2 ]. This demographic transition has been closely associated with escalating rates of obesity, metabolic syndrome prevalence, and delayed childbearing patterns in younger populations. While total hysterectomy with bilateral salpingo-oophorectomy remains the standard treatment, surgical menopause induces a series of estrogen-deficiency disorders, including accelerated osteoporosis, vasomotor symptoms and heightened cardiovascular mortality [ 3 ]. Consequently, young EC patients express demands for ovarian preservation. This paradigm shift aligns with evolving National Comprehensive Cancer Network (NCCN) guidelines recommendations emphasizing fertility-sparing considerations in premenopausal EC management [ 4 ], necessitating precision surgical stratification protocols to reconcile oncological safety with endocrine/quality-of-life outcomes. However, ovarian preservation entails a critical clinical trade-off: while avoiding premature menopause, it carries a 5–25% risk of concurrent adnexal malignancy (CAM) [5], with premenopausal patients facing a 3.17-fold higher risk compared to postmenopausal counterparts [ 5 ]. The risk of potentially occult tumors is worse than for surgical menopause, so the decision to preserve the ovaries must be made carefully (9), and it is important to accurately predict the risk of CAM before surgery. Current CAM detection primarily relies on imaging methods, but its diagnostic accuracy remains suboptimal and there is a certain rate of missed diagnosis. Existing prediction models face some limitations: 1) depended on postoperative histopathological parameters (e.g., myometrial invasion depth, lymph vascular space invasion), rendering them inapplicable for preoperative decision-making; 2) Conventional logistic regression fails to capture complex nonlinear interactions among multivariate predictors; 3) No validated tools specifically address the unique risk profile of premenopausal patients for ovarian preservation. Machine learning (ML) offers a promising alternative by leveraging high-dimensional data to improve predictive accuracy and support clinical decision-making. ML algorithms have demonstrated superior performance across various medical domains, including oncology, by integrating diverse data sources such as clinical records, imaging signals, and biomarker profiles [ 6 ]. Yet no dedicated ML tool currently exists for perioperative CAM risk stratification in premenopausal EC patients. To address this gap, we developed an interpretable ML-based decision support system that integrates multimodal perioperative features to quantitatively assess CAM risk. By employing model explanation techniques such as SHAP, our approach not only delivers high predictive accuracy but also provides clinically transparent insights into feature contributions. This study aims to facilitate personalized surgical planning through a computationally robust and clinically actionable tool aligned with the emerging paradigm of precision gynecologic oncology.