Machine Learning Models Using Routine Clinical Data versus Conventional Clinical Scores for Survival Prediction in Hepatocellular Carcinoma: A Single-Center Study
Journal of Hepatocellular Carcinoma · 2026-11 · Volume 13 · 1-18
DOI: 10.2147/jhc.s651173 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Hepatocellular Carcinoma, 2026-11. Verfasst von Baver Ordu und 7 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 3.0).
Abstract (Originalsprache)
Background
Hepatocellular carcinoma (HCC) has a poor prognosis, and conventional clinical scores have limited accuracy in predicting individualized survival.
Purpose
We aimed to determine whether machine learning (ML) models using routine clinical data discriminate survival better than established clinical scores in HCC and to compare the prognostic profiles of MASLD- and HBV-associated HCC.
Methods
We conducted a single-center retrospective cohort study of 233 patients with HCC diagnosed between 2010 and 2026. Patients with valid survival data and ≤ 20% missingness across 29 pre-specified predictor variables were included. During a median follow-up of 15 months, 143 patients died. Random Survival Forest, XGBoost, a Hybrid ensemble model, and penalized Cox regression were trained using 5-fold cross-validation. Model discrimination was assessed using Harrell’s C-index and the fixed-horizon AUROC. The score-excluded V2 analysis removed clinical scores from ML inputs. SHAP values quantified feature contributions.
Results
The best-performing ML model (Hybrid) achieved a C-index of 0.797, higher than BCLC stage (0.735), Child-Pugh class, and ALBI grade, but only modestly higher than penalized Cox regression with the same variables (0.773). In the score-excluded sensitivity analysis, the Random Survival Forest maintained a strong performance (C-index 0.784). The best fixed-horizon 12-month AUROC was 0.878. SHAP analysis identified AFP, ECOG performance status, albumin, tumor diameter, and total bilirubin as the most influential predictors. Among 57 patients with MASLD-associated HCC, the median survival was 23.0 months vs 54.0 months in 116 patients with HBV-associated HCC (p=0.0328); this difference was not explained by age, comorbidities, tumor characteristics, or treatment.
Conclusions
In this single-center cohort, models using routinely available variables discriminated survival in HCC better than conventional clinical scores by combining more information. Keywords: Machine learning, HCC, prognosis, MASLD
Zitieren
APA
Ordu, B., Dertsiz, B., Ertugrul, V., Karatas, M. O., Duran, O. F., Yesilyurt, M. C., Adali, G., & Mahmutoglu, M. A. (2026). Machine Learning Models Using Routine Clinical Data versus Conventional Clinical Scores for Survival Prediction in Hepatocellular Carcinoma: A Single-Center Study. Journal of Hepatocellular Carcinoma, Volume 13, 1-18. https://doi.org/10.2147/jhc.s651173
Vancouver
Ordu B, Dertsiz B, Ertugrul V, Karatas MO, Duran OF, Yesilyurt MC, et al. Machine Learning Models Using Routine Clinical Data versus Conventional Clinical Scores for Survival Prediction in Hepatocellular Carcinoma: A Single-Center Study. Journal of Hepatocellular Carcinoma. 2026;Volume 13:1-18. doi:10.2147/jhc.s651173
BibTeX
@article{Ordu2026,
author = {Ordu, Baver and Dertsiz, Berkay and Ertugrul, Volkan and Karatas, Muhammet Osman and Duran, Osman Firat and Yesilyurt, Merve Cinar and Adali, Gupse and Mahmutoglu, Mustafa Ahmed},
title = {Machine Learning Models Using Routine Clinical Data versus Conventional Clinical Scores for Survival Prediction in Hepatocellular Carcinoma: A Single-Center Study},
journal = {Journal of Hepatocellular Carcinoma},
year = {2026},
volume = {Volume 13},
pages = {1--18},
doi = {10.2147/jhc.s651173}
}
EUVIMED