• ML/DL methods did not significantly outperform Cox models for HCC risk prediction. • Cox models offered interpretable and efficient predictions with less overfitting. • Variable selection was consistent across top-performing models. Hepatocellular carcinoma (HCC) is a leading cause of cancer-related death worldwide, with chronic hepatitis B virus (HBV) infection being a major risk factor. To date, existing predictive scores of HCC are mainly based on traditional Cox proportional hazard (CPH) models. This study aimed to compare the variable selection process and performance of CPH models with those of machine learning (ML) and deep learning (DL) algorithms in predicting HCC among patients with chronic HBV infection. We used data from 4,370 individuals with chronic HBV infection enrolled in the French prospective multicentre ANRS CO22 HEPATHER cohort, of which 56 (1.3%) developed an HCC. Two published CPH-based scores (ADAPTT and SADAPTT) were compared to Random Survival Forest (RSF), Survival Support Vector Machine (SVM), Survival XGBoost, and DeepSurv algorithms. Models were evaluated using Harrell’s C-index, Inverse-Probability-of-Censoring Weighting win ratio statistic, and time-dependent area under the ROC curve at 3, 5, and 8 years. The same set of covariables was used to build all the models. CPH models demonstrated similar or higher performances (C-index 95% confidence interval: 0.84 0.82–0.85) for HCC prediction compared to ML and DL models, with less overfitting. Survival SVM and RSF performed similarly (0.81 0.79–0.83 and 0.81 0.79–0.82, respectively) without outperforming CPH models. Variable selection was consistent across top-performing models, though CPH models more effectively captured the predictive value of certain behavioural factors, such as soft drink intake. In this dataset with a limited sample size and strongly imbalanced outcome, traditional CPH models provided robust, interpretable, and computationally efficient predictions for HCC risk. ML and DL methods did not outperform traditional models, reinforcing the validity of traditional statistical approaches in small to medium datasets.
Ramier et al. (Thu,) studied this question.