Gamma: periodic and aperiodic gamma in meditation EEG

Is the meditation gamma effect an oscillation or a broadband shift?

This repository contains a re-analysis of OpenNeuro ds003969 (Braboszcz, Cahn, Levy, Fernandez & Delorme 2017, PLoS ONE), a 64-channel EEG dataset in which practitioners of three meditation traditions (Himalayan Yoga, Isha Shoonya, Vipassana) and controls alternated eyes-closed meditation and mind-wandering ("thinking") blocks. Spectral parameterisation (specparam, formerly FOOOF) is used to separate the periodic (oscillatory) and aperiodic (1/f background) components of the reported 60–110 Hz gamma increase, and to test in a single dataset whether the state (Jerbi lab, MEG) and trait (Ray lab, EEG) aperiodic findings conflict.

Presented as a poster at CuttingGardens 2026, IIT Mandi (Ananya Pandey).

• Interactive walkthrough: web/ is a React site explaining the method, the simulation and the results, deployable to Vercel (see Web walkthrough).
• Results: results/REPORT.md, generated from the committed tables in results/tables/.

Poster

Click the image for the full-resolution A0 PDF.

Summary of findings

Sample: 94 participants analysed (63 meditators, 31 controls; four participants in the undocumented tm group of the 98-subject dataset were excluded), 340 usable blocks in the parieto-occipital ROI.

• Fits were good (specparam mean R² 0.97) and the specparam exponent agreed with IRASA (r = 0.96).
• A countable 60–110 Hz peak was present in 21 of 340 blocks (6 %); a further 20 fitted peaks were pinned at a width limit and not counted.
• The state × trait interaction on the aperiodic exponent was β = −0.109 (95 % CI −0.195 to −0.023), p = .013 uncorrected; p = .027 after Benjamini-Hochberg FDR and p = .052 after Holm correction across the four fitted primary measures. It therefore holds under FDR control only. Controls' exponent was steeper during meditation than thinking; meditators' barely changed.
• Periodic gamma power was equivalent between groups within ±0.5 SD (TOST p = .041).
• Adjusting for scalp-muscle (EMG) proxies did not weaken the trait effect on the exponent.
• Years of practice (meditators, exploratory, adjusted for tradition): the meditation-vs-thinking effect does not depend on experience (years × state on the exponent p = .59; no measure significant). The resting level of gamma band power does: naive gamma power rises by 0.087 log10 units per doubling of years (p = .002; FDR .005, Holm .009), and the increase is entirely aperiodic; periodic gamma power is unrelated to years (partial ρ = .10, p = .44). Practitioners with more than 10 years differed from controls (exponent −0.23, FDR p = .022; Holm .050), whereas those with 10 years or fewer did not (FDR p = .23). Results are in results/tables/stats_practice.csv.

The ground-truth simulation shows that the method attributes an effect to the correct component (a pure slope change moved naive band power by d = +3.04 and periodic power by 0.00). It does not establish sensitivity to small peaks: simulated effects are far larger than those in the real data. The aperiodic exponent is treated as a candidate marker of cortical state, not a direct readout of excitation/inhibition balance. These are preliminary results from a single dataset.

Pipeline

00_simulate      ground-truth validation (no data needed)
01_download      fetch task BDFs from OpenNeuro (~600 MB / subject)
02_preprocess    filter, bad channels, average ref, ICA + ICLabel, epochs, EMG proxies
03_spectra       Welch PSD per preprocessing variant, IRASA on the primary ROI
04_parameterize  specparam fits per ROI -> long table; optional multiverse and per-channel fits
05_stats         mixed 2x2 (meditator x condition) with FDR/Holm correction, TOST,
                 dose-response and years-of-practice effects, EMG-confound, specification curve
06_figures       spectra, interaction plots, specification curves
07_stream        download -> process -> delete raw, per subject; replaces 01-03 with ~1 GB peak disk
08_report        write results/REPORT.md from whatever tables exist

All analysis decisions live in configs/default.yaml; the specification-curve grid is in configs/multiverse.yaml.

Multiple comparisons

medgamma.stats.adjust_pvalues adds Benjamini-Hochberg (p_fdr_bh) and Holm (p_holm) adjusted p-values to the mixed-model table. Each term (trait, state, interaction) is corrected across the primary measures that were fitted. IRASA measures are cross-checks of the specparam estimates and are not part of the correction family. The report shows uncorrected and both adjusted p-values side by side.

Setup

Requires Python ≥ 3.10.

python -m venv .venv
.\.venv\Scripts\python -m pip install -e ".[dev]"
.\.venv\Scripts\python -m pytest

Running the analysis

Full pipeline (resumable; add --overwrite for a clean rerun of all subjects):

.\.venv\Scripts\python scripts\run_all.py

Stage by stage, using the streaming stage, which replaces 01–03 and needs about 1 GB of disk at any time (each subject is downloaded, processed to ~4.5 MB of spectra, and its raw files deleted):

.\.venv\Scripts\python scripts\00_simulate.py                       # validation, no data needed
.\.venv\Scripts\python scripts\07_stream.py --n-first 20 --n-jobs 4  # ~6 min per subject
.\.venv\Scripts\python scripts\04_parameterize.py --multiverse --channels
.\.venv\Scripts\python scripts\05_stats.py
.\.venv\Scripts\python scripts\06_figures.py
.\.venv\Scripts\python scripts\08_report.py

Omit --n-first to process all subjects (~10 h). Stages 01–03 remain available when you want to keep the raw files and pre-ICA epochs on disk (01_download.py, 02_preprocess.py --n-jobs 8, 03_spectra.py --n-jobs 8).

Every script accepts --config, --subjects, --n-first N and --overwrite.

Google Colab: notebooks/colab_run.ipynb checkpoints outputs to Google Drive so the run survives session limits; budget ~15 min per subject on the free tier.

Data and storage

ds003969 is released under CC0 and is not committed here; 01_download.py fetches it to dataset.bids_root in configs/default.yaml. Raw data and derivatives go under the directory in the GAMMA_DATA environment variable (default ~/GammaData):

setx GAMMA_DATA D:\GammaData        # Windows, persistent (open a new terminal afterwards)

export GAMMA_DATA=/data/gamma        # macOS / Linux

Keep this directory outside any cloud-synced folder (OneDrive, Dropbox): the raw data are ~58 GB and the derivatives ~30 GB, and sync clients that lock and re-upload each file as it is written slow the pipeline by about two orders of magnitude.

Web walkthrough

web/ is a Vite + React site with interactive spectra (the periodic/aperiodic model and the simulated scenarios). To run it locally:

cd web
npm install
npm run dev

Deploying to Vercel: import the repository and set Root Directory to web. Vercel then uses web/vercel.json: Vite, npm run build, output dist, Node 22. No environment variables are needed.

Repository layout

medgamma/   package: config, data, preprocess, spectra, parameterize, simulate, stats, plots
scripts/    numbered pipeline stages
configs/    YAML analysis decisions
tests/      unit tests (run without data)
results/    tables/ (small CSVs, committed), REPORT.md; derivatives/ is ignored
figures/    generated figures (figures/poster/ at print resolution)
docs/       poster (PDF and JPG)
web/        interactive walkthrough (React, Vite)

Citation and licence

Code is released under the MIT License. If you use this pipeline, please cite the dataset (Braboszcz et al. 2017; OpenNeuro ds003969) and specparam (Donoghue et al. 2020, Nature Neuroscience).