Statistics and Its Interface

Volume 17 (2024)

Number 3

Associations between EEG-defined subgroups and antidepressant response: a joint mixture of probabilistic multilinear principal component analysis modeling approach

Pages: 397 – 411

DOI: https://dx.doi.org/10.4310/23-SII780

Authors

Peng Yu (University of Alberta)

Kaiqiong Zhao (University of Alberta)

Bei Jiang (University of Alberta)

Eva Petkova (New York University)

Thaddeus Tarpey (New York University)

R. Todd Ogden (Columbia University)

Abstract

Joint mixture models have become popular tools to uncover distinct latent subgroups in some manifest variables and to simultaneously relate such heterogeneity to an outcome of interest, while explicitly taking into account the uncertainty of latent class memberships. In this paper, we extend existing latent mixture models to incorporate matrix-variate data built on a mixture model for probabilistic multilinear principle component analysis. The work is motivated by a depression study to investigate the patient heterogeneity based on baseline electroencephalograph (EEG) and the association of such EEG defined subgroups and the antidepressant response. Specifically, there are three levels of structure in our proposed model. First, the uncertainty of latent class membership in the matrix-variate EEG data is specified through a multinomial logistic model. Second, the class-specific EEG data is assumed to follow a probabilistic multilinear principle component analysis model. Third, under the assumption of conditional independence given the latent class membership, the association of the baseline EEG and the antidepressant response is established through the latent classes of baseline EEG. Applying the proposed model to our motivating depression study, four distinct patient subpopulations are identified that differ with respect to their baseline EEG patterns and the risk for positively responding to antidepressant treatment. In contrast, other existing clustering methods cannot lead to findings of such patient subpopulations.

Keywords

matrix-variate, latent class, multilinear principle component analysis, joint modeling, antidepressant response

Received 2 December 2022

Accepted 28 January 2023

Published 19 July 2024