CATS-BIT-ME Score: A Machine Learning and Conventional Risk Model for Prognosis in Hepatocellular Carcinoma

The prognosis of hepatocellular carcinoma (HCC) is influenced by various factors. This study aimed to develop and validate a novel machine learning (ML)-based risk score to stratify HCC patients into prognostic groups and compare its predictive accuracy with conventional staging systems and biomarkers.

This retrospective study included 4038 HCC patients diagnosed between 2012 and 2023. Patients were randomly divided into training (n = 2827) and validation (n = 1211) cohorts in a 7:3 ratio. Prognostic factors for overall survival were identified using conventional Cox proportional hazards models and ML-based least absolute shrinkage and selection operator Cox regression. Variables significant in both methods were incorporated into a risk score, with each parameter scaled from 0 to 100, based on multivariable Cox regression coefficients. Patients were categorized into 3 risk groups using the 33rd and 66th percentiles.

After a median follow-up of 42 months (interquartile range: 33.6-50.4), 1931 deaths occurred, with a 5-year overall survival rate of 26.0%. Eleven significant prognostic variables were identified: age, maximum tumor size, extrahepatic metastasis, macrovascular invasion, serum albumin, bilirubin, creatinine, aspartate aminotransferase, lymphocyte-to-monocyte ratio, alpha-fetoprotein level, and treatment modality. The CATS-BIT-ME score demonstrated high predictive accuracy for overall survival (area under the receiver operating characteristic curve: 0.800) and effective risk stratification within the 5 years (area under the receiver operating characteristic curve range: 0.852-0.884). It outperformed 13 conventional staging systems and biomarkers in prognostic precision.

The ML-based CATS-BIT-ME score provides an accurate and reliable tool for stratifying HCC patients into distinct prognostic groups, enhancing personalized disease management and clinical decision-making.