A bayesian model-selection approach for determining the number of spectral peaks in neural power spectra

This paper introduces a data-driven Bayesian model-selection approach using the Bayesian Information Criterion to automatically determine the optimal number of spectral peaks in neural power spectra, thereby reducing reliance on manual parameter settings and enhancing the robustness, replicability, and interpretability of neural signal analysis.

The Gist

The Problem: Tuning a Radio with Too Many Knobs

Imagine you are trying to listen to a specific radio station (a rhythmic brain signal) while driving through a city full of static and background noise (the chaotic, non-rhythmic brain activity).

For years, scientists have had a tool called specparam that helps them separate the "music" (rhythmic brain waves) from the "static" (background noise). However, this tool has a major flaw: it requires the user to manually turn a "knob" to decide how many radio stations they think are playing.

• If you turn the knob too low: You might miss a real station.
• If you turn the knob too high: The tool gets greedy. It starts hearing "stations" that aren't there, mistaking static for music. This is called overfitting.

Because every researcher turns this knob differently, two scientists looking at the same brain data could get completely different results. This makes science hard to repeat and trust.

The Solution: A Smart, Self-Adjusting Radio

The authors of this paper, led by Luc E. Wilson and Jason da Silva Castanheira, created a new version of the tool called ms-specparam.

Instead of asking the human to guess how many stations are playing, this new tool acts like a smart, self-adjusting radio. It uses a mathematical rule called the Bayesian Information Criterion (BIC) to ask itself: "What is the simplest explanation that still fits the data?"

Think of it like packing a suitcase:

• The Old Way: You guess you need 10 shirts, so you pack 10. If you only needed 5, you wasted space. If you needed 12, you didn't have enough.
• The New Way (ms-specparam): The suitcase weighs itself. It keeps adding shirts until the weight is just right to explain the trip, then it stops. It finds the "Goldilocks" number of peaks—no more, no less.

How They Tested It

The team tested their new tool in two ways:

1. The "Fake" Test (Synthetic Data): They created 5,000 computer-generated brain signals where they knew exactly how many "stations" (peaks) were hidden inside.

   • Result: The old tool often heard extra fake stations (false alarms). The new tool was much better at ignoring the static and only counting the real stations. It was more accurate and less likely to make mistakes.

2. The "Real" Test (Human Brains): They analyzed brain scans from 606 real people of different ages.

   • Result: The new tool found fewer, but more reliable, brain waves. It also showed that the "static" (background noise) in the brain changes as we age, but the amount of change depends heavily on which tool you use. The new tool suggests the aging effect is real but perhaps slightly different than previously thought.

Why This Matters

1. Less Guesswork: Scientists no longer need to be experts at guessing the right settings. The data tells the tool what to do.
2. Better Science: Because everyone uses the same "smart" logic, different labs can compare their results more easily. It makes research more reproducible.
3. Clearer Picture: By stopping the tool from inventing fake brain waves, we get a cleaner, more accurate map of how our brains actually work.

The Bottom Line

This paper introduces a "smart filter" for brain signals. Instead of letting human bias decide how many brain rhythms exist, the new method lets the data decide for itself. It's like upgrading from a manual radio with a sticky knob to a digital tuner that automatically finds the clearest signal, ensuring that what we hear is real music, not just static.

Technical Summary

1. Problem Statement

Neural power spectra consist of two primary components: aperiodic (arrhythmic) activity, which follows a $1/f^\alpha$ power law, and periodic (rhythmic) activity, modeled as Gaussian-shaped peaks. Current spectral parameterization tools (e.g., specparam, SPRiNT) require users to manually define hyperparameters, most critically the maximum number of spectral peaks ($N_G$) to fit.

• The Core Issue: Manual selection of $N_G$ is arbitrary and user-dependent.
  • If set too low, genuine oscillatory peaks are missed (underfitting).
  • If set too high, the algorithm fits noise fluctuations as spurious peaks (overfitting), particularly in noisy data.
• Consequence: This reliance on user expertise compromises the replicability, robustness, and interpretability of neurophysiological findings, as results vary based on arbitrary parameter choices rather than the data itself.

2. Methodology

The authors propose ms-specparam, a data-driven model selection procedure that automates the determination of the optimal number of spectral peaks using the Bayesian Information Criterion (BIC).

