Feature Extraction And Selection Research Of Sub-health Recognition Based On Pulse Wave

Sub-health lies between health and disease,and it is the transition stage before illness.The effective diagnosis of sub-health is of great help to the prevention of disease.The body's all weak changes of physiological and psychological can be reflected in the pulse.Through the pulse wave to study whether the human body is in sub-health state has attracted more and more attention.Sub-health recognition based on pulse wave includes five parts:signal acquisition,preprocessing,feature extraction,feature selection and sub-health recognition.The following are the main tasks of this paper:The health and sub-health state of people according to the sub-health assessment table were judged in the thesis,and Hua Ke HK-2000C digital pulse sensor was used to collect the two kinds of pulse wave signals.The Complete Ensemble Empirical Mode Decomposition with Adaptive Noise combining with Permutation Entropy was used to de-noise pulse waves.Introducing the fourier series coefficient and multiscale entropy into the feature set was proposed in thesis,then 54 pulse wave features from three dimensions of time domain,frequency domain and nonlinearity were extracted.A hybrid feature selection algorithm(Laplacian mRMR SVM-RFE,LMSR)was proposed to achieve dimensionality reduction of the feature matrix.The algorithm calculated the Laplacian score and the maximum correlation minimum redundancy(mRMR)score for each feature in the original feature set,and obtained feature subsets with scores below the average of Laplacian scores and above the average of mRMR scores respectively.Taking the union and intersection of the two subsets as inputs of the Support Vector Machine Recursive Feature Elimination Algorithm respectively.Two feature sorting sets in descending order of importance were obtained,which contain 40 and 18 features respectively.A subset of the increasing dimension is selected from the feature sorting sets to train the SVM classifier,and the corresponding accurate recognition rate was calculated.Experimental results showed that the recognition rate and reduction effect in dimension are better than the intersection when subset was merged in a union in LMSR algorithm.Besides,when the feature set consisted of 6 features of multi-scale fuzzy entropy(1 features)and Fourier series coefficient(5 features),the classifier had the highest accurate recognition rate,which reached 81.58%.