Attention Matching Network for few-shot learning in the syndrome differentiation of cerebral stroke

Treatment based on Syndrome Differentiation is one of the most important characteristics of Traditional Chinese medicine (TCM). In recent years, Artificial Intelligence (AI) and Deep Learning has played a great role in promoting the TCM computer aided diagnosis technology. However, the performance of previous intelligent TCM AI models is limited due to their dependence on clinical medical records. In view of the above problems, this paper improves the framework of traditional few-shot learning framework and proposes a Hybrid Dual-Attentive Matching Network for syndrome differentiation of TCM six meridians, which can explore the law of syndrome differentiation from a small number of medical records and TCM literature resources. To be specific, the model involves two stages: pre-training stage and mate-learning stage. In the first stage, the network trains a classifier on pre-training dataset to get an encoder with some priori knowledge. In the second stage, there are three modules: fusion embedding module for the learning of meta information, hybrid dual-attention module for the getting of key characters and correlation between features and matching module for classification. When testing on our data set with 1134 traditional Chinese medical record of Cerebral Stroke patients, the result of accuracy could be the average precision of 82.39%. The paper is the first attempt of few-shot learning in the syndrome differentiation of six meridians tasks, demonstrating the feasibility of this method in TCM prediction. Besides, the dual-attention mechanism enhances the representation of samples by assigning amplifying the importance weights to each feature, which facilitates the classification performances.