Pulse Classification Based On Hidden Markov Model

Traditional Chinese Medicine (TCM) has been proved to be a worthwhile and clinically valid method over 2000 years. Traditional Chinese Pulse Diagnosis (TCPD) is the most distinctive one of the four Traditional Chinese Medical Diagnoses. However, the traditional way to diagnosis the disease is very vague and objective that is palpate the patient's pulses with fingertips and use the experience to give out a result. That restricted its application and further development. It's necessary to make use of the modern technology to pulse diagnosis, the objectifying of TCPD is urgent.A lot of research on time-domain features has been done and give out some good result, while research was mostly on the local feature not the global feature. This paper takes global feature into consideration. First, AR model parameters are used to represent the small segment of the pulse signal. Then Hidden Markov Model is used to simulate the relationship between the segments. We take both the specific features and overall characteristics of pulse signals into account; it is helpful for the diversity of the pulse signal and also meets the overall concept of TCM.The environment electromagnetic interference may introduce high frequency noise, which is called the non-stationary white noise. In this paper, we use db3 wavelet to remove it, the baseline drift caused by the collector's breathing and body movement is also a big problem. The cubic spline interpolation is used to remove it. Also pulse signals are normalized due to the different environment and different operations.In the feature extracting stage, we use the AR model which is superior in the time series domain to express the segments. And the mean characteristics and the first order difference features of the small segment are also used as the features. In the pattern classification stage, Hidden Markov Model which is very suitable for modeling the time signals is used in the experiments. Segmental K-means method and F_L algorithm are used to improve the HMM model. To avoid the randomicity introduced by segmental K-means method, we combined F_L algorithm and the segmental K-means method together. We also compared the result to other classification methods. The classification result shows that the joint algorithm obtained the best results. Finally, this method is applied to disease data set, and also received good results.