Objective A random survival forest algorithm was applied to explore the prognostic factors and develop the prognosis model for patients with unresectable hepatocellular carcinoma (HCC) after transcatheter arterial chemoembolization (TACE).

Methods Retrospective selection of 636 HCC patients treated with TACE as first-line treatment in the Department of Hepatology, Zhongshan Hospital, Fudan University from January 2014 to December 2017. The patients were divided into a training set (n=445) and a validation set (n=191) in a 7∶3 ratio. Based on the clinical data, laboratory indicators and follow-up survival of patients, the Cox proportional-hazards regression model and the random survival forest model based on machine learning algorithm was developed, and the predictive ability of the two models was evaluated.

Results The tumor burden, age, baseline gamma-glutamyl transpeptidase (GGT) level, baseline alpha-fetoprotein (AFP) level and albumin-bilirubin grade (ALBI) were independent factors affecting the prognosis of unresectable HCC patients treated with TACE. In the Cox model, the 1-year, 3-year and 5-year AUC of the training set was 0.782, 0.796 and 0.791, respectively, and the validation set was 0.750, 0.766 and 0.766, respectively. The 1-year, 3-year and 5-year AUC of the training set in the random survival forest model was 0.896, 0.894 and 0.875, respectively, and validation set was 0.743, 0.763 and 0.770, respectively. Random survival forest model could distinguish patients into good prognosis group and poor prognosis group, and the overall survival of these two groups was significantly different (P < 0.05). The decision curve analysis showed that the net benefit of the random survival forest model was better than that of the Cox proportional-hazards model.

Conclusions The random survival forest model is a reliable tool for predicting the prognosis of unresectable HCC patients treated with TACE.