🧠 neuroscience

Whole-brain normative modeling across multiple functional connectivity metrics in MEG

This paper introduces M/EEG-FUNCOIN, a whole-brain normative modeling framework that integrates multiple functional connectivity metrics from MEG data to characterize age- and sex-related neural patterns and identify meaningful individual deviations linked to cognitive ability.

Original authors: Kobbersmed, J. R. L., Gohil, C., Woolrich, M. W., Vidaurre, D.

Published 2026-10-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kobbersmed, J. R. L., Gohil, C., Woolrich, M. W., Vidaurre, D.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The human brain is not a static machine but a dynamic network, with billions of cells constantly communicating to produce thought, memory, and movement. For decades, scientists have tried to map these connections to understand what happens when the system goes wrong, such as in Alzheimer's disease or schizophrenia. A major challenge in this field is distinguishing between normal changes that happen as we age and signs of actual illness. To solve this, researchers have developed a method called "normative modeling." Imagine a growth chart for children: doctors use it to see if a child is growing at a typical rate for their age and height. Normative modeling does the same thing for the brain, using data from thousands of healthy people to create a map of what "normal" brain activity looks like at every stage of life. When a new patient is scanned, their results are compared against this map. If their brain activity deviates significantly from the expected pattern, it flags a potential problem that might otherwise be missed.

While this approach has been successful with brain imaging techniques that measure blood flow, such as functional MRI, it has been much harder to apply to electrical brain activity measured by magnetoencephalography, or MEG. MEG sensors sit outside the head and detect the tiny magnetic fields produced by neurons firing, offering a view of brain activity that is far more precise in time than blood-flow scans. However, the electrical signals are incredibly complex, shifting rapidly across many frequencies, which makes it difficult to build a single, reliable standard for what is normal. Previous attempts to create these standards often looked at individual connections between brain regions one by one or averaged them all together, but these methods frequently failed to capture the subtle, large-scale changes that occur as the brain matures and ages.

In a new study, researchers have introduced a powerful new tool called M/EEG-FUNCOIN to solve this problem. They turned to a massive, publicly available dataset from the Cambridge Centre for Ageing and Neuroscience, which includes resting-state MEG recordings from 612 healthy volunteers ranging in age from 18 to 88 years. The team did not just look at how strong the connections between brain areas were; they analyzed three different ways these areas communicate. They measured how the amplitude of signals in specific frequency bands moved together, how the timing of the signals stayed in sync, and a more complex method that captures both timing and strength without needing to sort the data into specific frequency bands first. By feeding all this information into their new model, they were able to identify broad, whole-brain patterns of connectivity that change predictably with age and sex.

The results showed that looking at the brain as a whole network reveals changes that simpler methods miss. When the researchers tried to model the brain by looking at individual connections or by averaging the strength of all connections, they found very little relationship to age or sex. In contrast, the new whole-brain model successfully identified specific dimensions of brain activity that shift steadily as people get older. For instance, they found that certain patterns of synchronization between brain regions change significantly with age, a detail that was invisible when the data was averaged out. This suggests that the aging brain does not just get "weaker" or "stronger" in a uniform way; rather, the way different parts of the brain coordinate with each other undergoes a specific, structured transformation.

To test whether these new models could actually tell something meaningful about how the brain works, the researchers looked at the cognitive performance of the volunteers. They separated the participants into groups based on their scores in a wide battery of tests covering memory, language, and executive function. They then used the new brain models to see if they could distinguish between those with the highest cognitive abilities and those with the lowest. The models built from the whole-brain patterns were successful at this task, identifying specific deviations in brain connectivity that correlated with cognitive performance. This indicates that the framework can capture individual differences in brain health that go beyond simple averages, offering a way to spot subtle variations that might be linked to how well a person thinks and remembers.

The study also addressed the technical hurdles of working with such complex data. Because the electrical signals from the brain are so intertwined, standard mathematical tools often struggle to process them without errors. The researchers adapted their software to handle these high-dimensional signals efficiently, ensuring that the patterns they found were real and not just mathematical artifacts. They confirmed that the model was stable and consistent, producing reliable results even when tested on different subsets of the data. This robustness is crucial for any tool intended for clinical use, where accuracy is paramount.

By demonstrating that these whole-brain patterns are both measurable and meaningful, the study opens a new door for understanding brain health. The ability to map the normal trajectory of brain connectivity across the lifespan provides a much sharper reference point for identifying disease. If a patient's brain activity falls outside the expected range for their age and sex, doctors may be able to detect early signs of pathology long before symptoms become obvious. The researchers emphasize that while this is a significant step forward, the framework is flexible enough to be refined further, potentially incorporating other variables or focusing on specific types of brain disorders. For now, the work establishes that the brain's electrical symphony has a predictable score that changes with time, and we have finally found a way to read the sheet music.

Technical Summary: Whole-brain normative modeling across multiple functional connectivity metrics in MEG

Problem Statement
Normative modeling has emerged as a powerful approach in neuroimaging to characterize individual deviations from healthy population norms, typically using structural and functional MRI (s/fMRI). However, applying this framework to magneto- and electroencephalography (MEG/EEG) remains challenging. While M/EEG offers superior temporal resolution and direct measurement of brain activity, the complexity of electrophysiological signals—characterized by nested oscillations across broad frequency ranges—necessitates a diverse set of functional connectivity (FC) metrics beyond standard covariance. Previous normative models in electrophysiology have largely focused on non-FC measures (e.g., power spectral density) or have employed edge-wise approaches (modeling individual connections separately) or averaged FC metrics. These existing methods have yielded limited associations with demographic covariates like age and sex, potentially failing to capture systemic, network-level effects that drive higher-order functions and disease phenotypes.

