Abstract: The gradual increase in coal-mine excavation depth leads to the significant rise in the in situ stress in deep surrounding rock and escalating risks of gas desorption and accumulation, causing a higher likelihood of coal-gas outbursts. In this light, the present study develops a deep-learning-based predictive model for coal-gas outbursts. First, the collected data were preprocessed using the Local Outlier Factor (LOF) and Multiple Imputation by Chained Equations (MICE), and employed Kendall's rank correlation coefficient to select those factors exhibiting strong correlation as the predictive indicators for gas outbursts. Next, a convolutional neural network (CNN) architecture was constructed, and optimized its hyperparameters via an enhanced dung beetle optimization algorithm (MSADBO). This algorithm incorporates an improved sine-based dynamic search-step adjustment, an adaptive Gaussian-Cauchy hybrid mutation to bolster global and local search capabilities, and a Bernoulli chaotic-map strategy to increase population diversity. Finally, comparative models were established; accuracy and other evaluation metrics were compared across models, and the safety of the predictions was analyzed via confusion matrices. Results demonstrate that the MSADBO-CNN model achieved an accuracy of 98.7 % on the training set and 91.67 % on both the validation and test sets, thereby attaining the highest predictive precision while also exhibiting superior robustness, generalization ability, and operational safety.