Clustering technique is an effective tool for medical data analysis as it can work for disease prediction, diagnosis record mining, medical image segmentation, and so on. This paper studies the kernel-based clustering method which can conduct nonlinear partition on input patterns and addresses two challenging issues in unsupervised learning environment: feature relevance estimate and cluster number selection. Specifically, a kernel-based competitive learning paradigm is presented for nonlinear clustering analysis. To distinguish the relevance of different features, a weight variable is associated with each feature to quantify the feature's contribution to the whole cluster structure. Subsequently, the feature weights and cluster assignment are updated alternately during the learning process so that the relevance of features and cluster membership can be jointly optimized. Moreover, to solve the problem of cluster number selection, the cooperation mechanism is further introduced into the presented learning framework and a new kernel clustering algorithm which can automatically select the most appropriate cluster number is educed. The performance of proposed method is demonstrated by the experiments on different medical data sets.