Abstract: The accurate estimation of the thickness of the excavation damaged zone (EDZ) is a fundamental and indispensable parameter for evaluating the stability of roadway surrounding rock and for ensuring the safety of underground engineering structures. In this study, five key factors that sufficiently reflect the characteristics of the EDZ were carefully selected as predictive indicators. These factors include the burial depth of the roadway, the span of the roadway, the cross-sectional area of the excavation, the uniaxial compressive strength of the rock mass, and the degree of joint development. Based on the dung beetle optimization (DBO) algorithm, a multi-strategy enhanced version, referred to as the improved dung beetle optimization algorithm (IDBO), was developed by integrating Chebyshev chaotic mapping, the golden sine strategy, and a dynamic weighting coefficient. The integration of these methods and strategies provides a reinforced optimization mechanism within the DBO framework, which improves both the global search capability and the overall convergence behavior of the algorithm. Subsequent use of the IDBO algorithm to optimize the initial weights and thresholds of a backpropagation (BP) neural network resulted in the construction of the IDBO–BP neural network model, which is specifically designed for predicting the thickness of the EDZ. A total of 102 sets of EDZ engineering case data were used to verify the applicability, reliability, and predictive accuracy of the proposed model. Spearman correlation analysis—applied to examine the degree of association between the selected predictive indicators and EDZ thickness—allows for a quantitative evaluation of the relative influence of each factor. Several commonly used evaluation metrics, including the coefficient of determination (R²), mean absolute error (MAE), mean bias error (MBE), and root mean square error (RMSE), were employed to systematically assess the predictive performance of the IDBO–BP neural network model. Together, these evaluation metrics provide a comprehensive assessment of the prediction accuracy, generalization ability, and overall performance of the model. Application of the model to predict the EDZ thickness of roadway sections with different burial depths in a gold mine provided additional verification of its practical engineering applicability and predictive effectiveness. The results indicate that among the five predictive indicators, the degree of joint development exhibits the strongest correlation with the EDZ thickness, whereas the cross-sectional area of the excavation demonstrates the weakest correlation. Compared with the BP and DBO–BP models, the IDBO–BP model demonstrates markedly smaller deviations for both the training and testing datasets by reflecting improved predictive accuracy and more reliable behavior. When compared with other widely used models, including support vector machine (SVM), random forest (RF), partial least squares (PLS), BP, and DBO–BP, the R² value of the IDBO–BP model is closest to 1, while this model simultaneously yields the lowest MAE, MBE, and RMSE values. Moreover, when applied to three horizontal roadways, the IDBO–BP model has an average prediction error of 8.2%, which represents reductions of 9.3% and 5.3% relative to the BP and DBO–BP models, respectively. These findings collectively indicate that the proposed IDBO-BP model has high predictive performance, iteration speed, and reliability, which further confirm its practical applicability and engineering value.