A. The Algorithm (ms-specparam)

1. Iterative Fitting: The algorithm starts by fitting an aperiodic-only model. It then iteratively adds Gaussian peaks in decreasing order of amplitude, similar to the standard specparam approach.
2. Objective Function: Instead of minimizing only the least-squares error, the method minimizes the negative log-likelihood (NLL) of the model, assuming zero-mean Gaussian noise.
3. Model Selection via BIC: For every model complexity (number of peaks $P$), the BIC is calculated:
   $$BIC = 2 \cdot NLL + \ln(N) \cdot k$$
   Where $N$ is the number of frequency bins and $k$ is the total number of parameters ($k = 3P + 2$, accounting for peak center, amplitude, bandwidth, plus aperiodic exponent and offset).
4. Optimization: The algorithm selects the model with the lowest BIC, which represents the best trade-off between goodness-of-fit and model parsimony (simplicity).
5. Bayes Factor (BF): The method computes a Bayes Factor to quantify evidence for periodic activity against an aperiodic-only null hypothesis:
   $$BF_{01} = e^{(BIC_0 - BIC_1)/2}$$
   Where $BIC_0$ is the aperiodic-only model and $BIC_1$ is the best-fitting model with peaks.

B. Validation Datasets

1. Synthetic Ground-Truth Data: 5,000 simulated power spectra with known aperiodic parameters and 0–4 ground-truth peaks, tested under low, medium, and high noise conditions.
2. Empirical Data: Resting-state Magnetoencephalography (MEG) data from 606 participants (Cam-CAN dataset, ages 18–90).
   • Preprocessing: Standard pipeline including line noise removal, artifact correction (SSP), and source mapping to 148 cortical regions (Destrieux atlas).
   • Analysis: Comparison of ms-specparam against default-specparam (fixed max peaks = 6).

3. Key Contributions

• Automation: Eliminates the need for users to pre-specify the maximum number of peaks, making the analysis fully data-driven.
• Parsimony: Systematically prevents overfitting by penalizing unnecessary model complexity via BIC.
• Quantitative Evidence: Provides a statistical framework (Bayes Factors) to objectively adjudicate the presence of rhythmic activity in specific frequency bands.
• Open Source: The algorithm is released as a free, open-source plugin for Brainstorm (MATLAB) and a Python library.

4. Key Results

A. Synthetic Data Performance

• Positive Predictive Value (PPV): ms-specparam significantly outperformed default-specparam in avoiding false positives.
  • PPV: 96% (ms-specparam) vs. 63% (default-specparam) in moderate noise.
  • Overfitting: default-specparam overestimated peak counts by 59%, whereas ms-specparam underestimated by only 13%.
• Parameter Accuracy: ms-specparam yielded significantly lower estimation errors for all parameters (aperiodic exponent, offset, peak frequency, amplitude, and bandwidth) compared to the default method.
• Robustness: Performance remained superior across varying noise levels and peak separations.

B. Empirical MEG Data (Cam-CAN)

• Model Fit: ms-specparam produced models with significantly lower residual variance (Mean Squared Error) across all cortical regions, particularly at the frequency spectrum edges (<5 Hz and >35 Hz).
• Peak Detection: On average, ms-specparam fitted 0.37 fewer peaks per spectrum than the default method, indicating a more parsimonious model. It detected fewer low-amplitude, potentially spurious peaks (especially in frontal theta ranges).
• Aperiodic Parameters: The choice of algorithm significantly influenced the estimation of age-related changes.
  • Age Interaction: A significant interaction was found between age and the algorithm choice. The magnitude of the age-related decline in the aperiodic exponent and offset was smaller when using ms-specparam compared to default-specparam. This suggests that previous findings of steep age-related flattening may have been partially inflated by overfitting noise as peaks in the default method.
• Spatial Distribution: Bayesian evidence for rhythmic activity was highest in occipital regions (e.g., cuneus) and lowest in medial frontal cortices.

5. Significance and Implications

• Reproducibility: By removing user-dependent hyperparameter tuning, ms-specparam enhances the reproducibility of spectral parameterization studies.
• Physiological Interpretation: The method suggests that previous estimates of the aperiodic exponent (linked to E/I balance) and offset (linked to neuronal firing) may have been biased by overfitting. The new method provides a more accurate, conservative estimate of these physiological markers.
• Scientific Rigor: The integration of BIC and Bayes Factors moves spectral analysis from a descriptive curve-fitting exercise to a rigorous statistical inference framework, allowing researchers to quantify the evidence for oscillatory activity rather than assuming its presence based on arbitrary thresholds.
• Practical Trade-off: The authors note that ms-specparam is approximately 8 times slower than the default method (due to iterative model testing) but remains computationally feasible for standard research workflows.

In conclusion, this paper presents a principled, automated solution to a pervasive problem in neurophysiology, ensuring that spectral decompositions reflect true neural dynamics rather than algorithmic artifacts or user bias.