Mixture model-based statistical pattern recognition of clustered or longitudinal data

Abstract

Mixture models implemented via the expectationmaximization (EM) algorithm are being increasingly used in a wide range of problems in statistical pattern recognition. For many applied problems in medical and health research, the data collected may exhibit a hierarchical structure. The independence assumption in the maximum likelihood (ML) learning of mixture models is no longer valid. Ignoring the correlation between hierarchically structured data can lead to misleading pattern recognition. In this paper, we consider the extension of Gaussian mixtures to incorporate data hierarchies via the linear mixed-effects model (LMM). Clustered and longitudinal data hierarchy settings in medical and biological research are considered.