Patterns and Clusters—Multivariate Approach for Interpreting Clinical Chemistry Results

Abstract

Abstract Quantitative assessment of the relationship among clinical observations requires the use of statistical models constructed for this purpose, i.e., multivariate models. Some multivariate methods of pattern cognition are reviewed. A pattern is defined as a combination of laboratory test values simultaneously observed in a patient. A single pattern may be noticed because it is suddenly realized that it is a highly abnormal one. This demonstration requires a multivariate reference region in place of the more commonly used univariate reference intervals. This concept is illustrated by an example in which the multivariate gaussian distribution is used as a model. A group of patterns belonging to a major category of patterns may be delimited, because patterns from this group are in some sense distinct from all other patterns of the category; that is, they form a cluster. Numerical taxonomy comprises the methods by which patterns or symptoms are compared and sorted into groups on the basis of their overall similarity. It is emphasized that many methods for cluster detection are unsound from a theoretical point of view, and if patients are classified the problem of the assignment of weights to variates is still unsolved. It is concluded from a review of the applications that the methods may be useful for the generation of hypotheses but they should not be used as the only methods in the analysis of multivariate data. Additional clustering methods oriented toward one specific goal are reviewed as well.