Manifold Learning for fMRI time-varying FC

Area of Science:

• Neuroscience
• Data Science
• Machine Learning

Background:

• Whole-brain functional connectivity (FC) measured with fMRI evolves over time (tvFC), but its high dimensionality (thousands of features) limits exploration.
• Dimensionality reduction techniques, particularly manifold learning (ML), are sought to create lower-dimensional representations (e.g., 2D/3D plots) of tvFC data.
• Understanding the intrinsic dimensionality and optimal ML techniques for tvFC is crucial for extracting meaningful neuro-biological information.

Approach:

• Estimated the intrinsic dimension (ID) of tvFC data manifolds, finding it ranges from 4 to 26, varying between rest and task states.
• Evaluated three state-of-the-art ML techniques—Laplacian Eigenmaps (LE), T-distributed Stochastic Neighbor Embedding (T-SNE), and Uniform Manifold Approximation and Projection (UMAP)—for tvFC data representation.
• Assessed the ability of MLTs to capture subject identity and task performance, and their robustness to hyperparameter selection and feature normalization.

Key Points:

• tvFC data exhibits an intrinsic dimension between 4 and 26, which differs significantly between resting and task states.
• UMAP and T-SNE effectively capture concurrent information on subject identity and task, whereas LE captures only one aspect at a time.
• MLT performance varies considerably based on the chosen algorithm and hyperparameter settings; feature normalization is critical for cross-subject analysis.

Conclusions:

• Manifold learning techniques show promise for generating meaningful, lower-dimensional representations of labeled tvFC data.
• UMAP and T-SNE are more effective than LE for capturing complex relationships within tvFC data.
• Applying MLTs to unlabeled tvFC data, such as resting-state scans, remains a significant challenge requiring further methodological development.