Abstract：In recent years,methods for improving the quality of medical care by mining large amounts of data and knowledge in electronic medical record systems have received widespread attention.The treatment engine can help doctors formulate a suitable treatment plan.The treatment engine predicts the patient′s next treatment prescription by modeling the patient′s historical hospitalization information.However,the timeliness and multi-modal characteristics of the data pose challenges to the effectiveness of the prediction results of the treatment engine.In order to meet the above challenges,an attention-based bidirectional heterogeneous LSTM treatment engine is proposed.This method proposes an end-to-end neural network that fully retains global time and multi-modal data information through a bidirectional heterogeneous LSTM structure.On the basis of the bidirectional heterogeneous LSTM,a bidirectional attention mechanism is established to make the treatment engine pay attention to the global time information,and predict the prescription of the next treatment course through the fully connected layer.In particular,the model pays attention to the time and multimodality of the data,and has high robustness and high predictive performance.It was tested on a large real critical medical data set MIMIC-III to verify the effectiveness of the treatment engine.Experimental results show that this method is superior to the most advanced method.
2021, Vol. 42 
