TL;DR
This paper introduces a novel "dual regression" method to precisely map and quantify the influence of cardiac and respiratory cycles on fMRI signals, particularly in the context of aging. They found that while the overall strength of cardiac-related signals in the brain doesn't change with age, the area of the brain affected by these signals, specifically in white matter, significantly expands in older adults.
What Problem Does This Paper Address?
Functional Magnetic Resonance Imaging (fMRI) is a powerful tool for studying brain activity. However, the fMRI signal (specifically the BOLD signal) is not purely a reflection of neural activity. It's also influenced by physiological processes like breathing and the heartbeat, which cause fluctuations that can mask or mimic neural signals. These physiological "noise" sources are particularly problematic when trying to study subtle changes, such as those associated with aging. Traditional methods for removing this noise often rely on external physiological recordings, which can be imperfect, or data-driven approaches that might not be specific enough. This paper aims to develop a more robust and refined method to disentangle these physiological signals from the neural signal, and then use this method to investigate how these physiological influences change with healthy aging.
Methods
The core of this paper is a novel application of dual regression to model the power spectra of fMRI data. Here's a breakdown:
- Power Spectra: Instead of looking at the raw time series of fMRI data, the researchers analyze its frequency content. They hypothesize that physiological signals (cardiac and respiratory) will manifest as distinct peaks or patterns in the power spectrum of the fMRI signal, especially at higher frequencies (above ~0.2 Hz) where neural signals are less dominant.
- General Linear Model (GLM): They use a GLM to model the fMRI power spectrum in each voxel as a combination of a baseline thermal noise, a cardiac component, and a respiratory component.
- Iterative Dual Regression: This is the key innovation.
- Initial GLM: They first run a GLM using external recordings of cardiac and respiratory signals as initial "explanatory variables" (regressors) to estimate the spatial maps of cardiac and respiratory influence.
- Refined Spectra: They then use these estimated spatial maps to derive brain-specific cardiac and respiratory power spectra.
- Iteration: Steps 1 and 2 are repeated, using the newly derived brain-specific spectra as the regressors in the GLM. This iterative process refines the estimates of both the brain-specific spectra and their spatial maps, effectively "learning" the physiological signals directly from the fMRI data.
- Data-Driven Variant: They also developed a version of this dual regression that doesn't require external physiological recordings. Instead, it uses initial estimates of cardiac and respiratory spectra derived directly from the fMRI data itself, making it applicable to existing datasets.
Datasets
The study used resting-state fMRI data from 21 healthy subjects ranging in age from 19 to 77 years. This dataset was acquired using a high-temporal resolution EPI sequence (TR = 328 ms) on a 7T scanner. External physiological recordings (respiratory pad and pulse-oximeter) were also collected at 1 kHz and sub-sampled to the TR. No new benchmarks were introduced; the focus was on developing and applying a novel analysis method to existing data types.
Key Insights and Ideas
- Brain-Specific Physiological Spectra: The study highlights that the physiological signals (cardiac and respiratory) measured within the brain are not identical to those measured externally (e.g., pulse-oximeter). The brain's cardiac spectrum, in particular, shows higher power at the dominant cardiac frequency than external measurements. This suggests that the dual regression approach is crucial for accurately capturing these signals as they manifest in brain tissue.
- Dual Regression's Power: The iterative dual regression method proves effective in disentangling and mapping these physiological influences, even outperforming simpler GLM approaches with external regressors.
- Data-Driven Approach Viability: The data-driven variant of dual regression shows comparable results to the informed approach, suggesting it's a valuable tool for analyzing datasets where external physiological recordings are unavailable.
- Aging and White Matter: The most striking finding is the age-related expansion of cardiac-modulated regions in white matter. This suggests that while the brain's fundamental response to the cardiac cycle might not change in strength, the reach of this influence into white matter increases with age.
Key Results
- Cardiac Spectra: The informed dual-regressed cardiac spectrum in white matter showed significantly higher power at the dominant cardiac frequency compared to the external pulse-ox spectrum. Grey matter also showed higher power than white matter.
- Respiratory Spectra: Mismatches between external and dual-regressed respiratory spectra were observed, but these were highly variable between subjects and showed no systematic differences.
- Spatial Overlap: The spatial maps of cardiac and respiratory influences derived from the informed and data-driven dual regression methods showed very high overlap (99.6% for cardiac in white matter, 99.7% in grey matter).
- Age Effect in White Matter: A significant increase in the spatial extent of cardiac cycle-induced fluctuations (Re) was found in white matter with age (p < 0.002 for both informed and data-driven methods).
- No Age Effect on Cardiac Strength: The strength of cardiac-related EPI power (βc) in white matter did not show a significant age effect when controlling for factors like cardiac rate and motion.
- Comparison to Previous Metric: The study's findings align with previous work using a fixed spectral window for cardiac pulsatility, which also showed an age-related increase in white matter. However, the authors argue their dual regression method is more robust as it's independent of spectral shape and focuses on significantly modulated voxels.
Limitations and Future Work
- Small Sample Size: The study acknowledges that the number of subjects is small, particularly when splitting into "younger" and "older" groups for visualization. This limits the generalizability of the findings.
- Age-Related Atrophy: The authors note that age-related atrophy needs to be accounted for in statistical analysis, which they attempted to do using partial volume estimates.
- Future Work: The authors suggest that the data-driven dual regression approach is promising for analyzing existing fast fMRI datasets where external physiological recordings are absent. They also highlight the need for further research with larger cohorts to confirm their findings on aging and physiological influences. The paper also hints at the potential of this method for studying pathological aging conditions like Alzheimer's disease, building on prior work.
Summary updated