Mining Interesting Patterns in Time-series Medical Databases: A Hybrid Approach of Multiscale Matching and Rough Clustering

Abstract In this paper, we present an analysis method of time-series laboratory examination data based on multiscale matching and rough clustering. We obtain similarity of sequences by multi-scale matching, which compares two sequences throughout various scales of view. It has an advantage that connectivity of segments is preserved in the matching results even when the partial segments are obtained from different scales. Given relative similarity of the sequences, we cluster them by a rough-set based clustering technique. It groups the sequences based on their indiscernibility and has an ability to produce interpretable clusters without calculating the centroid or variance of a cluster. In the experiments we demonstrate that the features of patterns were successfully captured by this hybrid approach.