<p>In this study, laboratory rockburst simulation experiments were conducted to collect complete-process acoustic emission (AE) signals. After data preprocessing, an improved catch fish optimization algorithm (ICFOA) was developed and integrated with an extreme gradient boosting (XGBoost) model to construct a rockburst stage prediction framework. Furthermore, SHAP analysis was employed to investigate the variation characteristics of AE signals during rockburst evolution. The results show that the AE feature dataset constructed in this study incorporates multiple dimensions, including time-domain, frequency-domain, and time–frequency features. Compared with single-dimensional approaches, it captures a broader range of information highly correlated with rockburst stage evolution. The ICFOA algorithm demonstrates superior performance in terms of convergence accuracy and global search capability. The ICFOA-XGBoost model constructed on this basis achieves higher accuracy than conventional optimization models in identifying four typical rockburst stages, with an overall recognition accuracy of approximately 86%. Among them, AE indicators such as the average signal level (ASL) are crucial predictive parameters, exhibiting distinct variation patterns across different rockburst stages. In conclusion, this study developed an ICFOA-XGBoost prediction model based on multidimensional AE feature parameters, achieving accurate identification of different rockburst evolution stages and providing an in-depth analysis of AE-signal variation characteristics. The findings offer key technical support for dynamic monitoring, trend assessment, and intelligent early warning of rockburst disasters, demonstrating strong engineering applicability and promising prospects for practical implementation. ICFOA-XGBoost code can be found here: <a href="https://github.com/qiqiqi-77/ICFOA-XGboost">https://github.com/qiqiqi-77/ICFOA-XGboost</a></p>