Extracting Approximate Biclusters/Patterns from Time Series Medical Data using Suffix Trees

Abstract

Time series medical data contains many null values and is collected over a long period of time. The focus is on extracting longer decreasing / increasing patterns/biclusters that may be of interest to medical experts in analysing drug responses and therapies, as well as predicting certain disease occurences. We apply the technique of biclustering to extract new, interesting patterns from this data. Given the data for each patient, we discretize it to obtain a symbolic representation using statistical methods. We then proceed to efficiently construct a compact generalized suffix tree over the entire dataset. The algorithm presented in this work extends the problem of common motif searching as applied in microarray experiments to extract approximate biclusters from within the suffix tree utilizing a form of string edit distance restricted to substitution and deletion, and the concept of valid models.