Comparison Between Cox Proportional Hazards and Machine Learning Models for the Prognostication of Recurrence and Survival Following Liver Resection for Hepatocellular Carcinoma

ABSTRACT Background A robust prognostication model after liver resection for hepatocellular carcinoma (HCC) can guide clinical management. We aimed to develop a prognostication model for HCC recurrence and survival following liver resection, comparing between Cox proportional hazards (CPH) and supervised machine learning models. Methods We studied all patients who underwent liver resection for HCC between January 1, 2000 and October 31, 2022 at our institution. We aimed to predict recurrence‐free survival following resection and identify risk categories for HCC recurrence. The CPH model and two supervised machine learning models (random survival forest [RSF] and extreme gradient boosting [XGB]) were used. Model performance was assessed with C‐index, time‐dependent area under curve (tdAUC) and Brier score. Results We studied 1290 patients, with 737 (57.1%) experiencing an event (HCC recurrence or death) over a median follow‐up duration of 19.2 months. The CPH model had the overall best performance (C‐index: 0.663, tdAUC at 6 months: 0.752; 1 year: 0.740; 2 years: 0.722; 5 years: 0.624). Using this model, patients stratified based on risk score could be discriminated between low, intermediate, and high‐risk groups ( p < 0.001). Conclusion A CPH‐derived prognostication model was effective for predicting and risk stratifying recurrence and survival following liver resection for HCC.