L-FAME-Dataset-Benchmark/L-FAME

L-FAME: Longitudinal Focused Attention Meditation EEG Dataset and Benchmark

A longitudinal 64-channel EEG dataset and benchmark for studying focused attention meditation (FAM) and how its neural signatures evolve across a six-week training period. 74 healthy adults were recorded at a pre-intervention baseline; 44 of them returned for a post-intervention follow-up. Three FAM techniques are systematically compared: Hare Krishna mantra (HK), SA-TA-NA-MA mantra (SA), and Breath Focus (BF).

The dataset accompanies the paper "L-FAME: Longitudinal Focused Attention Meditation EEG Dataset and Benchmark." It defines three benchmark tasks: (1) cognitive-state decoding (rest vs. meditation), (2) fine-grained technique classification (HK / SA / BF), and (3) cross-session adaptation across the longitudinal gap.

Highlights

• —64-channel scalp EEG recorded at 250 Hz, BIDS-EEG 1.9.0 compliant
• —Two sessions per participant: pre- and post-intervention (6-week gap)
• —Five fixed task segments per session: restOE, restCE01, Medita, restCE02, slMedita
• —Three meditation paradigms compared head-to-head (HK, SA, BF)
• —Three preprocessing tiers shipped: raw BIDS, EEGLAB-cleaned, and ML-ready tensors
• —Three benchmark tasks with reference baselines (PSD+SVM, FBCSP+SVM, ShallowConvNet, DeepConvNet, EEGNet, EEG-Conformer)

Cohort and demographics

Attrition (30 / 74 = 40.5 %) was statistically Missing-Completely-At-Random with respect to baseline demographics and pre-intervention EEG features (see paper Appendix A.2).

Experimental paradigm

[image]

Each EEG session (≈ 32 min) contains five task segments recorded in fixed order:

The artefact-free silent block slMedita is the primary input for benchmark Tasks 1 and 3. Event codes embedded in the BrainVision marker files: Medita=1, restCE01=2, restCE02=3, restOE=4, slMedita=5.

Longitudinal design. Pre-intervention EEG → 6-week daily home practice (5 min/day week 1, 10 min/day week 2, 15 min/day weeks 3–6) with a mid-point check-in → post-intervention EEG.

Recording setup

• —Amplifier: mBrainTrain Smarting Pro X
• —Cap: 64-channel EASYCAP, Ag/AgCl, extended international 10-10 layout
• —Reference: FCz | Ground: FPz | Impedance: < 20 kΩ
• —Sampling rate: 250 Hz | Power-line: 60 Hz
• —Electrolyte: abralyt HiCl gel
• —Electrode digitisation: CapTrak coordinate system (per-session *_electrodes.tsv and *_coordsystem.json)

Psychometric assessments

Alongside the EEG, each participant completed three validated self-report instruments at both the pre- and post-intervention sessions:

• —Perceived Stress Scale (PSS) — 10-item scale of perceived life stress (lower is better).
• —Five Facet Mindfulness Questionnaire — Short Form (FFMQ-SF) — 24-item trait-mindfulness scale across five facets: Observing, Describing, Acting with Awareness, Non-judging of Inner Experience, Non-reactivity to Inner Experience (higher is better).
• —Multidimensional Assessment of Interoceptive Awareness, v2 (MAIA-2) — 37-item interoceptive-awareness scale across eight domains: Noticing, Not-Distracting, Not-Worrying, Attention Regulation, Emotional Awareness, Self-Regulation, Body Listening, Trusting (higher is better).

Longitudinal comparisons in the paper use a strict paired cohort of 43 participants (one subject's questionnaire packet is missing at both timepoints). Paired pre → post group means:

Headline findings (paper Appendix B.4.1):

• —PSS: stress dropped in all three groups; SA and HK reductions reached significance, BF was directionally consistent but did not reach significance in the small paired BF sub-cohort.
• —MAIA-2: interoceptive awareness rose across all groups. SA gains were strongest in Not-Distracting, Trusting, and Attention Regulation; HK in Attention Regulation, Self-Regulation, and Emotional Awareness; BF in Self-Regulation, Emotional Awareness, and Trusting, with negligible change in Noticing.
• —FFMQ-SF: trait mindfulness rose across all groups. SA showed small uniform gains (slight dip in Describing); HK broadly expanded across all five facets; BF gained most in Describing, Observing, Acting with Awareness, and Non-reactivity, with a discernible dip in Non-judging.

Raw per-participant scores are shipped in questionnaire_summary.xlsx at the repo root, with three sheets — PSS, FFMQ, MAIA-2 — each containing one row per (participant × session) with the assigned meditation technique, item-level responses, and pre-computed totals / subscale scores.

Repository structure

Subject IDs are 2-digit (sub-01…sub-74) under the raw BIDS root and 3-digit (sub-001…sub-074) under all derivatives/ tiers, following BIDS-Derivatives convention.

participants.tsv and participants.json at the root document each subject's group, session-completion status, age, sex, and handedness. Together with questionnaire_summary.xlsx, they cover all non-EEG per-subject metadata.

Preprocessing details

EEGLAB tier (`eeglab_preproc/`):

1. 1.Zero-phase 1 Hz Butterworth high-pass
2. 2.Zapline-plus 60 Hz line-noise removal (spectral integration, no notch artefact)
3. 3.Artifact Subspace Reconstruction (ASR, burst SD = 25, max-bad-channel = 0.2; bad channels spherically interpolated)
4. 4.Common-average re-reference including FCz (65 channels)
5. 5.Infomax ICA → ICLabel (artefact threshold ≥ 0.9) → component removal

Both pre-ICA (_preproc_preica) and post-ICA (_preproc_icrm) versions are kept.

