We investigated whether combining the pre-arrest serum albumin level could improve the performance of the Good Outcome Following Attempted Resuscitation (GO-FAR) score for predicting neurologic outcomes in in-hospital cardiac arrest patients. Adult patients who were admitted to a tertiary care hospital between 2013 and 2017 were assessed. Their pre-arrest serum albumin levels were measured within 24 h before the cardiac arrest. According to albumin levels, the patients were divided into quartiles and were assigned 1, 0, 0, and, - 2 points. Patients were allocated to the derivation (n = 419) and validation (n = 444) cohorts. The proportion of favorable outcome increased in a stepwise manner across increasing quartiles (p for trend < 0.018). Area under receiver operating characteristic curve (AUROC) of the albumin-added model was significantly higher than that of the original GO-FAR model (0.848 vs. 0.839; p = 0.033). The results were consistent in the validation cohort (AUROC 0.799 vs. 0.791; p = 0.034). Net reclassification indices of the albumin-added model were 0.059 (95% confidence interval CI - 0.037 to 0.094) and 0.072 (95% CI 0.013-0.132) in the derivation and validation cohorts, respectively. An improvement in predictive performance was found by adding the ordinal scale of pre-arrest albumin levels to the original GO-FAR score.