Seoul St. Mary's Hospital develops AI model to predict liver function decline before cancer treatment

Seoul St. Mary's Hospital announced Wednesday that it has developed an AI model capable of predicting the risk of sudden liver function deterioration in hepatocellular carcinoma patients before they begin systemic treatment.

A research team led by Professor Han Ji-won of the hospital's Division of Gastroenterology analyzed data from 2,026 hepatocellular carcinoma patients treated at eight hospitals under the Catholic Medical Center network between 2010 and 2024, building a model called the Machine Learning-based Hepatic Safety Score, or MHSS.

The model draws on a comprehensive set of variables — including blood test results, liver function indicators, platelet counts, tumor size and number, vascular invasion status and tumor markers — to help clinicians select safe and effective personalized treatment plans.

The model is designed to assess the risk of liver function deterioration and variceal bleeding before a patient undergoes systemic treatment for hepatocellular carcinoma. While existing evaluation tools have focused primarily on blood test values and liver function indicators, MHSS also incorporates tumor-specific characteristics such as size, number and vascular invasion.

As a result, the model predicted variceal bleeding and post-treatment liver function decline more accurately than existing tools. It also produced stable results in an independent validation cohort composed of patients from other institutions.

Patients classified as high-risk by the model faced significantly elevated hazards compared with low-risk patients: a 3.25-fold higher risk of liver function deterioration during treatment, a 4.90-fold higher risk of variceal bleeding, and a 2.21-fold higher risk of death.

The hospital said the research team's analysis of outcomes by treatment choice also yielded meaningful clinical insights.

Simulations showed that a specific immunotherapy combination — atezolizumab plus bevacizumab — offered superior survival benefits over other treatments in low-risk patients, but in high-risk patients the regimen was associated with an increased risk of variceal bleeding that offset any clear survival advantage.

Based on those findings, the team ran a personalized treatment simulation that prioritized the immunotherapy combination for low-risk patients while favoring treatments with a relatively lower bleeding risk for high-risk patients. The simulation projected a 24 percent reduction in the risk of liver function deterioration, a 40 percent reduction in variceal bleeding risk and a 26 percent reduction in overall mortality risk compared with a non-optimized approach.

Han, who led the study, said the research "established an objective foundation for presenting safe and rational treatment pathways for individual patients by comprehensively evaluating tumor characteristics, liver function, and portal hypertension risk within a single AI framework." She added that the team plans to "develop this into a personalized precision medicine tool — validated through prospective studies and diverse real-world data — so that patients can safely continue treatment in actual clinical settings."

The study was funded through the Global Physician-Scientist Training Program supported by the Ministry of Health and Welfare and the Korea Health Industry Development Institute, and published in npj Digital Medicine (impact factor: 18.0), an international peer-reviewed journal in digital healthcare and medical AI. The research team has also made the predictive model freely available as a web-based calculator to improve clinical accessibility for patients and healthcare providers.