In deep learning, attention mechanisms enhance the efficiency and performance of handling complex data structures by focusing on key information in the input data. In brain-computer interface (BCI) decoding, attention mechanisms can accurately identify important features in electroencephalogram (EEG) signals, thereby improving decoding accuracy and efficiency. Currently, there is a lack of systematic reviews in this field. This paper reviews the applications of attention mechanisms in BCI decoding and their impact on model performance. The categories discussed include self-attention mechanisms, channel attention mechanisms, spatiotemporal attention mechanisms, and task-optimized attention mechanisms. Additionally, this paper analyzes the performance of various decoding models incorporating attention mechanisms across four different paradigms, aiming to reveal their potential application value in BCI decoding research.