Coupled generator decomposition for fusion of electro- and magnetoencephalography data

TL;DR

The paper addresses multisubject, multisource fusion of neural data (EEG/MEG) by introducing a coupled generator decomposition that enforces a shared latent generator while allowing source-specific mixing across subjects and modalities.It extends sparse principal component analysis to a multisource setting, and implements a PyTorch-based gradient optimization that matches traditional quadratic programming in accuracy but with greatly reduced computation time.Through a multimodal multisubject face-perception study, the method reveals a ~170 ms fusiform-related component that differs between real and scrambled faces, and demonstrates competitive performance relative to archetypal analyses, with a practical toolbox for data fusion.Overall, the approach provides a flexible, fast framework for extracting shared and source-specific neural features in complex, multimodal neuroimaging data, with broad potential extensions to other analysis paradigms.

Abstract

Data fusion modeling can identify common features across diverse data sources while accounting for source-specific variability. Here we introduce the concept of a \textit{coupled generator decomposition} and demonstrate how it generalizes sparse principal component analysis (SPCA) for data fusion. Leveraging data from a multisubject, multimodal (electro- and magnetoencephalography (EEG and MEG)) neuroimaging experiment, we demonstrate the efficacy of the framework in identifying common features in response to face perception stimuli, while accommodating modality- and subject-specific variability. Through split-half cross-validation of EEG/MEG trials, we investigate the optimal model order and regularization strengths for models of varying complexity, comparing these to a group-level model assuming shared brain responses to stimuli. Our findings reveal altered $\sim170ms$ fusiform face area activation for scrambled faces, as opposed to real faces, particularly evident in the multimodal, multisubject model. Model parameters were inferred using stochastic optimization in PyTorch, demonstrating comparable performance to conventional quadratic programming inference for SPCA but with considerably faster execution. We provide an easily accessible toolbox for coupled generator decomposition that includes data fusion for SPCA, archetypal analysis and directional archetypal analysis. Overall, our approach offers a promising new avenue for data fusion.

Figures (4)

• Figure 1: Variability in ERP waveform across subjects for a chosen right occipital EEG and MEG channel.
• Figure 2: Boxplot of model convergence across stochastic optimization in PyTorch and quadratic programming inference techniques, as well as effect of initialization on validation loss. Each model was run across 10 random initializations using a regularization coefficient pair determined on the lowest attained validation loss. We note that there are, in fact, three boxes for each number of components; the boxes are very small due to diminishing variability.
• Figure 3: Lineplot of model performance across number of components, $K$. The models were evaluated on a test set using regularization coefficients determined on the validation set, and the average and standard deviation (shaded area) across 10 random initializations is shown. We note that the variance across random initializations is too small to be clearly distribguishable.
• Figure 4: Sparse PCA and archetypal analysis results on data from a multimodal multisubject face perception neuroimaging experiment. The coupled generator decomposition computes a shared generator matrix ${\mathbf G}$ (top row) and subject and modality-specific topographical maps ${\mathbf X}{\mathbf G}$, here shown as an average across subjects. For the group formulation, the mixing matrix ${\mathbf S}$ is also shared, while we show the average mixing matrix across subjects for the multimodal, multisubject models. EEG units are in $\mu V$ and MEG units are in $fT$.