TCM-KDIF: An Information Interaction Framework Driven by Knowledge–Data and Its Clinical Application in Traditional Chinese Medicine

The effectiveness of traditional Chinese medicine (TCM) has been proved by various researches in decades, especially in the COVID-19 pandemic. Numerous TCM-AI interdisciplinary researches have been proposed for trying to modeling its mechanism and knowledge, assisting efficient decision making of human doctors. Currently, most of the works are focus on supervised learning paradigm. Such methodology leads to the fact that models fail in scenario of rare disease (few samples are available). In this work, we focus on the knowledge–data-oriented mechanism and design a framework enables the model ability to interact the information between knowledge and samples, called TCM-KDIF. We build a TCM knowledge graph (KG) with the TCM concepts (macroscopic) and molecular biology (microcosmic). Based on it, models can interact the information of training samples with external KG by TCM-KDIF. The proposed framework extracts the features of training samples and its related knowledge subgraph first. Then, these two types of information communicate in both directions between samples and knowledge subgraph iteratively. The TCM-KDIF is evaluated on the TCM prescription generation task. The experimental results demonstrate that the TCM-KDIF outperforms all comparison baselines, reduces model’s dependency on training samples, and reveal the possible interact mechanisms between medicine and symptoms.