Explainable Machine Learning Integrating Metabolic and Inflammatory Signatures for Personalized Prognosis in Resected Intrahepatic Cholangiocarcinoma
Menée à partir de données portant sur 690 patients atteints d'un cholangiocarcinome intrahépatique traité par résection, cette étude évalue la performance de modèles d'apprentissage automatique, intégrant des facteurs clinicopathologiques, métaboliques et immuno-inflammatoires, pour prédire la survie sans maladie et la survie globale
Background : The precise prognostic stratification of intrahepatic cholangiocarcinoma (iCCA) remains challenging. We aimed to develop and validate interpretable machine learning (ML) models that integrate clinicopathological, metabolic, and immune-inflammatory factors to personalize prognosis prediction.
Methods : We retrospectively collected data from 690 iCCA patients across five centers. Patients from four centers were assigned to training/testing sets (n=597, 7:3 split), and another single center as an external validation set (n=93). After feature selection, five survival models were developed and compared for predicting overall survival (OS) and disease-free survival (DFS) using the concordance index (C-index), time-dependent ROC, Kaplan-Meier, calibration and decision curves. SHapley Additive exPlanations (SHAP) interpreted predictions, and a clinically applicable web-based tool was developed.
Results : The survival support vector machine (SSVM) model achieved the best predictive performance for both OS and DFS prediction. The SSVM_OS model achieved a C-index of 0.754, and the SSVM_DFS model achieved a C-index of 0.709. Both models showed excellent performance in the external validation set and demonstrated good clinical utility. The models effectively stratified patients into distinct risk groups and outperformed the AJCC-TNM staging system. SHAP analysis identified gamma-glutamyl transferase, triglyceride-glucose index, lymph node metastasis, and carcinoembryonic antigen as the most influential predictors for both overall survival and disease-free survival. The optimal models were deployed as an online tool to provide individualized risk estimates for death and recurrence, supporting clinical decision-making.
Conclusions : We developed and externally validated explainable ML models to predict postoperative risk for iCCA patients. The best-performed SSVM models were implemented as a clinical decision-support tool to guide personalized surveillance.
European Journal of Surgical Oncology , article en libre accès, 2026