LACE-UP: An ensemble machine-learning method for health subtype classification on multidimensional binary data

Abstract

Disease and behavior subtype identification is of significant interest in biomedical research. However, in many settings, subtype discovery is limited by a lack of robust statistical clustering methods appropriate for binary data. Here, we introduce LACE-UP [latent class analysis ensembled with UMAP (uniform manifold approximation and projection) and PCA (principal components analysis)], an ensemble machine-learning method for clustering multidimensional binary data that does not require prespecifying the number of clusters and is robust to realistic data settings, such as the correlation of variables observed from the same individual and the inclusion of variables unrelated to the underlying subtype. The method ensembles latent class analysis, a model-based clustering method; principal components analysis, a spectral signal processing method; and UMAP, a cutting-edge model-free dimensionality reduction algorithm. In simulations, LACE-UP outperforms gold-standard techniques across a variety of realistic scenarios, including in the presence of correlated and extraneous data. We apply LACE-UP to dietary behavior data from the UK Biobank to demonstrate its power to uncover interpretable dietary subtypes that are associated with lipids and cardiovascular risk.