Total correlation explanation of toxic metal concentrations and physiological biomarkers amongst NHANES participants

Abstract

Introduction Unravelling the health effects of multiple pollutants presents scientific and computational challenges. CorEx is an unsupervised learning algorithm that can efficiently discover multiple latent factors in highly multivariate datasets. Here, we used the CorEx algorithm to perform a hypothesis free analysis of demographic, biochemical, and toxic metal biomarker data.

Methods Our data included 77 variables from 2,750 adult participants of the National Health and Nutrition Examination Survey (NHANES 2015-2016). We used an implementation of the CorEx algorithm designed to deal with the features of bioinformatic datasets including mixed data-types. Models were fit for a range of possible latent variables and the best fit model was selected as that which resulted in the largest Total Correlation (TC) after adjustment for the number of parameters. Successive layers of CorEx were run to discovered hierarchical data structure.

Results The CorEx algorithm identified 20 variable clusters at the first layer. For the majority clusters, the associations between variables were consistent with known associations – e.g. gender and the hormones, estradiol and testosterone were included in the first cluster; blood organic mercury and blood total mercury were grouped in cluster 4, and cluster 6 included the liver function enzymes ALT, AST and GGT. At the second layer, 3 branches of were identified reflecting hierarchical structure. The first branch included numerous physiological biomarkers and several exogenous biomarkers. The second branch included a number endogenous and exogenous variables previously associated with hypertension, while the third branch included mercury biomarkers and some related endogenous biomarkers.

Discussion We have demonstrated the CorEx algorithm as a useful tool for hypothesis free exploration of a biomedical dataset. This work extends previous implementations of CorEx by allowing mixed data-types to be modelled and the results showed that CorEx detected meaningful hierarchical structure. CorEx may facilitate exploration of novel datasets in future.