Related Experiment Video
Updated: Jun 24, 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 and validation of a machine learning-based postoperative staging system for hilar cholangiocarcinoma: a
Qingxiang Gao1, Yingjun Wu1, Gen Li2
1Department of Biliary Tract Surgery I, Eastern Hepatobiliary Surgery Hospital, 225 Changhai Road, Yangpu District, Shanghai, 200438, China.
Background:
Hilar cholangiocarcinoma (HCCA) is a malignant tumor at the hepatic hilum with poor prognosis and survival despite surgical resection. While existing staging systems, including AJCC, Gazzaniga T, MSKCC and Blumgart, provide prognostic frameworks, their predictive accuracy remains suboptimal. This study aimed to develop and validate a novel machine learning-based postoperative staging system to predict overall survival (OS) and enhance prognostic stratification for HCCA patients undergoing surgery.
Methods:
We retrospectively analyzed 426 HCCA patients who underwent surgical resection at Eastern Hepatobiliary Surgery Hospital (EHBH) between February 2011 and May 2022. Seventy clinical variables, encompassing demographic characteristics, laboratory parameters, and imaging data, were evaluated. Eight machine learning algorithms were employed for feature selection and model development. A multivariable Cox regression analysis was subsequently performed to establish a staging system. External validation was performed using an independent cohort of 88 HCCA patients for further evaluation.
Results:
The final model incorporated key prognostic factors including tumor stage, vascular invasion, and tumor biomarkers (CA199,CA125,CEA). The CatBoost algorithm demonstrated superior performance on the validation set (AUC = 0.88). The developed EHBH-HCCA staging system based on CatBoost stratified patients into four risk levels, significantly differentiating 3-year survival outcomes (P < 0.001). Median survival times were: Level I (> 36 months), Level II (25.0 months), Level III (18.0 months), and Level IV (< 12 months). The system showed superior discrimination (Harrell's C-index: 0.717, 95% CI: 0.662-0.793) and predictive accuracy (Brier score: 0.144) compared with conventional staging systems. In the external validation cohort, the CatBoost model demonstrated good discriminative performance for 3-year survival prediction, with an AUC of 0.84. For survival time prediction, the TabNet model achieved an RMSE of 4.71 months, MAE of 3.49 months, and R² of 0.53.
Conclusions:
The EHBH-HCCA staging system provides enhanced prognostic stratification for HCCA patients following surgical resection. This novel approach offers potential for improved clinical decision-making and personalized treatment regimen.
