Guided data reduction for flow cytometry

A statistical procedure for identifying subpopulations in large samples is discussed in the context of the analysis of flow cytometric data. The PINDEX algorithm is proposed for analyzing such multimodal data. PINDEX is a clustering algorithm which considers the covariance structure of clusters. When such an algorithm is run in a brute force manner, the number of runs required for convergence may involve computation which, for large samples, is prohibitive. A solution to this problem using starting centroids chosen by an expert system is proposed. Sample runs were made using data sets for lymphoid cells; these data sets are of modest dimensionality and contain four subpopulations. The results demonstrate that substantive economy of processing can be achieved by having the expert system provide start-up cluster centroids close to the final coordinates