Correlation analysis and recurrence evaluation system for patients with recurrent hepatolithiasis: a multicentre...

Area of Science:

• Hepatobiliary surgery
• Medical artificial intelligence
• Clinical prognosis

Background:

• Accurate prognosis prediction for recurrent hepatolithiasis (RH) post-biliary surgery is limited.
• Existing methods lack dynamic risk assessment capabilities.
• Need for advanced predictive models incorporating complex clinical data.

Purpose of the Study:

• Develop a machine learning (ML) model for dynamic risk prediction of RH recurrence.
• Utilize high-order correlation data for improved accuracy.
• Create a user-friendly tool for clinical decision support.

Main Methods:

• Collected data from RH patients across five centers (2015-2020).
• Developed nine ML models, including Extreme Gradient Boosting and LightGBM, forming the CARES system.
• Employed k-fold cross-validation and a separate testing set for robust evaluation.
• Utilized Shapley Additive Explanations (SHAP) for variable importance interpretation.

Main Results:

• ML models significantly outperformed traditional regression for RH recurrence prediction.
• XGBoost and LightGBM achieved an Area Under the ROC Curve (AUC) > 0.9.
• Models demonstrated strong performance on testing sets, indicating no overfitting.
• Key predictors identified: immediate/final stone clearance, prior surgeries, and preoperative CA19-9.

Conclusions:

• The CARES model, based on ML, effectively predicts RH recurrence post-hepatectomy.
• An online version of CARES facilitates clinical decision-making and personalized surveillance.
• The study provides a valuable tool for managing patients with recurrent hepatolithiasis.