Wrist Pulse Diagnosis Based On Complex Network And Scattering Convolution Network

Pulse signal contains rich information about the health status and pulse diagnosis isone of the four important diagnosis methods in the traditional Chinese medicine (TCM).It has been successful applied for thousands of years, however, pulse diagnosis is asubjective skill and the diagnosis result relies on the personal experience of thepractitioner. To overcome these limitations, computerized pulse diagnosis has beenproposed in recent years. Pulse diagnosis usually involves in four modules, i.e., signalacquisition, preprocessing, feature extraction and classification. In our work, we focuson the preprocessing and feature extraction modules. First, we add a pulse signal qualityassessment process in the preprocessing module to remove the pulse signal with lowquality. And then we proposed two new effective feature extraction methods based onthe complex network and scattering convolution network.The common preprocessing process usually contains high frequency interferenceremoval part and baseline wander removal part, which aims to improve the quality ofthe sampled pulse signals. However pulse signal quality is not evaluated, therefore can’tensure the quality of the signal used in model training. In this work we proposed theentropy of second order difference method to evaluate the quality of the pulse signalfrom the perspective of the complexity, and experiment results show that this method isvery effective, leading to90%of the low quality signals in the dataset were removedusing our proposed quality evaluation method.We also proposed two feature extraction methods in this work i.e., the complexnetwork method and scattering convolution network method. Complex network methodfocuses on the extraction of inter-cycle variations features of pulse signal, whichprovides useful information in the diagnosis of the diseases with period variations of thepulse signal, such as arrhythmia. Scattering convolution network is a feature extractionmethod which has many good traits, such as stability to deformation and informationpreservation. The experiment results show that the complex network feature hasadvantage in the arrhythmia diagnosis and scattering convolution network performbetter than other features in the other experiments.Furthermore, based on these pervious methods, we in this work developed a pulsediagnosis system, including the pulse signal quality evaluation module, feature extraction module that can extract complex network feature and scattering convolutionfeature and classification module with SVM classifier. The system can be used in pulsesignal quality evaluation and classification.