The prediction of karst groundwater level is prone to significant errors due to the strong nonlinear and non-stationary fluctuation characteristics of karst groundwater. Addressing to the poor prediction accuracy problem, an EMD-LSTM coupled model is proposed. Firstly, empirical mode decomposition (EMD) is used to decompose the karst groundwater level of Baotu Spring into five components (four intrinsic mode function terms and one residual term), in order to eliminate non-stationary fluctuations in water level data. At the same time, a long short-term memory (LSTM) neural network model is constructed, and the indices that are closely related to the dynamic change of groundwater level, such as rainfall (representing the aquifer recharge term), monthly average temperature, monthly maximum temperature, monthly minimum air temperature, and water vapor pressure (representing the aquifer discharge term), are used as input items to predict the five components respectively, and finally the prediction results of the components are summed up to obtain the prediction value of the groundwater level. The results indicate that EMD can significantly eliminate the non-stationary fluctuation characteristics of karst groundwater level; the EMD-LSTM coupled model can effectively improve the accuracy of karst groundwater level prediction, with a root mean square error reduction of 28.22% and 59.94% compared to the LSTM model and ARIMA model, respectively. Overall, the proposed EMD-LSTM coupling model has strong reliability and stability, which can provide some reference for accurate prediction of karst groundwater level.