Identifying subgroups of eating behavior traits unrelated to obesity using functional connectivity and feature representation learning

Abstract

Eating behavior is highly heterogeneous across individuals and cannot be fully explained using only the degree of obesity. We utilized unsupervised machine learning and functional connectivity measures to explore the heterogeneity of eating behaviors measured by a self-assessment instrument using 424 healthy adults (mean ± standard deviation [SD] age = 47.07 ± 18.89 years; 67% female). We generated low-dimensional representations of functional connectivity using resting-state functional magnetic resonance imaging and estimated latent features using the feature representation capabilities of an autoencoder by nonlinearly compressing the functional connectivity information. The clustering approaches applied to latent features identified three distinct subgroups. The subgroups exhibited different levels of hunger traits, while their body mass indices were comparable. The results were replicated in an independent dataset consisting of 212 participants (mean ± SD age = 38.97 ± 19.80 years; 35% female). The model interpretation technique of integrated gradients revealed that the between-group differences in the integrated gradient maps were associated with functional reorganization in heteromodal association and limbic cortices and reward-related subcortical structures such as the accumbens, amygdala, and caudate. The cognitive decoding analysis revealed that these systems are associated with reward- and emotion-related systems. Our findings provide insights into the macroscopic brain organization of eating behavior-related subgroups independent of obesity.

Keywords: autoencoder; eating behavior; functional connectivity; integrated gradient; manifold learning; representation learning; subgroup.

FIGURE 1

Subgroup identification using the dimensionality reduction technique and autoencoder‐based feature representation. (a) Schematic of functional connectome organization (left) and group averaged functional connectivity matrix (middle) are reported. Template eigenvectors were generated using dimensionality reduction techniques, and three dominant eigenvectors (E1, E2, E3) were selected (right). (b) The autoencoder model learned latent features of the eigenvectors after controlling for age and sex (left top), and loss values are plotted for each epoch (middle). We calculated linear correlations between the original (E) and reconstructed (E') eigenvectors of the test dataset, and correlation coefficients across the subjects are reported with mean ± SD (right). We defined subgroups using the latent features of the autoencoder, where the number of clusters was determined using the consensus coefficient (left bottom). (c) Distribution of BMI and eating behavior scores of each subgroup is plotted. Significant differences in scores between subgroup pairs are indicated by asterisks. BMI, body mass index; FDR, false discovery rate.

FIGURE 2

Characteristics of latent features using the integrated gradient technique. (a) The integrated gradient technique estimates the attribution of input toward predicting the output by averaging contributions while changing input intensities. (b) Spatial maps of each eigenvector and results of the integrated gradient technique are plotted on the brain surfaces. (c) Effects of the integrated gradient are summarized according to functional communities. IG, integrated gradient.

FIGURE 3

Between‐group differences in the cortico‐cortical and subcortico‐cortical connectivity. (a) Between‐group differences in the cortico‐cortical connectivity based on the integrated gradient maps among the subgroups are visualized on the cortical surfaces, where the findings were multiple comparisons corrected using FDR <0.05. Effects were stratified according to seven intrinsic functional communities. (b) We visualized between‐group differences in subcortico‐cortical nodal connectivity strengths and stratified the effects according to each subcortical structure. FDR, false discovery rate.

FIGURE 4

Cognitive associations. (a) We conducted cognitive decoding using the F‐statistic map of cortico‐cortical and subcortico‐cortical connectivity differences across the subgroups using Neurosynth. (b) Correlation coefficients between the between‐group difference maps and 24 different cognitive state maps are shown with bar plots.

FIGURE 5

Reproducibility analyses. (a) Distribution of BMI and eating behavior scores of each subgroup using the replication dataset are shown. Significant differences between the scores of subgroups are indicated by asterisks. (b) We compared profiles of the BMI and eating behavior scores of two different datasets. BMI, body mass index; FDR, false discovery rate.