Comparison of machine learning and Cox regression models for prognostic analysis in hepatocellular carcinoma patients with distant metastasis

Abstract

Background

With the development of conversion therapy, there has been a significant improvement in advanced stage hepatocellular carcinoma (HCC) patients' survival outcomes. Accurate prognostic assessment of patients with distant metastasis (DM) is therefore pivotal in improving quality of life, guiding treatment, and optimizing patient management.

Methods

This study extracted patients with distant metastatic HCC from the Surveillance, Epidemiology, and End Results database. Univariate and multivariate Cox regression were used to identify prognostic factors. Then, Cox regression, DeepSurv, Decision Tree, and Random Survival Forests models were used to predict overall survival. Model performance was evaluated by area under the curve (AUC), decision curve analysis, calibration curve, and Brier score. The visualization of Cox regression and machine learning algorithms utilized nomogram and Shapley additive explanations, respectively.

Results

The study included 3051 HCC patients with DM. Factors such as tumor size, lung metastasis, N stage, ace, chemotherapy, radiotherapy, AFP, fibrosis, treatment interval, and number of metastases were independently associated with patient prognosis. Among all models, Cox regression and Random Survival Forest models showed stable performance, achieving AUCs of 0.746/0.760, 0.745/0.749, and 0.729/0.718 at 3, 6, and 12 months, respectively. Meanwhile, Cox regression showed the lowest Brier score (0.180 and 0.125) at 6 and 12 months.

Conclusions

Cox regression and Random Survival Forest models demonstrated robust prognostic performance for HCC, with Cox regression exhibiting superior temporal stability. The Cox-based nomogram provides an intuitive tool for rapid 3-, 6-, and 12-month survival stratification in metastatic HCC patients.