Multi-view manifold learning of human brain-state trajectories

Area of Science:

• Neuroscience
• Data Science
• Computational Biology

Background:

• Human brain activity is complex and often appears high-dimensional.
• Existing nonlinear dimensionality reduction methods struggle with the dynamic structure of brain activity data, limiting their use with functional magnetic resonance imaging (fMRI).

Purpose of the Study:

• To introduce Temporal Potential of Heat-Diffusion for Affinity-based Transition Embedding (T-PHATE), a novel nonlinear manifold learning method for time-series data.
• To demonstrate T-PHATE's ability to denoise, reveal dynamic trajectories, and improve analysis of fMRI data.

Main Methods:

• Developed T-PHATE, a nonlinear manifold learning technique for time-series data.
• Applied T-PHATE to three functional magnetic resonance imaging (fMRI) datasets.
• Compared T-PHATE's performance against state-of-the-art dimensionality reduction benchmarks.

Main Results:

• T-PHATE recovers low-dimensional intrinsic manifold geometry from time-series data.
• T-PHATE effectively denoises data and reveals dynamic trajectories by exploiting autocorrelative structure.
• T-PHATE significantly improved data visualization, classification, and segmentation compared to existing methods on fMRI datasets.

Conclusions:

• T-PHATE is a powerful tool for analyzing high-dimensional time-series data, particularly fMRI.
• The method's ability to preserve dynamic structure offers significant advantages over current techniques.
• T-PHATE has broad potential applications for other complex, temporally diffuse processes.