Methodology
The authors propose M/EEG-FUNCOIN, a whole-brain normative modeling framework adapted from their previous fMRI work (FUNCOIN) to handle the specific mathematical and computational challenges of electrophysiology.

Data: The study utilized resting-state MEG data from 612 healthy subjects (aged 18–88) from the Cam-CAN dataset. Data were preprocessed using source reconstruction (LCMV beamformer) and parcellated into 52 anatomical regions of interest (ROIs).
Functional Connectivity Metrics: Three distinct FC measures were evaluated to capture different aspects of brain coupling:
1. Amplitude Envelope Correlation (AEC): Correlation of signal amplitudes within predefined frequency bands.
2. Phase-Locking Value (PLV): Consistency of phase differences between regions within predefined frequency bands.
3. Time-Delay Embedding (TDE) Covariance: A data-driven approach capturing both power and phase relationships across frequencies without predefined band filtering, utilizing time-shifted signal copies.
Modeling Framework: M/EEG-FUNCOIN performs simultaneous dimensionality reduction and regression. It identifies shared components ( $\gamma$ $γ$ ) across subjects whose strengths are maximally associated with covariates (age and sex).
- To address rank-deficiency common in electrophysiological FC matrices (e.g., due to high correlation in TDE signals), the authors implemented the Moore-Penrose pseudoinverse and singular value decomposition (SVD) to handle non-full-rank matrices.
- The model optimizes a heteroscedastic, log-linear likelihood function to find components where the variance is explained by covariates.
- Bias correction was applied by refitting coefficients while keeping identified components fixed.
Validation and Application:
- Consistency: Components were validated via 10-fold cross-validation (10 repetitions) to ensure stability across folds.
- Comparison: The framework was compared against edge-wise models and models based on averaged FC strength.
- Cognitive Classification: To test the utility of the normative deviations, the authors trained logistic regression classifiers to distinguish between "low" and "high" cognitive performers (top and bottom 15% of a composite cognitive score) using Z-scores derived from the normative models. The models were trained on the middle 70% of subjects to ensure out-of-sample testing.

Key Results

Identification of Network-Level Effects: M/EEG-FUNCOIN successfully identified multiple age- and sex-dependent components across all three FC measures (AEC, PLV, TDE). The model revealed strong correlations between FC component strengths and age ( $r \geq 0.37$ ), with TDE showing the highest single-component $R^2$ values.
Superiority Over Existing Methods:
- Edge-wise models: Individual connection models showed very low $R^2$ values, indicating that modeling connections separately fails to capture significant demographic associations.
- Averaged FC models: Averaging FC values across the whole brain (as suggested in previous EEG studies) resulted in weak or non-existent associations with age and sex, except for delta-band PLV.
- M/EEG-FUNCOIN: In contrast, the whole-brain approach captured network-level aging effects that were invisible to localized or averaged models, particularly in AEC and PLV across multiple frequency bands.
Predictive Power for Cognitive Ability: The Z-scores derived from M/EEG-FUNCOIN components were significant predictors of cognitive performance.
- Components from delta and alpha band PLV, as well as TDE, significantly contributed to classifying high vs. low cognitive performers.
- Notably, the TDE component was a consistent predictor across all model settings.
- Some M/EEG-FUNCOIN components (e.g., beta-band AEC) became significant predictors only when combined with other features, suggesting they capture unique variance not explained by average FC or other metrics.
Statistical Validity and Scalability: The residuals of the identified components were approximately normally distributed with homogeneous variance, validating the Z-score calculations. The framework demonstrated scalability in handling large datasets and high-dimensional matrices (e.g., from TDE).

Significance and Claims
The paper claims that M/EEG-FUNCOIN provides a robust, scalable, and statistically valid framework for whole-brain normative modeling in electrophysiology. Its primary significance lies in:

Overcoming Limitations of Previous Approaches: It challenges the notion that the resting-state electrophysiological connectome is preserved in healthy aging (a conclusion drawn from edge-wise or averaged models) by demonstrating that network-level FC structures do change with age and sex.
Capturing Systemic Effects: By moving away from edge-wise modeling, the framework captures systemic, network-level deviations that are crucial for characterizing complex brain pathology and individual differences.
Versatility: The framework successfully integrates diverse FC metrics (AEC, PLV, TDE), offering researchers flexibility to choose metrics based on specific research questions and signal-to-noise ratios.
Clinical Potential: The ability to identify meaningful deviations in individuals with extreme cognitive performance suggests the framework has potential for applications in characterizing and subtyping heterogeneous brain disorders, although the current study was limited to a healthy cohort.

The authors conclude that M/EEG-FUNCOIN establishes normative ranges for FC across measures and frequency bands, offering a promising tool for future studies on complex, network-level brain pathology.

Technical Summary: Whole-brain normative modeling across multiple functional connectivity metrics in MEG

More like this