Clinical Rational Drug Use Recommendation System Based On Deep Learning

Objective: Currently,the main recommendation system for rational drug use in medical practice is the rational drug use monitoring system.Clinical drug recommendation to assist clinical drug therapy is widely used in medical practice.However,the traditional recommendation systems based on knowledge bases such as drug instructions are essentially traditional expert systems,which are difficult to adapt to the individualized medication needs of patients.In this paper,we propose a model of drug recommendation system based on large-scale clinical medical record data and deep learning architecture of graph convolutional networks.Method:In this study,a patient medication recommendation system based on electronic medical record data and graph convolutional neural networks was developed.The system was trained and constructed using pediatric patient medical record data.Firstly,the pediatric electronic medical records written in natural language were processed for named entity recognition,and the best result of pre-training model was selected after testing and experiments to extract patient features from the medical records and construct a structured pediatric patient database.And on this basis,drug-drug association network,drugsymptom association network and symptom-symptom association network are constructed.Based on the graph structure-based patient feature network,the multimodal graph convolutional neural network architecture is used to learn the features of this graph structure,obtain the association network of individual patient features and drugs,and make medication recommendations accordingly.This model breaks the limitation that traditional graph convolutional neural networks can only process a single graph structure,and can process and learn multiple different types of graph structures simultaneously.The system can provide clinical medication recommendations based on the individual patient’s symptoms.To validate the effectiveness of the system,two approaches are taken: prediction by pre-clustering and prediction without pre-clustering based on disease diagnosis.And the results of traditional link prediction and graph convolutional neural network algorithms were compared with the model results in this study to verify the effectiveness of the model.Results:After experimental comparison,RoBERTa-wwm-ext was selected as the pretrained model for named entity recognition,and the named entity recognition module was constructed by fine-tuning the model and combining it with Bi GRU-CRF to extract medically relevant named entities from electronic medical records(F1-score=99.27%).Based on the efficient and accurate extraction of Chinese medical record features using named entity recognition technology,three multimodal graph is constructed and a multimodal graph convolutional neural network architecture is used for deep learning to obtain the Deep Graph Medicine Recommendation System(DGMRS).Recommendation System(DGMRS)".We use 4541 real clinical electronic medical records as all data to test the effectiveness of the model for medicine recommendation.The average recall rate of DGMRS without grouping disease diagnoses was 49.15% and the average accuracy rate was 31.14%,while the average recall rate of DGMRS with grouping disease diagnoses was 58.69% and the average accuracy rate was 37.15%,which was improvement.In comparison with the traditional link prediction and graph convolutional neural network algorithms,the traditional link prediction algorithm had an average recall of 1% and an accuracy of 0.62%,while the traditional graph neural network algorithm had an average recall of 6.55% and an accuracy of 4.09%.DGMRS was much more effective than the traditional models and methods.Conclusion:The named entity model of electronic medical records selected in this study possesses high accuracy,recall,and F1-score,and can adequately extract medically relevant named entities of patients.The drug recommendation system obtained by training based on multimodal graph convolutional neural network and electronic medical record data possesses much higher recommendation recall and accuracy than traditional models and methods.This result also validates that the system can learn from clinicians’ experience and provide them with treatment recommendations based on clinical experience.