Angle Basis: a Generative Model and Decomposition for Functional Connectivity.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Functional connectivity (FC) is a key input for fMRI predictive models, but lacks robust theoretical generation models.
• Current FC analysis methods may not fully capture the underlying neural dynamics.
• Existing decomposition techniques often require large population datasets.

Approach:

• We introduce a novel decomposition of FC into basis sine wave states plus a jitter component.
• This method decomposes FC for individual subjects without needing population data, unlike PCA or Factor Analysis.
• The decomposition is validated by its predictive performance and ability to identify subjects.

Key Points:

• The decomposition captures predictive power comparable to original FC using 5-10 basis states.
• The decomposed components and residual show similar predictive value, outperforming FC alone by up to 5% AUC in ensemble models.
• The residual component achieves high subject identifiability (97.3%) for fingerprinting, significantly exceeding FC's performance (62.5%).

Conclusions:

• This decomposition offers a new perspective on group differences in patient populations by revealing two equally predictive FC components.
• The method facilitates the generation of synthetic fMRI data, potentially reducing acquisition costs and aiding research.
• This approach advances the understanding and application of functional connectivity in neuroimaging analysis.