Analytical Choices Impact the Estimation of Rhythmic and Arrhythmic Components of Brain Activity

This study demonstrates through simulations and empirical validation that using specparam to model rhythmic power provides a more accurate and independent quantification of rhythmic and arrhythmic brain components compared to detrending methods, which conflate the two and produce spurious correlations.

The Gist

The Big Picture: Listening to the Brain's Radio

Imagine your brain is a busy radio station. It broadcasts two very different types of signals at the same time:

1. The Rhythmic Signal (The Music): These are the clear, catchy tunes—like a drumbeat or a melody. In the brain, these are the "oscillations" (like Alpha or Beta waves) that help us focus, remember, or see things. Scientists want to measure how loud these "songs" are.
2. The Arrhythmic Signal (The Static): This is the background hiss, the crackle, and the static that fills the airwaves. In the brain, this is the "aperiodic" activity (the 1/f noise). For a long time, scientists thought this was just useless noise and tried to ignore it. But recently, we've realized this "static" actually tells us a lot about the brain's health, age, and how excited or calm it is.

The Problem:
When you try to listen to the "music" (rhythms), the "static" (background noise) makes it hard to hear clearly. If the static gets louder, the music might sound quieter, even if the music itself hasn't changed.

For years, scientists have used different tools to try to separate the music from the static. Some tools just try to "turn down the volume" on the static (a method called detrending). Others try to mathematically build a model of the music and the static separately (a method called modeling).

The Discovery:
This paper asks a crucial question: "Which tool gives us the truest picture?"

The authors ran thousands of computer simulations (creating fake brain signals with known answers) and then tested these tools. They found that the tools used to "turn down the static" were actually lying to us.

The Analogy: The "Fake Noise" Trap

Imagine you are a judge trying to measure how much a singer is singing (the rhythm) versus how much wind is blowing through a microphone (the background noise).

• The Old Way (Detrending): You try to subtract the wind noise from the recording. But here's the catch: if you guess the wind noise wrong, your subtraction messes up the singer's volume.

  • If you think the wind is louder than it actually is, you subtract too much, and the singer sounds quieter.
  • If you think the wind is quieter, you subtract too little, and the singer sounds louder.
  • The Result: You start seeing a fake connection. You might think, "Oh, whenever the wind gets stronger, the singer gets quieter!" But that's not true! The singer didn't change; your math just got confused. The paper calls this a "spurious correlation" (a fake relationship).

• The New Way (Modeling): Instead of just subtracting, you build a mathematical model. You say, "Okay, the wind follows a specific curve, and the singer follows a specific bell shape." You fit both shapes to the data at the same time.

  • The Result: This method correctly identifies that the wind and the singer are independent. It doesn't get confused. It tells you the true volume of the singer, regardless of how loud the wind is.

What the Paper Found

1. Detrending is Dangerous: The popular method of simply "subtracting" the background noise (detrending) creates fake relationships.

   • It made it look like brain rhythms and background noise were negatively linked (when one went up, the other went down).
   • It made it look like different brain rhythms (like Alpha and Beta waves) were connected to each other, when they actually weren't.
   • It made it look like how these signals changed with age was different depending on which tool you used.

2. Modeling is the Hero: The method that uses Gaussian modeling (fitting specific shapes to the peaks, like the specparam tool mentioned in the paper) is much more accurate.

   • It successfully separated the "music" from the "static."
   • It showed that in real life, brain rhythms and background noise are actually positively linked (they tend to go up and down together), which makes more sense biologically.

Why Should You Care?

This isn't just about math; it changes how we understand the brain.

• If we use the wrong tool: We might think a disease causes the brain's "static" to rise while the "music" falls. We might think aging makes the brain's rhythms disappear. We might design treatments based on these fake connections.
• If we use the right tool: We realize that the brain's background noise and its rhythms are actually working together. This helps us understand how the brain ages, how diseases like Parkinson's or Alzheimer's affect us, and how we perceive the world.

The Takeaway

The authors are saying: "Stop trying to just subtract the background noise. It creates illusions."

Instead, we should use modeling tools that treat the brain's rhythm and its background noise as two separate, independent things that happen to exist in the same space. By doing this, we stop seeing fake patterns and start seeing the real story of how our brains work.

In short: Don't just turn down the volume on the static; learn to distinguish the song from the noise so you don't get fooled by the radio.

Technical Summary

1. Problem Statement

Brain activity consists of two distinct components: rhythmic (periodic) oscillations (e.g., alpha, beta bands) and arrhythmic (aperiodic) background activity (typically following a $1/f^\chi$ power law). While both components are increasingly linked to behavior, cognition, aging, and disease, accurately separating them remains methodologically challenging.

Current literature employs divergent methods to quantify rhythmic power:

• Model-based approaches: Using algorithms like specparam to fit Gaussian peaks to the power spectrum, where rhythmic power is defined by the amplitude (height) of the Gaussian peak above the modeled $1/f$ background.
• Detrending approaches: Subtracting the modeled $1/f$ arrhythmic component from the raw power spectrum (either in log-log or linear space) and calculating the residual power (mean or max) within a specific frequency band.

The core problem is that these analytical choices may introduce spurious correlations between rhythmic and arrhythmic parameters. If the methods used to isolate these components are flawed, findings regarding their independent contributions to behavior or disease may be artifacts of the analysis rather than true neurophysiological phenomena.

