Unsupervised knowledge discovery in medical databases using relevance networks

Abstract

Increasing amounts of data exist in medical databases. When multiple variables are measured for each case in a data set, there exists an underlying relationship between all pairs of variables, some highly correlated and some not. This report describes a technique that creates networks of related variables, or relevance networks, by dropping links with either too weak correlation or too few data points to defend the relationship. The paper describes how applying this methodology to the domain of laboratory results allows the generation of meaningful relations between types of laboratory tests. These relations could be used as the basis of further exploratory research.