Clustering Human Wrist Pulses For Traditional Chinese Medicine

In this thesis, we study the clustering methods for human wrist pulses. First, each collected pulse is preprocessed and represented by one period, by which features are extracted with characteristic points and sparse coding. Then the minimum number of clusters is estimated by observing the visualization result obtained from principal component analysis, and the maximum number of clusters is set to a reasonable value.Given the feature and the range of cluster number, six clustering algorithms, namely k-means, k-medoids, Gaussian mixture model, spectral clustering, kernel k-means and self-organized map, are initialized to generate a number of clustering alternatives. The goodness of these alternatives are sequentially comprehensively evaluated by 9 cluster validity indices, e.g. Calinski-Harabasz, C Index, Dunn's, Davies-Bouldin, Ray-Turi,Silhouette, SD Validity, Wemmert-GaKaraski, Xie-Beni. Finally, selecting a good clustering is modeled as a multiple criteria decision making(MCDM) problem, solved by the technique for order preference by similarity to ideal solution(TOPSIS). The optimal one is picked as the final clustering. According to the TOPSIS rank, clustering the data set into 11 clusters via kernel k-means is the optimal clustering. The optimal clustering result can be further aggregated into two groups, namely the healthy and the unhealthy,laying the foundation for health identification in traditional Chinese medicine.