TL;DR
This study investigates whether the standard Pearson's correlation coefficient is sufficient for measuring functional connectivity (FC) in fMRI data, or if a newer method, Multiscale Graph Correlation (MGC), offers crucial insights into nonlinear relationships. While Pearson's correlation is generally reliable for within-network connections, MGC reveals localized, potentially nonlinear dependencies between networks and in altered brain states, suggesting it can provide a more nuanced view of brain interactions.
What Problem Does This Paper Address?
Functional connectivity (FC) in fMRI is typically measured using Pearson's correlation coefficient, which quantifies linear relationships between brain region activity over time. However, brain activity is complex and may involve nonlinear interactions that Pearson's correlation misses. This limitation could lead to an incomplete understanding of how brain regions communicate, especially in dynamic or altered states. The paper addresses this by comparing Pearson's correlation with Multiscale Graph Correlation (MGC), a method designed to capture both linear and nonlinear dependencies across different "scales" of temporal data, and to identify an "optimal scale" where these dependencies are strongest.
Methods
The researchers employed two primary methods for measuring functional connectivity:
- Pearson's Correlation Coefficient (Pearson's r): This is the standard, widely used method that measures the linear relationship between the time series of two brain regions.
- Multiscale Graph Correlation (MGC): This newer method assesses statistical dependencies across multiple temporal scales. It provides two key outputs:
- MGC statistic: A measure of the strength of dependence, similar to correlation but capable of capturing nonlinearities.
- Optimal scale: This indicates the specific temporal window (or "scale") within the fMRI data where the dependency between two regions is strongest. A "global" scale suggests the relationship is consistent across the entire scan, while a "local" scale suggests the relationship is prominent only within a specific segment of the data, hinting at nonlinear dynamics.
The study also utilized Dynamic Time Warping (DTW) as an additional nonlinear measure for comparison. Computational cost was also evaluated.
Datasets
The study utilized data from three sources:
- Human Connectome Project (HCP) test-retest dataset: Used for comparing Pearson's r and MGC, assessing reliability, and examining the impact of data quantity.
- Nathan Kline Institute-Rockland Sample (NKI-RS): Used for examining the impact of different repetition times (TRs) on FC measures and for brain-behavior association studies.
- Macaque dataset: Used to investigate FC under different anesthesia states (awake vs. varying levels of isoflurane sedation), providing a distinct experimental condition.
Key Insights and Ideas
- Within-network FC is largely linear: For connections within established brain networks, Pearson's r and MGC statistics showed high similarity, and the optimal scale was often "global," suggesting these relationships are stable and predominantly linear. This implies Pearson's r is often sufficient for these common analyses.
- Between-network FC may be nonlinear: Connections between different brain networks showed more localized optimal scales with MGC, even when Pearson's r indicated weak or negative linear correlations. This suggests that MGC can reveal complex, nonlinear interactions that Pearson's r overlooks.
- Anesthesia reveals nonlinear dynamics: In macaque studies, deeper anesthesia led to a shift from global to local optimal scales in MGC, indicating that altered brain states can induce more localized, time-varying, and potentially nonlinear dependencies. This highlights MGC's utility in studying dynamic brain states.
- Data quantity matters: Both Pearson's r and MGC showed improved reliability and stronger observed connections with increased amounts of fMRI data. This underscores the importance of sufficient data for robust FC analysis.
- Pearson's r is more reliable with limited data: While MGC offers deeper insights into nonlinearities, Pearson's r generally demonstrated higher test-retest reliability, especially when data was limited. MGC's reliability improved with more data.
Key Results
- High spatial similarity: Pearson's r and MGC statistics captured highly similar spatial connectivity patterns across datasets, with correlations often exceeding 0.83 at the individual level and 0.938 at the group level.
- Reliability: Pearson's r generally showed higher discriminability (a measure of reliability) than MGC statistics and optimal scale, particularly with limited data. However, both methods improved with more data. For example, with 800 time points, whole-brain discriminability reached ~0.99 for Pearson's r and ~0.98 for MGC statistics.
- Optimal scale insights: Local optimal scales were frequently observed for between-network connections and in anesthetized macaque states, suggesting nonlinear dependencies. Global optimal scales were common for within-network connections.
- Brain-behavior associations: Both Pearson's r and MGC were equally effective in capturing age-related changes in functional connectivity, with MGC offering no significant advantage in this specific domain.
Limitations and Future Work
- Computational Cost: MGC requires significantly more computational resources than Pearson's correlation.
- Reliability with Limited Data: MGC's reliability, particularly its optimal scale measure, can be lower than Pearson's r when data is scarce, potentially due to sensitivity to noise.
- Brain-Behavior Association: MGC did not outperform Pearson's r in detecting brain-behavior associations, suggesting that dominant relationships in this context may be linear.
The authors suggest that researchers should carefully weigh the trade-offs between MGC's added complexity and computational cost against the potential for deeper insights into dynamic or nonlinear brain processes. Future work could explore MGC's utility in other experimental conditions and clinical populations where nonlinear dynamics are suspected.
Summary updated