The prognostic role of an optimal machine learning model based on clinical available indicators in HCC patients

Although methods in diagnosis and therapy of hepatocellular carcinoma (HCC) have made significant progress in the past decades, the overall survival (OS) of liver cancer is still disappointing. Machine learning models have several advantages over traditional cox models in prognostic prediction. This study aimed at designing an optimal panel and constructing an optimal machine learning model in predicting prognosis for HCC. A total of 941 HCC patients with completed survival data and preoperative clinical chemistry and immunology indicators from two medical centers were included. The OCC panel was designed by univariate and multivariate cox regression analysis. Subsequently, cox model and machine-learning models were established and assessed for predicting OS and PFS in discovery cohort and internal validation cohort. The best OCC model was validated in the external validation cohort and analyzed in different subgroups. In discovery, internal and external validation cohort, C-indexes of our optimal OCC model were 0.871 (95% CI, 0.863-0.878), 0.692 (95% CI, 0.667-0.717) and 0.648 (95% CI, 0.630-0.667), respectively; the 2-year AUCs of OCC model were 0.939 (95% CI, 0.920-0.959), 0.738 (95% CI, 0.667-0.809) and 0.725 (95% CI, 0.643-0.808), respectively. For subgroup analysis of HCC patients with HBV, aged less than 65, cirrhosis or resection as first therapy, C-indexes of our optimal OCC model were 0.772 (95% CI, 0.752-0.792), 0.769 (95% CI, 0.750-0.789), 0.855 (95% CI, 0.846-0.864) and 0.760 (95% CI, 0.741-0.778), respectively. In general, the optimal OCC model based on RSF algorithm shows prognostic guidance value in HCC patients undergoing individualized treatment.