Research On The Generation Method Of Traditional Chinese Medicine Formula Based On Policy Gradient

Prescription of traditional Chinese medicine is the main method of traditional Chinese medicine.The use of artificial intelligence technology to generate reasonable and effective Chinese medicine prescriptions has always been one of the hot topics in the field of Chinese medicine.It can not only assist doctors in making clinical diagnosis and treatment decisions,but also promote the research and development of new Chinese medicines.However,there are still many challenges in the generation of traditional Chinese medicine prescriptions:(1)the herbs and symptoms in traditional Chinese medicine prescriptions show a serious long-tailed distribution,and most of the herb data in the label head are adjuvant herbs;(2)the multi-label dimension If it is too high,it is difficult to classify,and the dimension of herbs is much higher than the dimension of symptoms;(3)In traditional Chinese medicine prescriptions,the same symptom set generally corresponds to a variety of different prescriptions,that is,the situation of one symptom with multiple prescriptions.Based on the above problems,it is difficult to directly predict traditional Chinese medicine prescriptions through symptom sets.Therefore,based on the TCM prescription benchmark dataset,this paper starts with the composition structure and indications of TCM prescriptions,decomposes the TCM prescription generation problem into three sub-tasks,and reversely explores TCM prescriptions that meet the expected indications.The main research contents of this paper are as follows:(1)Aiming at the prediction of indications of traditional Chinese medicine prescriptions,this paper builds a prescription indication prediction model based on multilabel text classification to predict the main functions and indications of traditional Chinese medicine prescriptions.In this paper,the multi-label classification problem is regarded as a sequence generation problem,and a sequence generation model is constructed to solve the problems of high label dimension,sparse label quantity,and close relationship between labels in TCM prescription data,and introduce beam search,global embedding and other methods to reduce small error.The experimental results show that the prescription indication prediction model can accurately predict the indications of traditional Chinese medicine prescriptions according to the sequence of traditional Chinese medicine prescriptions.Its absolute match rate and Micro F1 value reach 18.68%and 25.27%,respectively,which are overall better than all baseline methods.In addition,using attention mechanism visualization to observe the effect of different herbs on symptoms,the results are basically consistent with the theory of traditional Chinese medicine.(2)Aiming at the generation of Chinese medicine prescriptions that conform to the compatibility law,a prescription generation model is constructed based on the autoregressive language model to learn the herb compatibility laws in Chinese medicine prescriptions,so as to generate Chinese medicine prescriptions that conform to the theory of traditional Chinese medicine.The model consists of an embedding layer,an LSTM layer,and a fully connected layer.The experimental results show that the autoregressive language model achieves 11.48 on the perplexity evaluation index,which is better than the N-Gram language model.In addition,this paper generated 25,703 prescriptions based on the model,and further demonstrated the model from the perspective of traditional Chinese medicine by counting the number of invalid prescriptions,the number of novel prescriptions,the similarity with the original prescription,and the frequency and length of the generated herbs.Rationality and novelty of the generated TCM prescriptions.(3)Aiming at the problems of prescription recommendation and multiple prescriptions for one symptom,an optimal TCM prescription structure exploration framework is proposed to generate TCM prescriptions suitable for the target symptom set.The framework contains pretrained generative and predictive models,and uses a policy gradient algorithm to jointly train the generative and predictive models.This paper conducts experiments on the generation of prescriptions for TCM comorbidities,and the F1 value of this framework reaches 33.37%.The case study shows that the framework can solve the problem of one symptom and multiple parties in the prescription generation task,which further demonstrates the practical value of the model.