Research Of Medication Combination Recommendation And Optimization Methods Based On Time-series EHR Data

In recent years,with the widespread usage of Electronic Health Records(EHR)and the rapid development of computer technologies,deep learning techniques have made great achievements in prediction tasks in the medical field,among which medication combination recommendation based on patient’s historical visit records is a research hotspot in the field of medical informatics,and accurate and effective recommendation of medication combinations are of great significance to help physicians make appropriate clinical decisions.Therefore,this thesis focuses on how to improve the performance of recommended medication combinations by mining patients’ visit patterns based on time-series EHR data.Currently,most studies have addressed this problem by modeling patient visit sequences using deep learning models.However,existing methods still have some shortcomings:on the one hand,they mainly focus on obtaining global dependency information of medical data,ignoring local dependency information,and they also do not take into account the impact of irregular time intervals existing in time-series EHR data on the performance of medication recommendation;on the other hand,some drugs have Drug-Drug Interaction(DDI)relationships,while the use of medication combinations with adverse DDI relationships may lead to patient exacerbation or even death,so it is necessary to consider adverse DDI relationships in medication recommendation tasks.In addition,unstructured textual information such as drug descriptions and physician prescriptions have positive effects on improving the accuracy of medication recommendations.However,current approaches do not consider fusing these unstructured textual information to improve the modeling performance of deep learning models.To address the above issues,this thesis provides an in-depth study of a deep learning-based medication combination recommendation model.1.A Time-Aware Hierarchical Dependency Network for Medication Recommendation(TAHDNet)based on time-aware hierarchical dependency learning is proposed.In response to the irregular time interval feature in EHR data,the model introduces a dynamic time-aware module to capture the changes of patients’ conditions with time intervals.Besides,the model learns two different levels of dependencies,global dependency at medication and diagnosis level and local dependency at patient’s visit level,by using a Transformer-based and a onedimensional convolutional neural network,respectively,to enhance the representation learning capability,which improving its performance on medication recommendation task.2.A medication combination recommendation optimization model based on information enhancement and adverse DDI information is proposed.Based on Research Component 1,the model performs medication combination recommendation optimization by incorporating unstructured textual information and DDI relationships.Specifically,the model uses Transformer network to represent the unstructured text information for learning,and introduces the DDI loss function to jointly determine the results of medication combination recommendation based on TAHDNet model.In addition,this thesis designs a text-aware skip connection module to fuse unstructured text information features with structured EHR features.In addition to enriching the information representation,this module can further optimize the performance of the model by avoiding the gradient disappearance and gradient explosion problems due to the excessive complexity of the model as much as possible.Finally,the performance of the model is validated on the MIMIC-Ⅲ public dataset in this thesis.The experimental results show that the models proposed in this thesis all achieve better experimental results compared with the existing medication recommendation methods.