Due to the particularity of loess engineering geology, loess railway tunnel accidents occur frequently. Based on the theoretical basis of risk management, this study evaluated the service performance of loess railway tunnels. Based on the improved TOPSIS method, the indirect proximity degree of each risk factor was compared, and the appropriate service performance evaluation index was selected. Based on the ISM model and the causality graph modification method, the dependency relationship between nodes was obtained and the Bayesian network evaluation model was constructed. By constructing the database, the EM algorithm was used for data learning, and the model was trained and verified, and the overall accuracy (ACC) and F₁ value are used to comprehensively evaluate the training and prediction effect of the model. Finally, the evaluation model was applied to a tunnel case. The results show that the established Bayesian network model has a high accuracy of 92%, which is easy to operate, effective and practical, and it is also applicable in situations of incomplete index statistical data.