Background and Aims: Gastric cancer, characterized by high degree of malignancy and early metastasis, often leads to poor clinical prognosis, and particularly the gastric cancer liver metastasis (GCLM) is the main cause of death of the patients. However, there are still certain deficiencies in the evaluation methods for the prognosis of patients with GCLM at present. Therefore, this study was conducted to establish a prognostic evaluation model with good predictive ability by analyzing the clinicopathologic characteristics and prognostic risk factors of patients with GCLM based on the SEER database, so as to improve the individualized prognostic evaluation ability for the patients. 
Methods: The clinical data of GCLM patients diagnosed from 2010 to 2015 were extracted from SEER database. According to the inclusion and exclusion criteria, a total of 2 554 patients were included in the study after strict screening, and then the patients were randomly assigned to modeling set (1 790 cases) and validation set (764 cases) with a 7:3 ratio. The clinical baseline characteristics of patients in modeling set and the validation set were compared, and the independent risk factors for the overall survival (OS) and the cancer-specific survival (CSS) of GCLM patients were screened by Cox equal proportional regression model and Fine-Gray competitive risk model, respectively. Based on the results of multiple regression analysis of the modeling set Cox and Fine-Gray risk model and AIC optimization, the nomogram models for predicting the OS and CSS of GCLM patients were constructed. Finally, the reliability of the predictions obtained from the models were evaluated by C-index, ROC curve and calibration curve.
Results:  There were no significant difference in the baseline characteristic between the patients in modeling set and validation set. The results of analysis showed that age, chemotherapy, tumor grade, primary lesion resection and number of primary lesions were independent risk factors for OS, and chemotherapy, tumor grade, primary lesion resection and number of primary lesions were independent risk factors for CSS in GCLM patients. Based on the above variables, the nomogram models were constructed and evaluated, respectively. The C-index of either the nomogram model for predicting OS or for predicting CSS was remarkably higher than that of AJCC-TNM staging system (modeling set: 0.706 vs. 0.560 and 0.670 vs. 0.554; validation set: 0.769 vs. 0.534 and 0.744 vs. 0.518). Moreover, the ROC curve analysis showed that both prediction models had a relatively high accuracy. Finally, the calibration curve analysis showed that both nomogram models predicted OS or CSS had a good consistency with the actual observed values.
Conclusion: The constructed nomogram models based on SEER database have relatively high accuracy in predicting the OS and CSS in GCLM patients, which may help the clinicians to develop individualized treatment strategies for GCLM patients.