Development and Validation of a Machine Learning Model to Prognosticate Hepatocellular Carcinoma

BACKGROUND: Prognostic models for hepatocellular carcinoma (HCC) may have limited accuracy. We aimed to construct and validate a novel prognostic model for HCC that incorporates biomarkers for liver function and tumour characteristics. METHODS: Consecutive participants (n = 1102) with HCC from five international tertiary institutions in Asia and the U.S. comprised the derivation (n = 627), internal validation (n = 270) and external validation (n = 205) cohorts. The Liver Cancer Risk predictioN (LCRN) Index was constructed using a gradient-boosted decision tree model based on the Cox proportional hazards framework. The discriminative performance of the LCRN Index was evaluated using Harrell's concordance index (C-index) and compared to the albumin bilirubin grade (ALBI), Barcelona Clinic for Liver Cancer (BCLC) staging and Cox regression model. Model calibration was assessed using the integrated Brier score and calibration plots. RESULTS: and 77.2% were male. The LCRN Index comprised type 2 diabetes mellitus, ascites, hepatic encephalopathy, albumin, bilirubin, AFP and diameter of largest tumour nodule. In the external validation cohort, the LCRN Index (1-year area under the receiver operating curve [AUC]: 0.79; 3-year AUC: 0.77; 5-year AUC: 0.75) was numerically higher than the ALBI grade and BCLC stage for prognosticating HCC. The LCRN Index demonstrated improved calibration compared to the ALBI grade and BCLC stage. CONCLUSION: The LCRN index is a promising tool for prognosticating HCC. If further validated, these data may have potential clinical implications for the management of HCC.