Evolutionary Clustering Algorithm And Its Application To Medical Data Analysis

Recently, evolutionary clustering is an identified new and hot research topic in data mining. In medical data analysis field, mining medical data based on clustering algorithms not only improves the analysis accuracy of the typical huge amount of medical data, but also helps discover the hidden patterns and knowledge, promoting the further better understanding and the advance of the medical science.Traditionally, clustering on medical datasets usually is based on static views, using static clustering algorithms such as K-means andthe static spectral clustering algorithms. All of them just cluster datasets without considering the time information. With the time information having to be considered, the traditional approach applies the static clustering methods to the dataset collected at each time repeatedly, ignoring the interrelationships among the datasets of different times.In this thesis, we study evolutionary clustering algorithms to cluster dynamic medical data. Considering the relationship information of medical data at different times, the first algorithm we present here is based on the K-means algorithm. This method takes the history information of the medical data into consideration when we cluster the current data. Thus, it greatly improves the performance accuracy of clustering the dynamic medical data. Meanwhile, for most of current clustering algorithms, the number of clusters is assumed to be given in advance. Clearly this assumption is often too strong.. Here we present another algorithm based on the Dirichlet Process and the Hidden Markov Model to maximize the posterior probability to cluster the dynamic medical data. During clustering, the number of clusters can be adaptively updated as the new data arrive. We use the above two methods to cluster the medical data collected from a group of hepatitis patients using the artificial liver functions in comparison with the control group of patients without using the artificial liver functions. Comparing with the traditional methods, the experimental results show that both of the above two methods obtain a better performance.