Clustering time-series medical databases based on the improved multiscale matching

This paper presents a novel method called modified multiscale matching, that enable us to multiscale structural comparison of irregularly-sampled, different-length time series like medical data. We revised the conventional multiscale matching algorithm so that it produces sequence dissimilarity that can be further used for clustering. The main improvements are: (1) introduction of a new segment representation that elude the problem of shrinkage at high scales, (2) introduction of a new dissimilarity measure that directly reflects the dissimilarity of sequence values. We examined the usefulness of the method on the cylinder-bell-funnel dataset and chronic hepatitis dataset. The results demonstrated that the dissimilarity matrix produced by the proposed method, combined with conventional clustering techniques, lead to the successful clustering for both synthetic and real-world data.