Cluster Analysis of Time-Series Medical Data Based on the Trajectory Representation and Multiscale Comparison Techniques

This paper presents a cluster analysis method for multidimensional time-series data on clinical laboratory examinations. Our method represents the time series of test results as trajectories in multidimensional space, and compares their structural similarity by using the multiscale comparison technique. It enables us to find the part-to-part correspondences between two trajectories, taking into account the relationships between different tests. The resultant dissimilarity can be further used with clustering algorithms for finding the groups of similar cases. The method was applied to the cluster analysis of Albumin-Platelet data in the chronic hepatitis dataset. The results denonstrated that it could form interesting groups of cases that have high correspondence to the fibrotic stages.