Machine learning based hepatic safety score predicts decompensation in hepatocellular carcinoma systemic therapy

Abstract

Hepatocellular carcinoma (HCC) frequently coexists with portal hypertension, significantly increasing the risk of hepatic decompensation (HD) and variceal bleeding during systemic therapy. We developed a machine learning based hepatic safety score (MHSS) using data from 2026 patients with unresectable HCC to predict clinically significant portal hypertension (CSPH) and prognosis. A random forest model was trained in a derivation cohort (n = 1262) and validated in an independent cohort (n = 764). The MHSS demonstrated robust performance (AUROC 0.840) in CSPH and predicting HD. Stratification revealed that high MHSS patients faced significantly elevated risks of HD (HR 3.25), variceal bleeding (VB, HR 4.90), and mortality (HR 2.21). Crucially, while atezolizumab-bevacizumab offered a survival advantage in low MHSS patients, it was associated with high bleeding risk and no survival benefit in the high MHSS group compared to other regimens. A simulation of MHSS guided treatment selection demonstrated a 24% reduction in HD, a 40% reduction in VB, and a 26% reduction in mortality. In conclusion, the MHSS effectively predicts CSPH, decompensation, and survival in patients with HCC prior to systemic therapy. By enabling individualized risk stratification, the MHSS may guide personalized treatment selection between bevacizumab-containing and alternative regimens, ultimately improving patient outcomes. Clinical trial number: not applicable.