A novel transfer learning model for traditional herbal medicine prescription generation from unstructured resources and knowledge

Traditional Chinese medicine (TCM) is an essential part of the world's traditional medicine. However, there are still many issues in the promotion and development of TCM, such as a lot of unique TCM treatments are taught only between the master and an apprentice in practice, it takes dozens of years for a TCM practitioner to master them and the complicated TCM treatment principles. Intelligent TCM models, as a promising method, can overcome these issues. The performance of previously proposed AI models for intelligent TCM is restricted since they rely on clinical medical records, which are limited, hard to collect, and unavailable for intelligent TCM researchers. In this work, we propose a two-stage transfer learning model to generate TCM prescriptions from a few medical records and TCM documentary resources, called TCMBERT for short. First, the TCMBERT is trained on TCM books. Then, it is fine-tuned on a limited number of medical records to generate TCM prescriptions. The experimental results show that the proposed model outperforms the state-of-the-art methods in all comparison baselines on the TCM prescription generation task. The TCMBERT and the training process can be used in TCM tasks and other medical tasks for dealing with textual resources.