2. Methodology

The authors employed a two-pronged approach: simulation-based parameter recovery and empirical validation.

A. Simulations (Ground Truth)

• Data Generation: Over 20,000 synthetic neural time series were generated using the NeuroDSP toolbox. These series combined independent aperiodic ($1/f$) and periodic (Gaussian) components with known ground-truth parameters.
• Experimental Manipulations: The authors systematically varied specific parameters while holding others constant to test for independence:
  • Alpha center frequency (7–14 Hz).
  • Alpha amplitude (0.3–2.0 a.u.).
  • Arrhythmic exponent (0.6–1.5).
  • Alpha bandwidth (1–5 Hz).
  • Correlated relationships (simulating ground-truth correlations between alpha power and the exponent).
• Methods Compared: Three methods were applied to the same simulated data to estimate rhythmic alpha power:
  1. Modelled Power: The height of the Gaussian peak fitted by specparam.
  2. Log-Detrended Power: Residual power after subtracting the $1/f$ model in log-log space.
  3. Linear-Detrended Power: Residual power after subtracting the $1/f$ model in linear space.
• Missing Value Handling: The authors tested strategies for handling cases where specparam failed to fit a peak (e.g., excluding data vs. imputing zero amplitude).

B. Empirical Validation

• Dataset: Resting-state MEG data from 606 participants (ages 18–89) from the CamCAN repository.
• Preprocessing: Standard MEG preprocessing (filtering, artifact removal, source reconstruction using LCMV beamforming).
• Analysis: The three quantification methods were applied to cortical parcels. The authors examined the correlation between rhythmic alpha power and the arrhythmic exponent, as well as the relationship between alpha and beta power, across different age groups.

3. Key Contributions

1. Demonstration of Methodological Bias: The paper provides rigorous evidence that spectral detrending (both log and linear) introduces systematic spurious correlations between rhythmic power and the arrhythmic exponent, even when no such relationship exists in the ground truth.
2. Mechanism of Error: The authors identify that errors in estimating the arrhythmic background (offset and slope) propagate into the residual power calculation in detrending methods. Specifically, underestimation of the arrhythmic offset leads to overestimation of rhythmic power in detrended analyses.
3. Validation of Modelled Power: The study demonstrates that modelled Gaussian amplitude (from specparam) is the most robust method for independently recovering rhythmic and arrhythmic components, minimizing spurious relationships.
4. Handling Missing Data: The authors show that replacing missing peak values (where no Gaussian is fitted) with zero amplitude yields more accurate recovery of ground-truth relationships than excluding those data points.

4. Key Results

Simulation Findings

• Spurious Correlations: When ground truth was set to zero correlation ($\rho = 0$) between alpha power and the arrhythmic exponent:
  • Log-Detrended: Showed strong negative correlations ($\rho \approx -0.35$ to $-0.40$).
  • Linear-Detrended: Showed moderate negative correlations ($\rho \approx -0.17$).
  • Modelled Power: Accurately recovered the null relationship ($\rho \approx -0.03$).
• Cross-Rhythm Artifacts: Detrending methods also created spurious positive correlations between distinct rhythms (e.g., alpha and beta power), whereas modelled power correctly showed them as independent.
• Bandwidth Sensitivity: For broad spectral peaks (bandwidth > 3 Hz), all methods struggled, but modelled power remained superior for narrow peaks.
• Missing Data: Imputing missing peak amplitudes with zero improved the recovery of simulated correlations compared to listwise deletion.

Empirical Findings (CamCAN)

• Divergent Interpretations: The choice of method drastically altered the observed relationship between alpha power and the arrhythmic exponent across the cortex:
  • Log-Detrended: Showed a negative correlation, particularly in posterior regions.
  • Modelled: Showed a positive correlation across most of the cortex.
• Age Effects: The apparent relationship between alpha power and age was method-dependent. Detrending suggested age-dependent changes in the relationship, while modelled power suggested a consistent positive relationship across the lifespan.
• Alpha-Beta Correlation: Log-detrended power showed strong positive correlations between alpha and beta amplitudes (suggesting they are linked), whereas modelled power showed they were largely independent.

5. Significance and Recommendations

• Reproducibility Crisis: The findings suggest that many previous studies using detrending methods may have reported spurious relationships between brain rhythms and behavior, or between different rhythms, due to analytical artifacts.
• Theoretical Implications: The positive correlation found between alpha power and the arrhythmic exponent (using modelled power) supports theories that both components reflect cortical inhibition and Excitation/Inhibition (E-I) balance, rather than being independent or negatively coupled as previously suggested by detrending studies.
• Best Practices:
  • Researchers should prioritize modelled Gaussian amplitudes (e.g., via specparam) over detrending methods for quantifying rhythmic power.
  • When using modelled power, missing values (no peak fitted) should be handled by imputing zero rather than excluding data, provided the resulting distribution is handled appropriately (e.g., using non-parametric statistics).
  • Visual inspection of spectral fits is critical, especially for broad peaks, to ensure the arrhythmic background is accurately modeled.

Conclusion: The paper establishes that analytical choices are not neutral; they fundamentally shape the interpretation of neural data. By adopting model-based parameterization, the field can achieve more robust, reproducible, and physiologically accurate quantification of rhythmic and arrhythmic brain activity.