世界科学技术-中医药现代化2025，Vol.27Issue(6)：1750-1757,8.DOI:10.11842/wst.20240920003
基于深度学习Transformer和迁移学习的声诊体质辨识研究
Research on Sound Diagnosis Constitution Identification Based on Deep Learning Transformer and Transfer Learning
摘要
Abstract
Objective The identification of TCM constitution plays an important role in"treating and preventing diseases"of TCM.At present,the identification of damp-heat constitution and balanced constitution is mostly determined by questionnaire,and subjective factors have a great influence.Aiming at the identification of damp-heat constitution and balanced constitution in TCM,this paper utilizes voice signal to automatically realize the constitution identification task,in order to provide assistance for the clinical identification of TCM constitution.Methods Based on deep learning Transformer and transfer learning,a pure attentional mechanism model was designed for the identification of constitution in TCM sound diagnosis.We collected 700 voices from 34 subjects,pre-processed the voice data to obtain the corresponding Mayer spectrum diagram,and used the Transformer model pre-trained based on the public data set to improve the performance of the model for audio classification.Results The accuracy of the experimental results was 83.33%,the AUC was 92.16%,the sensitivity was 80.25%,and the specificity was 87.03%.Compared with the Convolutional Neural Network(CNN),the performance of the deep learning model was better.Conclusion In this paper,the damp-heat constitution and balanced constitution identification model Transformer has achieved better identification effect,indicating that it can improve the efficiency of TCM acoustic diagnosis of constitution identification,and promote the objective and intelligent development of constitution identification.关键词
中医声诊/体质辨识/深度学习/TransformerKey words
Sound diagnosis of TCM/Constitution identification/Deep learning/Transformer分类
医药卫生引用本文复制引用
门韶洋,陈绿洁,黄晓梅,文晓冰,林传权,张洪来..基于深度学习Transformer和迁移学习的声诊体质辨识研究[J].世界科学技术-中医药现代化,2025,27(6):1750-1757,8.基金项目
科技部"中医药现代化研究"重点专项(2019YFC1710402):辨证论治四诊延伸性电子化数据采集系统及设备集成,负责人:张洪来 （2019YFC1710402）
国家自然科学基金委员会专项项目(T2341009):基于"唾液生物标志物-舌苔微生物-舌象映射脾运化功能"构建2型糖尿病的智能预警模型,负责人:林传权 （T2341009）
广东省基础与应用基础研究基金委员会面上项目(2023A1515011316):基于声诊知识图谱和联合神经网络的多源声音智能识别方法研究及其在中医体质辨识中的应用,负责人:门韶洋. （2023A1515011316）
