Digital educational games exhibit substantial promise in advancing STEM education. Nevertheless, the empirical evidence on both the efficacy of digital game-based learning and its designs in STEM education is characterized by notable inconsistencies. Therefore, the current study aimed to investigate (1) the general effect of digital game-based STEM learning over STEM learning without digital game, and (2) the enhancement effect of added game-design elements against base game versions in STEM learning. Two meta-analyses were conducted in this study. Based on the 136 effect sizes extracted from 86 studies, the first meta-analysis revealed a medium to large general effect of digital game-based STEM learning over conventional STEM learning (g = 0.624, 95% CI [0.457, 0.790]). In addition, digital game-based STEM learning appeared to be differentially effective for different learning outcome, different types of game, and different subject. A total of 44 primary studies and 81 effect sizes were identified in the second meta-analysis. The results revealed a small to medium enhancement effect of added game-design elements over base game versions (g = 0.301, 95% CI [0.163, 0.438]). Furthermore, our results indicated that the game-design elements added for content learning were more effective than those added for gaming experience. Possible explanations for these findings, as well as the limitations and directions for future research were discussed.