Related Experiment Video
Updated: May 19, 2026
Comparison of Predictive Performance of Three Lymph Node Staging Systems in Colorectal Signet Ring Cell Carcinoma Based on Machine Learning Model
Published on: April 18, 2025
Development of an interpretable machine learning model for lymphovascular space invasion prediction in patients with
Meixuan Wu1, Ling Zhou1, Xiao Yang1
1Department of Obstetrics and Gynecology, Peking University People's Hospital, Beijing 100044, China.
Objective:
Lymphovascular space invasion (LVSI) is a high-risk factor for lymph node metastasis, relapse, and poor prognosis in patients with endometrioid endometrial carcinoma (EEC). However, the diagnosis of LVSI still relies on traditional pathological methods. Moreover, the high-risk factors and mechanism for LVSI remain unclear. Thus, this study developed an interpretable machine learning (ML) model to accurately predict LVSI status in patients with EEC.
Methods:
The study collected data from 832 patients with EEC at Peking University People's Hospital. Patients were randomly divided into training (n=582) and internal validation (n=250) cohorts. A prospective external validation cohort included 129 patients with EEC from Nanjing Drum Tower Hospital. Using 21 parameters, 6 ML strategies were used to build prediction models. The global and local interpretation of feature significance was performed using the SHapley Additive exPlanations (SHAP) approach. Data from NanoString nCounter evaluation was subjected to pathway enrichment and Spearman correlation analysis to investigate the mechanistic basis of LVSI.
Results:
Among the six ML models, the XGBoost model had the best performance. The XGBoost model correctly predicted the risk of LVSI in the training set [area under the curve (AUC): 0.982, 95% confidence interval (95% CI): 0.972-0.991], the internal validation set (AUC: 0.818, 95% CI: 0.776-0.860), and the external test set (AUC: 0.748, 95% CI: 0.618-0.879). The calibration curve indicated that the XGBoost model exhibited favorable consistency between the predicted and actual risks. SHAP analysis identified age, carbohydrate antigen 125 (CA125), low-density lipoprotein (LDL), and neutrophil as the top four variables contributing to XGBoost model predictions. Analysis of NanoString data indicated that LVSI may be closely associated with the PI3K-Akt signaling pathway.
Conclusions:
We developed an interpretable ML model for preoperative LVSI risk prediction in patients with EEC. This model may aid clinicians by informing individualized clinical decision-making.
