Abstract: The thickness of the excavation damaged zone is a fundamental parameter for assessing the stability of roadway surrounding rock, and its accurate prediction is of great significance to the safety of underground engineering. Therefore, five parameters including roadway buried depth, span, excavation cross-sectional area, uniaxial compressive strength of the rock mass, and joint development degree were selected as predictive indicators of the excavation damaged zone thickness. An improved dung beetle optimization (IDBO) algorithm was employed to optimize the back propagation (BP) neural network, and a predictive model for excavation damaged zone thickness based on the IDBO–BP neural network was established. Moreover, the applicability and reliability of the proposed model were validated using 102 sets of excavation damaged zone instance samples. Meanwhile, the performance of the proposed model was evaluated using four statistical metrics: the coefficient of determination (R2), mean absolute error (MAE), mean bias error (MBE), and root mean square error (RMSE). Finally, the model was applied to predict the excavation damaged zone thickness of different mining level roadways in a certain gold mine. The results indicate that among the five predictive indicators, the joint development degree exhibits the strongest correlation with the excavation damaged zone thickness, whereas the excavation cross-sectional area shows the weakest correlation. Compared with the conventional BP and DBO–BP models, the IDBO–BP prediction model exhibits smaller deviations between the predicted and actual values in both the training and testing datasets, indicating higher prediction accuracy. Furthermore, Compared with SVM, RF, PLS, BP, and DBO-BP, the R2 value of IDBO-BP is closest to 1, and the MAE, MBE, and RMSE values are the smallest, indicating that IDBO-BP has the best prediction performance, fastest iteration speed, and highest prediction accuracy. The average prediction error of the IDBO–BP model for the excavation damaged zone thickness across the three mining levels is 8.2%, representing reductions of 9.3% and 5.3% compared with the BP and DBO–BP models, respectively, further demonstrating the reliability and effectiveness of the proposed approach.