The stability of underground entry-type excavations (UETEs) is of paramount importance for ensuring the safety of mining operations. As more engineering cases are accumulated, machine learning (ML) has demonstrated great potential for the stability evaluation of UETEs. In this study, a hybrid stacking ensemble method aggregating support vector machine (SVM), k-nearest neighbor (KNN), decision tree (DT), random forest (RF), multilayer perceptron neural network (MLPNN) and extreme gradient boosting (XGBoost) algorithms was proposed to assess the stability of UETEs. Firstly, a total of 399 historical cases with two indicators were collected from seven mines. Subsequently, to pursue better evaluation performance, the hyperparameters of base learners (SVM, KNN, DT, RF, MLPNN and XGBoost) and meta learner (MLPNN) were tuned by combining a five-fold cross validation (CV) and simulated annealing (SA) approach. Based on the optimal hyperparameters configuration, the stacking ensemble models were constructed using the training set (75% of the data). Finally, the performance of the proposed approach was evaluated by two global metrics (accuracy and Cohen’s Kappa) and three within-class metrics (macro average of the precision, recall and F1-score) on the test set (25% of the data). In addition, the evaluation results were compared with six base learners optimized by SA. The hybrid stacking ensemble algorithm achieved better comprehensive performance with the accuracy, Kappa coefficient, macro average of the precision, recall and F1-score were 0.92, 0.851, 0.885, 0.88 and 0.883, respectively. The rock mass rating (RMR) had the most important influence on evaluation results. Moreover, the critical span graph (CSG) was updated based on the proposed model, representing a significant improvement compared with the previous studies. This study can provide valuable guidance for stability analysis and risk management of UETEs. However, it is necessary to consider more indicators and collect more extensive and balanced dataset to validate the model in future.