Incorporating Ontological Knowledge into the Multi-Label Classification of Traditional Chinese Medicine Symptom Entities Research (Preprint)

Abstract

In the realm of AI-assisted traditional Chinese medicine (TCM) syndrome differentiation and disease diagnosis, precise symptom recognition and classification pose significant challenges. This is because TCM heavily relies on a nuanced understanding of symptoms to guide treatment decisions. However, current entity recognition models grapple with a range of issues, including limitations stemming from label space, resource-intensive computations, a lack of domain expertise to cover diverse symptom descriptions, and an over-reliance on large annotated datasets. Furthermore, the world of TCM symptom labels is vast and complex, with intricate correlations and the added complexity of label imbalance, all contributing to the complexity of the task.

The goal of this study is to tackle the challenges associated with multi-class symptom entity recognition within the realm of TCM symptoms. To achieve this, we propose a two-stage entity classification approach and leverage ontology knowledge to enhance entity classification. Our approach aims to reduce the model's dependency on annotated data and address issues such as label imbalance and the expansive label space.

We introduce an innovative multi-label entity classification model designed for accurate TCM symptom recognition. We establish a comprehensive TCM symptom ontology framework to standardize symptom descriptions. We use the BERT+BiLSTM+CRF method to identify multiple symptom entities in TCM medical records. In order to gain a deeper understanding of the relationships between multiple symptom entities within the text and the connections between different category labels, we introduce a multi-level correlation feature fusion module. Finally, we adopt a multi-label classification method based on a hierarchical label tree, effectively mitigating the challenges associated with label imbalance within TCM text.

Using authentic Qihuang TCM electronic medical records, our model significantly enhances efficiency and accuracy in multi-label symptom classification, achieving a Hamming Loss of 2.932 * 10^-2 and a Micro-F1 score of 0.8452.

Our study delivers several notable contributions. Firstly, we construct a comprehensive TCM symptom ontology framework, successfully incorporating TCM domain knowledge into the model to bolster its foundational features. Secondly, we employ a multi-label classification approach for entity recognition, capturing the multiple labels and intricate relationships of symptom entities in TCM texts with heightened accuracy. Simultaneously, we introduce a hierarchical label tree to effectively mitigate the impact of symptom label imbalance on the model. Lastly, we introduce a multi-level correlation feature fusion module that comprehensively captures textual information, thereby improving model performance, enhancing label clustering, and ultimately elevating the overall quality and efficiency of the model. These contributions provide an effective methodology for TCM symptom extraction and are poised to make a significant impact on TCM research and practice.