Machine Learning Integration of Lymphovascular Space Invasion, Ki-67 Index, and Immunohistochemical Profiles for Preoperative Lymph Node Metastasis Prediction in Endometrial Cancer

Technical Summary: Machine Learning Integration for Preoperative Lymph Node Metastasis Prediction in Endometrial Cancer

Problem Statement
Endometrial cancer (EC) is a leading gynecological malignancy. Accurate preoperative prediction of lymph node metastasis (LNM) is critical for individualized surgical decision-making, specifically to avoid the overtreatment and associated complications (e.g., lymphedema, intestinal obstruction) of non-selective systematic lymph node dissection in low-risk patients. Current clinical prediction models often rely on postoperative pathological indicators (e.g., tumor grade, depth of invasion) or complex genomic data, creating a fundamental disconnect where "postoperative indicators guide preoperative decisions." Furthermore, existing models frequently omit key prognostic markers such as Ki-67 or fail to comprehensively integrate routine serological and immunohistochemical profiles available prior to surgery.

Methodology
This study employed a retrospective design analyzing 594 patients with endometrial carcinoma treated at Hainan General Hospital between October 2019 and October 2025.

• Data Collection: Variables included clinical baseline data, preoperative serological markers (AFP, CEA, CA125, CA19-9, CA72-4, etc.), postoperative pathological indicators (myometrial invasion, LVSI, tumor size, etc.), and immunohistochemical profiles (p53, ER, PR, Ki-67, MMR proteins).
• Feature Selection: Least Absolute Shrinkage and Selection Operator (LASSO) regression was utilized to screen variables and prevent overfitting. This process identified seven key predictors.
• Model Construction: Four machine learning algorithms were constructed and compared: Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and XGBoost.
• Validation Strategy: The dataset was split into a development set (n=416, 70%) and a test set (n=178, 30%). Model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC), sensitivity, specificity, calibration curves, and Decision Curve Analysis (DCA).
• Robustness Checks: Subgroup analyses were performed across various clinical characteristics. A sensitivity analysis was conducted by excluding p53-mutated cases to assess model stability.

Key Contributions

1. Preoperative Focus: The study developed a prediction model exclusively utilizing indicators available preoperatively, addressing the limitation of models relying on postoperative data.
2. Multimodal Integration: The model systematically integrates routine pathological parameters (LVSI), immunohistochemical markers (PR), and serological biomarkers (CA19-9, CA72-4). While Ki-67 was evaluated, it was not retained as an independent predictor in the final multivariate model.
3. Algorithmic Comparison: The research provides a comparative analysis of simple (Logistic Regression) versus complex (RF, XGBoost, SVM) machine learning models, demonstrating that simpler models may offer superior generalization in medical datasets with limited sample sizes.
4. Clinical Utility Assessment: Beyond statistical metrics, the study employs Decision Curve Analysis to quantify the net clinical benefit, ensuring the model's practical applicability in decision-making.

Results

• Predictors: LASSO regression identified seven independent predictors: preoperative CA19-9, CA72-4, myometrial invasion ≥1/2, total number of lymph nodes removed, LVSI, PR expression, and preoperative CA125. Notably, while CEA was significant in univariate analysis, it did not retain independent predictive significance in the multivariate analysis and was not among the final seven variables selected by LASSO.
• Model Performance: In the independent test set, the Logistic Regression model demonstrated superior performance with an AUC of 0.863 (95% CI: 0.788–0.938), sensitivity of 0.842, and specificity of 0.761.
  • This outperformed the Random Forest model (AUC = 0.848), XGBoost (AUC = 0.839), and SVM (AUC = 0.635).
  • The Logistic Regression model also surpassed established benchmarks: the ENDORISK system (AUC = 0.820) and Yang's stratification system (AUC = 0.845).
• Feature Importance: LVSI was identified as the most critical feature, followed by the total number of lymph nodes dissected, PR expression, and preoperative CA19-9.
• Robustness: Subgroup analysis confirmed consistent performance (AUC ≥ 0.8) across various menopausal, LVSI, and FIGO stage subgroups. After excluding p53-mutated cases, the test set AUC remained stable at 0.817.
• Clinical Utility: Decision Curve Analysis indicated that the model provides a significant net clinical benefit over the "treat all" or "treat none" strategies within threshold probabilities of 0.02 to 0.60.

Significance and Claims
The authors claim that a logistic regression model based on routine, preoperative indicators can effectively predict lymph node metastasis in endometrial cancer. The study posits that:

• Simplicity vs. Complexity: In medical datasets with limited samples, increasing model complexity does not necessarily enhance predictive performance and may lead to overfitting. The simpler Logistic Regression model demonstrated better generalization than complex ensemble methods.
• Cost-Effectiveness: The model offers an economical and accessible tool for preoperative decision-making, suitable for medical institutions at all levels, without requiring expensive genomic sequencing or postoperative data.
• Novel Marker Utility: The study highlights the independent predictive value of preoperative CA19-9 and CA72-4, suggesting they may identify aggressive subgroups prone to early lymphatic metastasis, potentially outperforming CA125 in specific contexts.
• Clinical Translation: By relying solely on preoperative data, the model resolves the "postoperative guide preoperative" paradox, providing a reliable foundation for personalized surgical planning.

The paper concludes that while the model shows high predictive accuracy and robustness, future work should focus on multicenter prospective validation and the development of user-friendly nomograms or online calculators to facilitate clinical adoption.