Prognostic Modeling of 1-Year Survival in Hepatocellular Carcinoma: A Comparative Evaluation of Machine Learning
Jinlian Wang1, Hui Li2, Changqing Ju2
1Department of Quantitative and Systems Health Sciences, University of Texas at Austin Dell Medical School, Austin, TX 78712, USA.
Objective:
To develop and comparatively evaluate machine learning models for predicting 1-year survival in hepatocellular carcinoma (HCC) patients, focusing on the prognostic interplay between chronic liver diseases and diabetes aligned with the EASL Clinical Practice Guidelines (2018), and to assess the impact of class balancing via Synthetic Minority Oversampling Technique (SMOTE) on model performance.
Methods:
Retrospective analysis was performed on the UCI HCC Survival Dataset (N = 165). Following median/mode imputation of ~10% missing values, four classifiers were evaluated: Logistic Regression (LR), Multilayer Perceptron (MLP), Gradient Boosting (GB), and Random Forest (RF), each trained with and without SMOTE oversampling (eight variants total). Models were trained on 16 EASL-aligned clinical and biochemical features using a stratified 80/20 train/test split (training N = 132, held-out test N = 33). Performance was evaluated via accuracy, AUC-ROC (with 95% bootstrap confidence intervals across 1000 iterations), precision, recall, and macro F1-score. Stratified 5-fold cross-validation and multiple imputation sensitivity analyses (MissForest, KNN, MICE, complete-case) were conducted. SHAP (SHapley Additive exPlanations) was applied to the top-performing model to provide patient-level interpretability.
Results:
Diabetes was present in 33.9% of patients (56/165; 28/102 survivors [27.5%] vs. 28/63 non-survivors [44.4%]). Diabetes was significantly associated with 1-year mortality in univariate analysis (χ2 = 5.02, p = 0.025) but not after multivariable adjustment (aOR = 0.50, 95% CI: 0.21-1.22, p = 0.129). GB + SMOTE achieved the highest discriminative ability (AUC-ROC: 0.727, 95% CI: 0.552-0.886), while LR + SMOTE yielded the highest overall accuracy (72.7%) and macro F1-score (0.723). Without SMOTE, all models exhibited poor minority-class detection (non-survival F1 range: 0.300-0.615). SMOTE markedly improved non-survival recall across LR (0.231 to 0.769) and GB (61.5% to 69.2%). Stratified 5-fold CV AUC on the training set was highest for RF (0.779 ± 0.100), followed by GB (0.744 ± 0.065). TreeSHAP analysis identified ECOG performance status (mean |SHAP| = 1.284) and ALT (1.060) as the most influential predictors; diabetes ranked 15th (mean |SHAP| = 0.014), consistent with minimal independent predictive contribution.
Conclusions:
Gradient boosting and regularized linear regression combined with SMOTE class balancing provide competitive, well-calibrated, and transparent risk stratification on small, imbalanced tabular HCC datasets, offering an exploratory proof-of-concept for personalized prognostic modeling that requires external validation.
