Research On Auxiliary Decision-making Of Pulse Diagnosis Of Traditional Chinese Medicine Based On Machine Learning

In Traditional Chinese Medicine(TCM),pulse diagnosis is one of the main ways to understand the internal changes of the disease.However,this diagnosis method suffers the problem of strong subjectivity.This thesis builds a machine learning model based on the patient’s pulse signal to excavate the potential knowledge in the pulse signal data,assist doctors in clinical diagnosis decision-making,and facilitate the objective development of pulse diagnosis.The research content of the thesis includes pulse signal based single-label disease prediction,multi-label disease prediction,unsupervised pulse signal representation learning,and the realization of a pulse diagnosis assistant decisionmaking platform.The main work of this thesis is presented as follows:1.To solve the problem that the pulse signal classification results are not interpretable,a multi-scale discriminant sub-sequence distance measurement method is proposed.First,the significant pulse signal sub-segments are obtained based on the discriminating sub-sequence positioning.Then,the distance between the signal subsegment pair is calculated in two modalities.Next,the distance measurement results of multiple discriminant signal sub-segments are calculated as the features to establish multi-scale distance feature space.Finally,the pulse signal classification is realized by the support vector machine.Experimental results show that the method achieves the best performance in 7 of 11 public data sets.The classification accuracy on the pulse signal data set is 0.917.2.Considering the fact that patients may have multiple diseases at the same time in clinical diagnosis and the importance of pulsation rhythm in pulse diagnosis,a multi-scale periodic information fusion network model for multi-label disease prediction is proposed.The model is composed of global signal convolutional encoder,periodic signal convolutional encoder,information fusion gate,and attention module.The model can extract the depth features of pulse signals at different scales,fuse global features with intra-period features,and assign weights to different periodic features.Experimental results show that the accuracy,recall,and F1 of the model reaches 0.923,0.941,and 0.932,respectively.3.In order to discover the potential patterns in the pulse signal and make full use of the unlabeled pulse signal data,an unsupervised representation learning method based on fusion loss is proposed to cluster the unlabeled pulse signals.The algorithm integrates pulse signal dimensionality reduction and pulse signal sequence clustering into an unsupervised end-to-end learning framework.The convolutional autoencoders are used to obtain potential feature representations to reduce the time dimension of the pulse signal.To form a cluster structure,the K-means loss is adopted to guide the potential feature representation learning.Experimental results show that the Rand index and the normalized mutual information of the proposed method reaches 0.903 and 0.653,respectively.4.A pulse diagnosis assistant decision-making platform is designed and implemented.The platform development is based on the Java open source lightweight framework using a Browser/Server architecture and My SQL database.The functions of the platform include system management,data management,intelligent auxiliary diagnosis,and diagnosis and treatment rule mining,which can effectively assist doctors in making diagnosis decisions and promote the objective development of pulse diagnosis.