Individual-specific fMRI-Subspaces improve functional connectivity prediction of behavior

Rajan Kashyap, Ru Kong, Sagarika Bhattacharjee, Jingwei Li, Juan Helen Zhou, B. T. Thomas Yeo

Abstract

Resting-state functional connectivity (RSFC) is a promising tool for predicting human behavior, but its effectiveness is often limited by shared group-level signals. This study utilizes Common and Orthogonal Basis Extraction (COBE) to decompose fMRI data from 803 Human Connectome Project (HCP) participants into common and individual-specific subspaces. By removing these shared components—which often correspond to state-specific effects like visual fixation or the Default Mode Network—the researchers achieved an 11.7% average improvement in predicting 58 behavioral measures spanning cognition, emotion, and personality.

Introduction

The quest to find a neural "fingerprint" for human behavior relies on resting-state functional connectivity (RSFC). However, a significant portion of the fMRI signal is shared across individuals, potentially masking the unique variations that correlate with specific cognitive or emotional traits. If everyone’s brain signals appear largely identical due to shared experimental conditions or general human physiology, the predictive power for individual differences is naturally diminished. This study hypothesizes that isolating and removing these "common subspaces" can sharpen the focus on individual-specific signals, thereby enhancing behavioral prediction.

Methods

The researchers analyzed the Human Connectome Project (HCP) S1200 release, focusing on 803 healthy adults with four resting-state fMRI runs. To isolate individual variance, they employed Common and Orthogonal Basis Extraction (COBE).

Schematic of the COBE decomposition (A) and the resulting time series filtering (B).

As illustrated in Figure 1, COBE decomposes each participant's fMRI data into two distinct parts: a common subspace shared across the entire group and an individual-specific subspace. This decomposition was applied to each of the four fMRI runs independently to account for potential "state-specific" changes over time. Behavioral prediction was then performed using Elastic Net regression, comparing raw data (NO-COBE) against various filtered versions where different numbers of common components were removed.

Results and Key Insights

The decomposition revealed that the common signals were not random noise but had distinct spatial localizations:

  • State-Specific Visual Effects: In the first run, the primary common component was localized to the visual cortex, likely reflecting the participants' initial focus on the fixation cross-hair. This component was notably absent in subsequent runs.
  • The DMN as a Group Constant: Across all runs, subsequent common components were highly similar and localized to the Default Mode Network (DMN), specifically the posterior cingulate cortex and precuneus. This suggests that a significant portion of DMN activity is a shared group phenomenon.

Removing these common components led to a substantial boost in predictive accuracy. The "COBE-2111" variant—removing two components from the first run and one from the others—performed best, improving behavioral prediction by an average of 11.7% across 58 measures. The mean prediction accuracy rose from r = 0.160 (unfiltered) to r = 0.179 (filtered).

Discussion

While filtering improves results, there is a "sweet spot." Removing too many components (e.g., the COBE-3222 variant) led to a decline in accuracy, suggesting that the algorithm began stripping away useful individual-specific information. Furthermore, while the common components align with known brain networks like the DMN, their exact biological nature—whether they represent shared cognitive states or purely physiological noise—remains a subject for future investigation. Nonetheless, the results demonstrate that refining the individual "fingerprint" by removing group-level commonalities is a powerful strategy for brain-behavior modeling.

Summary updated

BibTeX
@article{kashyap2019individual,
  author = {Rajan Kashyap and Ru Kong and Sagarika Bhattacharjee and Jingwei Li and Juan Helen Zhou and B. T. Thomas Yeo},
  title = {{Individual-specific fMRI-Subspaces improve functional connectivity prediction of behavior}},
  journal = {NeuroImage},
  year = {2019},
  doi = {10.1016/j.neuroimage.2019.01.069},
}
Figure