两阶段多标签分类探索中医证素辨证规律

Exploration of the Rules of Traditional Chinese Medicine Syndrome Elements Differentiation by Two-stage Multi-label Classification

探索证素辨证规律能更好地辅助临床决策和促进中医辨证理论的传承。中医文本句式结构复杂、表述标准不一,难以匹配符号规则,且神经网络黑盒特性又难以直接解释其辨证过程。为探索中医证素辨证规律,第一阶段使用神经网络模型对证素进行多标签分类,通过稀疏注意力捕获与证素相关的关键词及其权重生成证素表征;第二阶段使用随机森林对融入相关证素标签的证素表征进行分类训练,后对随机森林规则提取以探索辨证规律,提高证素辨证的可解释性。实验结果表明,该方法提升了证素辨识的准确率,同时F1保持较高水平,有利于探索证素辨证规律。

Exploring the rules of syndrome element differentiation can better assist clinical decision-making and promote the inheritance of TCM syndrome differentiation theory.The sentence structure of TCM texts is complex,with varying expression standards,making it difficult to match symbol rules,and the black box characteristics of neural networks are difficult to directly explain their differentiation process.In order to explore the rules of syndrome element differentiation in TCM,it uses a neural network model to classify syndrome elements with multiple labels in the first stage,and generates syndrome element representation by capturing keywords and their weights related to syndrome elements through sparse attention.In the second stage,the random forest is used to conduct classified training on syndrome element representation incorporating relevant syndrome element labels,and then random forest rules are extracted to explore the syndrome differentiation rules and improve the interpretability of syndrome element differentiation.The experimental results show that this method improves the accuracy of syndrome element identification,while maintaining a high level of F1,which is conductive to exploring the rules of syndrome element differentiation.