Design Of Bimodal Information Acquisition System Of Traditional Chinese Medicine And Research On Sub-health Detection

With the development of the times,people pay more and more attention to health problems.According to the survey,70% of people are in sub-health state.Traditional Chinese medicine is an important "preventive treatment of disease" and non-invasive diagnosis mode,which is very consistent with the definition that sub-health is based on personal feelings without obvious disease.With the deepening of the objective research of TCM diagnosis in recent years,it provides unlimited possibility for TCM to identify sub-health state.Based on the theory of traditional Chinese medicine,a compact bimodal physiological information acquisition system is designed,which mainly includes pulse acquisition subsystem and tongue acquisition subsystem.The pulse acquisition subsystem is equipped with three independent pressure adjustable pulse sensors,which can restore the doctor pulse diagnosis mode as much as possible,and realize the good performance of objectification and visualization of pulse diagnosis.The image acquisition subsystem precisely fixes the distance and height between the chin and the camera,and takes pictures of the tongue under sufficient illumination.More importantly,a simple and convenient infrared temperature sensor is used to obtain the body temperature data.This paper uses this system to collect the tongue and pulse data of sub-health and healthy students.In the part of pulse signal preprocessing,wavelet analysis is used to denoise,then normalization and period segmentation are carried out,and the time-domain and frequencydomain features of pulse signal are extracted.In the part of tongue image preprocessing,the tongue segmentation based on deep learning is studied,and then the reflective point detection and repair,the separation of moss and texture,color feature and texture feature extraction are carried out.In this study,methods such as Chi2-Tree,LR12,and SVMREF were used for feature selection.Four machine learning algorithms,such as Support Vector Machine,Random Forest,Light GBM and Cat Boost,are used to train and learn pulse feature data set,tongue feature data set,pulse + tongue + human feature data set and feature selection data set.According to the research comparison,Light GBM based on Chi2-Tree is the best comprehensive classification model,and its average recognition accuracy,sensitivity,specificity and micro average AUC area under ROC curve are basically better than other classification models.