Multilevels Hidden Markov Models For Temporal Data Mining Weiqiang Lin , Mehmet A. Orgun , and Graham J. Williams Department of Computing, Macquarie University Sydney, NSW 2109, Australia, Email: wlin, mehmet @comp.mq.edu.au CSIRO Mathematical and Information Sciences, GPO Box 664, Canberra ACT 2601, Australia, Email: Graham.Williams @cmis.csiro.au Abstract. This paper describes new temporal data mining techniques for extracting in- formation from temporal health records consisting of time series of diabetic patients’ treatments. In this new method, there are three steps for analyzing patterns from a lon- gitudinal data set. The first step, a structural-based pattern search, to find qualitative patterns (or, structural patterns). The second step performs a value-based search to find quantitative patterns. In the third step we combine results from the first two steps to form new model. The hidden Markov model has the expressive power of both qualita- tive analysis and data quantitative analysis. The global patterns can therefore be identi- fied from a DTS set. Keywords: temporal data mining, discrete-valued time series, similarity patterns, peri- odicity analysis, hidden Markov model 1 Introduction Temporal data mining is concerned with discovering qualitative and quantitative tem- poral patterns in a temporal database or in a discrete-valued time series (DTS) dataset. DTS commonly occur in temporal databases (e.g., the weekly salary of an employee). Recently, there are two kinds of major problems that have been studied in temporal data mining: 1. The similarity problem: finding fully or partially similar patterns in a DTS, and 2. The periodicity problem: finding fully or partially periodic patterns in a DTS. Although there are various results to date on discovering periodic patterns and sim- ilarity patterns in discrete-valued time series (DTS) datasets (e.g. [1]), a general theory and general method of data analysis of discovering patterns for DTS data analysis is not well known. In this paper we describe a new framework for discovering patterns from temporal health records using multilevel hidden Markov model(MHMM). There are three steps for discovering knowledge from the dataset in this approach. The first step of the framework consists of a Markov model analysis for discovering structural (qualitative) patterns. In this step, only the rough shapes of patterns are decided from the DTS. The patterns are grouped into clusters by Nearest Neighbour (NN)to, or the closest candidates of, given patterns among the similar ones selected. In the second