Frequency-Based Temporal Pattern Mining in Health Data

The low occurrence rate of adverse drug reactions makes it difficult to identify the risk factors from straightforward application of frequent pattern discovery in large databases. In this paper, we are interested in developing a data mining strategy that can fully utilize the information around rare events in sequence data in order to measure the multiple occurrences of patterns in the whole period of target and non-target data. We define an interestingness measure which exploits the difference between frequency of patterns in target and non-target sequence data. The proposed strategy guarantees the easy generation of candidate patterns from the target sequence data by applying existing association mining algorithms. Then these patterns can be evaluated by comparing their frequency in the target and non-target data. We also propose a ranking algorithm that takes into account both the rank of patterns as determined by the interestingness measure and the support in the target population, which can prune the patterns greatly and highlight more interesting results. Experimental results of a case study on angioedema show the usefulness of the proposed approach.