South Korean researchers have developed an artificial intelligence (AI) model that predicts the risk of rapid liver function deterioration before patients with hepatocellular carcinoma begin systemic treatment. The tool could help physicians evaluate not only which therapy is most effective, but also which treatment is safest for each individual patient.
Researchers at The Catholic University of Korea Seoul St. Mary’s Hospital, led by Professor Han Ji-won of the Division of Gastroenterology, developed the Machine Learning-Based Hepatic Safety Score (MHSS) after analyzing data from 2,026 patients with hepatocellular carcinoma treated at eight hospitals within the Catholic Medical Center network between 2010 and 2024.
Hepatocellular carcinoma accounts for approximately 90% of all primary liver cancers. The MHSS integrates multiple clinical factors—including blood test results, liver function indicators, platelet counts, tumor size and number, vascular invasion, and tumor biomarkers—to help physicians select the safest and most effective treatment strategy for each patient.
Traditionally, clinicians have relied on liver function assessment tools such as the Child-Pugh score, Albumin-Bilirubin (ALBI) score, Model for End-Stage Liver Disease (MELD) score, and FIB-4 index. These scoring systems primarily evaluate liver function using laboratory data but do not account for tumor-specific characteristics such as tumor burden or vascular invasion.
The newly developed AI model incorporates both liver function and tumor-related information, allowing it to predict treatment-related liver deterioration and variceal bleeding more accurately than conventional assessment tools. The model also demonstrated consistent performance in an independent external validation cohort.
Patients classified as high risk by the MHSS had 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 compared with low-risk patients.
The findings suggest that treatment-related liver deterioration is influenced not only by baseline liver function but also by tumor characteristics such as size and vascular invasion, highlighting the value of AI in capturing these complex interactions.
The research team also evaluated treatment selection through simulation analyses. Among low-risk patients, the combination immunotherapy of atezolizumab and bevacizumab produced superior survival outcomes compared with alternative treatments. However, among high-risk patients, the increased risk of variceal bleeding reduced the regimen’s overall survival benefit.
Based on these findings, the researchers simulated a personalized treatment strategy that prioritized the atezolizumab-bevacizumab combination for low-risk patients while recommending therapies with lower bleeding risks for high-risk patients. Compared with standard treatment selection, the personalized approach was projected to reduce the risk of liver function deterioration by 24%, variceal bleeding by 40%, and overall mortality by 26%.
The researchers said the study provides clinical evidence that treatment decisions for liver cancer should consider not only which therapy offers the greatest efficacy but also which option is safest for each patient. The model could also support more targeted endoscopic screening before treatment, improve bleeding prevention strategies, and help physicians tailor treatment intensity and follow-up plans.
Han said the AI model provides an objective framework for integrating tumor characteristics, liver function, and portal hypertension risk into a single assessment, enabling safer and more individualized treatment decisions.
She added that future prospective studies and broader real-world validation will help further develop the model into a precision medicine tool that can support safer treatment in routine clinical practice.
The study was supported by the Global Physician-Scientist Training Program, funded by South Korea’s Ministry of Health and Welfare and the Korea Health Industry Development Institute.
The findings were published in the international journal npj Digital Medicine (Impact Factor: 18.0), and the research team has made the prediction model freely available through a web-based calculator to improve accessibility for both clinicians and patients.
