This repository provides scripts to reproduce the main analyses and figures in the paper "Normative Modeling of MEG Brain Oscillations Across the Human Lifespan".

• Mohammad Zamanzadeh¹
• Ymke Verduyn¹
• Augustijn de Boer²
• Tomas Ros³
• Thomas Wolfers⁴
• Richard Dinga¹
• Marie Šafář Postma¹
• Andre F. Marquand²⁵
• Marijn van Wingerden¹
• Seyed Mostafa Kia¹²⁶

¹ Department of Cognitive Science and Artificial Intelligence, Tilburg University, Tilburg, the Netherlands
² Donders Institute for Cognition, Brain and Behavior, Radboud University, Nijmegen, the Netherlands
³ CIBM Center for Biomedical Imaging, University of Geneva, Geneva, Switzerland
⁴ German Center for Mental Health, University of Tübingen, Tübingen, Germany
⁵ Department of Cognitive Neuroscience, Radboud University Medical Center, Nijmegen, the Netherlands
⁶ Department of Psychiatry, UMC Utrecht Brain Center, University Medical Center, Utrecht, the Netherlands

Normative modeling has recently been put forward to characterize heterogeneity within cohorts with neuropsychiatric disorders and enable individual-level analysis. Despite their popularity, normative models of magnetoencephalography (MEG)-based functional neuroimaging-derived phenotypes (f-IDPs) are still lacking. In this project, we utilized the MEGaNorm and PCNToolkit to derive normative models for f-IDPs of resting-state MEG (rs-MEG) recordings for the first time. The models were trained on a large (1,846 participants), lifespan-spanning (ages 6–88), multi-site (6 scanner sites) dataset using hierarchical Bayesian regression (HBR) with a Sinh-Arcsinh (SHASH) likelihood. We incorporated age as a covariate, and sex and acquisition site as grouping effects. These normative ranges can support treatment outcome evaluation, subtyping, and diagnosis. The normative models retain individual variation and provide individual-level participant profiles.

• We first preprocessed the recordings and isolated periodic activity in the power spectrum. Features (relative theta, alpha, beta, and gamma power) were extracted and averaged across sensors.
  

• We compared two HBR models: Non-linear, heteroscedastic, and non-Gaussian, and Linear, homoscedastic, and Gaussian, using a range of model diagnostics. Our results showed better performance for the non-linear, heteroscedastic, and non-Gaussian model.

• To demonstrate clinical application, we used deviation scores to distinguish Parkinson’s disease (PD) patients from healthy participants. We reported AUC scores across f-IDPs by running the models 10 times on different train-test splits. Extreme deviation statistics showed higher positive deviations (Z > 2) in theta and negative deviations (Z < –2) in beta in PD patients compared to controls.
  

• Furthermore, we introduced a dimensional approach to characterize heterogeneity across PD patients, treating abnormalities as a spectrum rather than categorical distinctions.

• To support individual-level interpretation, we developed Individual-level Neuro-Oscillo Charts (I-NOCs)—visual tools that quantify and display individual deviations, enabling personalized assessment of functional brain dynamics.
  

The analysis is based on 1,846 resting-state MEG recordings from clinically undiagnosed participants, pooled from six distinct datasets:

1. Cambridge Centre for Ageing and Neuroscience (Cam-CAN)
   Taylor et al., 2017

2. Boys Town National Research Hospital (BTH)
   Rempe et al., 2023

3. The Open MEG Archive (OMEGA)
   Niso et al., 2016

4. The Human Connectome Project (HCP)
   Van Essen et al., 2012

5. National Institutes of Mental Health (NIMH)
   Nugent et al., 2022

6. Mother Of Unification Studies (MOUS)
   Schoffelen et al., 2019

# 1. Create and activate the environment
conda create --channel=conda-forge --strict-channel-priority --name mne python=3.12 mne
conda activate mne

# 2. Install MEGaNorm
pip install meganorm==0.1.0

# 1. Create and activate the environment
conda create --channel=conda-forge --strict-channel-priority --name mne python=3.12 mne
conda activate mne

# 2. Clone the repository and install MEGaNorm
git clone https://github.com/ML4PNP/MEGaNorm.git
cd MEGaNorm/
git checkout tags/v0.1.0
pip install .

git clone [email protected]:ML4PNP/MEG_Norm.git

git clone https://github.com/ML4PNP/MEG_Norm.git

Use the MEG_Norm_sequential.ipynb notebook to run the pipeline sequentially. Note that this may take considerable time due to computational complexity.

Use MEG_Norm_parallel.ipynb to run steps in parallel. This notebook uses SLURM for job scheduling.

Mohammad Zamanzadeh and Seyed Mostafa Kia. ML4PNP/MGaNorm: First public release, May 2025. URL https://doi.org/10.5281/zenodo.15441320.

This project is licensed under the terms of the GNU General Public License v3.0