Intelligent Generating Prescription Models Of TCM Based On Deep Learning

With the rapid development of artificial intelligence(AI),the application of deep learning in the medical field becomes more and more widespread.In this context,the government proposed the goal of promoting the modernization,informatization and intellectualization of traditional Chinese medicine(TCM).Although there have been some applications on the integration of AI and TCM,few studies focus on intelligent generating prescriptions(IGP)of TCM,and the only one research did not achieve the expected result.Intelligent generating prescription research is a very important task in the intellectualization of TCM.Therefore,this research did important explorations on building IGP models of TCM,the comparison of different IGP models on precision,recall and F1 score,and extracting sentence features with visualization methods from the trained model.The previous research on IGP model of TCM lacked the pre-training strategy and used complicated training data,so the final result was not good.This study critically inherited the advantages and disadvantages of the previous one,and established the definite directions as follows: making the transfer learning strategy,selecting the pre-training model,and using the small and refined training data which is the generally accepted classic in TCM.Based on this,our research performed the following works:1.According to TCM knowledge,we selected the Treatise on Febrile Diseases and Synopsis of Golden Chamber as based training data,the Zhu Jie Treatise on Febrile Diseases,Synopsis of Golden Chamber Qian Zhu and Ding Zheng Zhong Jing Quan Shu Synopsis of Golden Chamber Zhu as supplementary training data and the Canon of Internal Medicine and Classic on Medical Problems as fine-tuning training data.Meanwhile,the based training data was extended using the EDA data augmentation.2.The pre-training strategy selection was based on word embedding model with Si Ku Quan Shu database,BERT model with Chinese Wikipedia database and Ro BERTa model with Chinese Wikipedia and general extended database.Meanwhile,the TCM-BERT model was fine-tuned with the fine-tuning training data basing on a BERT model.3.We built several deep learning IGP models basing on transfer learning strategy and pretraining model,and made comparisons with several experiments using the based training data and supplementary training data among the IGP models.Finally,the Ro BERTa-large model showed the best performance,and the results were as follows: In based training data,testing precision was 92.22%,recall was 86.71%,F1 score was 89.38%,ten-fold cross-validation precision was 94.5±2.5%,recall was 90.47±4.1% and F1 score was 92.38±2.8%.In supplementary training data,testing precision was 80.82%,recall was 71.50%,F1 score was 75.87%,ten-fold cross-validation precision was 81.85±4.4%,recall was 71.28±3.2% and F1 score was 76.14±3.1%.4.We extracted the sentence language features with visualization methods from the trained Ro BERTa-large model using the classic terms in the Treatise on Febrile Diseases.In the results,we found that the model did not only remember the combination of symptoms,but rather stratified the sentence according to their semantics.Within each layer,it showed that there were some associations between the important symptoms,between the diseases and symptoms and between the commas.With each layer,there were some laws that displayed the dynamic symptom transformation or the recursion.Thus,the stratified analysis may be useful for the further studies on the Treatise on Febrile Diseases.In conclusion,our study did several important works,but there are still some deficiencies.In further research,we suggest that the BERT series models should combine with Knowledge Graph methods to get a better effect.In addition,the application of language feature extracting and visualization methods should be further used on the study of TCM ancient books.