Abstract : Objective To explore the value of the machine learning model based on routine laboratory data for the risk prediction of primary liver cancer. Methods A total of 298 cases of liver cancer group and 882 cases of non-hepatocellular carcinoma group were collected from the hospital A. The characteristic parameters for model were screened out to build a prediction model based on machine learning. A total of 178 cases of the liver cancer group and 315 cases of the non-hepatocellular carcinoma group were collected from the hospital B. The external validation for the constructed optimal model was conducted. Results The two optimal combinations of characteristic parameters were screened out by statistical methods. Model 1-5 and Model 6-10 were built by machine learning algorithm respec- tively. In the internal verification set, Model3 ( AUCROC = 0.952, accuracy=0.899) and Model8 ( AUCROL = 0.951, accuracy =0.897)constructed by XGBoost al gorithm exhibited the optimal perfomance indicators. The characteristice parameters shared by Model 3 and Model 8 included sex, age, alpha-fet oprotein, C-reactive protein and cystatin C. The characteristic parameters of Model 3 also included fibrinogen with AUCHOC = 0.823 and accuracy= 0.793 for the external validation set, and the characteristic parameters of Model 8 also included albumin with AUCRO a = 0.816 and accuracy= 0.793 for the external validation set ，respectively. Conclusion A risk prediction model for primary liver cancer could be constructed based on routine laboratory data.