Multi-view manifold learning of human brain state trajectories

Abstract

The complexity and intelligence of the brain give the illusion that measurements of brain activity will have intractably high dimensionality, rifewith collection and biological noise. Nonlinear dimensionality reduction methods like UMAP and t-SNE have proven useful for high-throughput biomedical data. However, they have not been used extensively for brain imaging data such as from functional magnetic resonance imaging (fMRI), a noninvasive, secondary measure of neural activity over time containing redundancy and co-modulation from neural population activity. Here we introduce a nonlinear manifold learning algorithm for timeseries data like fMRI, called temporal potential of heat diffusion for affinity-based transition embedding (T-PHATE). In addition to recovering a lower intrinsic dimensionality from timeseries data, T-PHATE exploits autocorrelative structure within the data to faithfully denoise dynamic signals and learn activation manifolds. We empirically validate T-PHATE on three human fMRI datasets, showing that T-PHATE significantly improves data visualization, classification, and segmentation of the data relative to several other state-of-the-art dimensionality reduction benchmarks. These notable improvements suggest many potential applications of T-PHATE to other high-dimensional datasets of temporally-diffuse processes.