Machine-learning-based decision-tree model for patients with single-large hepatocellular carcinoma

Background: Single-large hepatocellular carcinoma (SLHCC) is defined as a solitary tumor that is larger than 5 cm and lacks macrovascular invasion or extrahepatic spread. SLHCC is a distinct clinical subtype with considerable prognostic heterogeneity, and available staging systems offer limited predictive accuracy for this subgroup. Therefore, we aimed to develop a machine learning (ML)-based decision-tree model to improve individualized prognostic stratification of SLHCC. Methods: This retrospective study included patients with SLHCC who were diagnosed at Taipei Veterans General Hospital between January 2012 and January 2023. The patients were randomly assigned to a training cohort and a validation cohort. Prognostic factors for overall survival (OS) were identified using multivariate Cox regression and incorporated into a decision-tree algorithm. The model performance was evaluated using accuracy and the area under the receiver operating characteristic curve (AUROC). Results: Among the 477 patients, 307 (64.4%) received curative treatment, and 170 (35.6%) received non-curative therapy. The median age was 70 years, and 77.1% were male. After a median follow-up of 50 months, the 5-year OS rate was 42.0%. Six variables were independently associated with OS: tumor size > 10 cm, serum creatinine > 1 mg/dL, non-curative treatment, albumin bilirubin (ALBI) grade 2, fibrosis-4 (FIB-4) score ≥ 2.67, and serum alpha-fetoprotein (AFP) > 20 ng/mL. The decision-tree model incorporated four key variables: including treatment modality, creatinine, tumor size, and FIB-4. The model stratified patients into five risk groups. The model’s accuracy was 74.3% in the training cohort and 67.1% in the validation cohort, and the AUROCs were 0.756 and 0.706, respectively. Conclusion: The clinically interpretable ML-based decision-tree model effectively stratifies patients with SLHCC according to prognosis using routine clinical and laboratory data. This model complements conventional staging systems and could support personalized treatment planning and patient counseling in real-world clinical practice.