ARTIFICIAL INTELLIGENCE FOR SURVIVAL PREDICTION IN HEPATOCELLULAR CARCINOMA: DEVELOPMENT AND VALIDATION OF A CLINICAL DATA–DRIVEN MODEL IN A COHORT OF 129 PATIENTS

Abstract

Introduction and Objectives

To develop and validate a predictive survival model for patients with hepatocellular carcinoma (HCC) associated with metabolic dysfunction–associated steatotic liver disease (MASLD), using artificial intelligence applied to widely available clinical and laboratory data. Additionally, to compare the model’s performance with traditional prognostic scores commonly used in HCC risk stratification.

Materials and Methods

This retrospective study included 129 patients with confirmed HCC and underlying MASLD. Clinical, laboratory, and tumor-related variables were analyzed, including metabolic comorbidities, liver function markers, tumor burden, cirrhosis-related complications, and established prognostic scores (Child-Pugh, FIB-4, and ALBI). The predictive model was built using Cox proportional hazards regression with L2 regularization to manage high-dimensional data and minimize overfitting. The XGBoost (Extreme Gradient Boosting) algorithm was implemented, with random allocation of the dataset into a training cohort (80%) and an internal validation cohort (20%). DeepSurv, a deep learning–based survival model, was also explored as a complementary strategy.

Results

The regularized Cox model demonstrated robust predictive performance, achieving a concordance index (C-index) of 0.774 in the validation cohort. The variables most strongly associated with reduced survival included tumor thrombosis (HR 8.27), hepatic encephalopathy (HR 4.66), and spontaneous bacterial peritonitis (HR 6.51), all statistically significant. The proposed model outperformed widely used prognostic scores such as BCLC, CLIP, and ALBI, showing superior discriminative ability for survival prediction in patients with HCC-MASLD.

Conclusions

The AI-based model, built using easily accessible clinical and laboratory data, demonstrated superior performance in predicting survival in patients with HCC-MASLD. This approach enables more precise and scalable risk stratification, with direct applicability in real-world clinical practice.