A Deep Learning-based Method And Application For Reviewing Chinese Medicine Prescriptions

As the core of TCM treatment,TCM prescription is one of the crucial research objects in the tasks related to TCM machine learning research.Using deep learning as a research tool to study TCM prescriptions,we can provide new ideas for TCM modernization research by capturing the combination pattern of prescribed herbs through neural network models,and deeply analyzing and discovering their characteristics.The purpose of Chinese medicine prescription review is to detect potentially inappropriate and wrong herbs in prescriptions in a timely manner,to improve the quality of treatment in medical institutions,and to protect the life and health of patients.The study firstly obtained the distributed representation of prescription herbs by predicting the central herbs,and obtained the herb feature vector with the information of herb co-occurrence relationship.On this basis,by calculating the difference between the prediction result and the real herbs,we proposed a herb evaluation method to score the herbs quantitatively,so as to find the potentially unsuitable herbs in the prescription and realize the automated prescription review method.Combined with an electronic Chinese medicine diagnosis and treatment system,it can solve various problems in the current prescription review process to a certain extent,such as insufficient review ability due to shortage of talents and inadequate professionalism,the review process floating in a formal way,and the lack of consultation data items.The main contributions are as follows:1.A Transformer-based SET-RM method is proposed for the distributed representation of Chinese herbal medicines.Due to the non-temporal nature of the arrangement of herbs in TCM prescriptions and the need for overall feature extraction of prescriptions,the model highlights local features by Self-attention LSTM while Transformer extracts the overall features of prescriptions.The experimentally proposed SET-RM method performs better in representation learning relative to the Baseline model.2.Based on the feature vectors obtained from the distributed representation of herbs,we characterize the suitability of herbs in prescriptions by measuring the difference between the predicted results of central herbs and real herbs,and then propose a method for the machine evaluation of prescriptions and their herbs.The results of the study show that the evaluation of potentially inappropriate herbs in prescriptions has significantly lower properties than the overall evaluation of prescriptions.Based on the full discussion and validation of the experimental results,we implemented a TCM prescription review method using a visualization design and incorporated it into a TCM treatment system for guiding physicians to more efficiently identify potentially inappropriate herbs in prescriptions during the prescription review process.3.The model based on the self-attentive mechanism is able to learn feature information with strong explanatory power in the representation learning process.By visualizing the attention distribution of the model in the study of distributed representation of herbs,we found that the model attention tends to focus on the main medicine of the prescription,and this finding can provide new ideas for TCM prescription research.