ML tier (`ml_preproc_data/`):

• —0.5 Hz FIR high-pass + EEGLAB clean_rawdata (spatial-corr 0.9, line-noise 4 SD)
• —Spherical-spline interpolation of bad channels for uniform 64-channel topography
• —Per-segment epochs as float32 (C × T) NumPy arrays

ML continuous tier (`ml_continuous_tensors/`):

• —ICA-cleaned, full continuous signal stream (no segmentation), saved as *_desc-icacleaned_continuous.npy. Use this for sliding-window pipelines or self-supervised pre-training that needs the unsegmented stream.

Per-channel z-score normalisation is applied at run time by the dataset loader, not at storage time. Full parameter list: paper Appendix C.1.

Benchmark tasks

Reference benchmark code, configs, and baselines: <https://github.com/Angqi-Li/L-FAME-Benchmark>.

Quick start

The repository ships a one-call wrapper, load_benchmark, that handles dataset download, windowing, label assignment, and cross-validation splitting. Per-task auto-download keeps it to ~2–3 GB instead of the full 11 GB.

1. Install dependencies

pip install -U huggingface_hub torch numpy pandas scikit-learn

2. Pull the wrapper + dataset API

Both files live at the repo root and are downloaded once with hf_hub_download:

from huggingface_hub import hf_hub_download
for fn in ["dataset_api.py", "lfame.py"]:
    hf_hub_download("L-FAME-Dataset-Benchmark/L-FAME",
                    filename=fn, repo_type="dataset", local_dir=".")

3. Run a benchmark task in one call

import sys; sys.path.insert(0, ".")
from lfame import load_benchmark

# Task 1 — Cognitive State Decoding (rest vs. focused meditation)
train, val, test = load_benchmark(task_id=1, cv_strategy="inter")
# or LOSO holding out sub-05
train, val, test = load_benchmark(task_id=1, cv_strategy="loso", test_subject="05")

# Task 2 — Technique Classification (HK vs. SA vs. BF)
train, val, test = load_benchmark(task_id=2, cv_strategy="inter", task2_session="pre")

# Task 3 — Cross-Session Adaptation, 30-shot calibration on sub-12
train, val, test = load_benchmark(task_id=3, cv_strategy="intra_30_shot",
                                  test_subject="12")

# Standard PyTorch training loop
for X, y in train:
    print(X.shape, y.shape)   # torch.Size([64, 64, 1000]) torch.Size([64])
    break

The first call auto-downloads only the segments needed for that task_id into ~/.cache/huggingface/; subsequent runs are cache-hits.

Cross-validation strategies

Window length defaults to 4 s @ 250 Hz; overlap defaults are paper-faithful (50% for cross-subject, 87.5% for all intra-*). Override with window_sec=..., overlap_sec=....

Dev mode (small subset for pipeline debugging)

train, val, test = load_benchmark(
    task_id=1, cv_strategy="inter",
    download_subjects=["01", "02", "03"],   # only download these subjects (~150 MB)
    subject_filter=["01", "02", "03"],      # build the index from only these
)

Lower-level data access

If you need a different preprocessing tier, custom windowing, or just want raw files, skip the wrapper and use huggingface_hub.snapshot_download directly:

from huggingface_hub import snapshot_download
import numpy as np

# All ML-ready segmented tensors (≈ 11 GB)
root = snapshot_download(
    repo_id="L-FAME-Dataset-Benchmark/L-FAME", repo_type="dataset",
    allow_patterns=["derivatives/ml_preproc_data/**", "participants.tsv"],
)
arr = np.load(f"{root}/derivatives/ml_preproc_data/sub-001/"
              "sub-001_ses-premedita_task-slMedita_eeg_preproc.npy")
print(arr.shape, arr.dtype)   # (64, ~120000), float32

Note on the dataset viewer. The viewer panel on this page shows only participants.tsv. The raw BIDS files (.eeg/.vhdr/.vmrk/.set/.fdt/.npy) are not tabular and are intentionally excluded from automatic loading; download them via snapshot_download (above) or browse the Files and versions tab.

Ethics statement

All experimental procedures were approved by the Institutional Review Board (IRB) of Michigan State University. All participants provided written informed consent prior to the commencement of the study. The privacy rights of all human subjects have been observed throughout the research.

License

This dataset is released under CC BY-NC 4.0. You may copy, redistribute, remix, and build upon the material for non-commercial purposes with appropriate attribution. Commercial use requires explicit written permission from the authors. By downloading the dataset you agree not to attempt re-identification of any participant, and to comply with applicable data-protection regulations (HIPAA / GDPR where relevant).

Citation

If you use this dataset or benchmark, please cite:

@misc{li2026lfamelongitudinalfocusedattention,
      title={L-FAME: Longitudinal Focused Attention Meditation EEG Dataset and Benchmark},
      author={Angqi Li and Ab Basit Rafi Syed and Hamzeh Alzweri and Taosheng Liu and Barry H. Cohen and Saiprasad Ravishankar},
      year={2026},
      eprint={2605.22893},
      archivePrefix={arXiv},
      primaryClass={eess.SP},
      url={https://arxiv.org/abs/2605.22893},
}

Authors and acknowledgments

Authors: Angqi Li, Ab Basit Rafi Syed, Hamzeh Alzweri, Taosheng Liu, Barry H. Cohen, and Saiprasad Ravishankar.

We thank Devin O'Rourke and Sidharth Chhabra at The Harmony Collective (Ypsilanti, MI) for expert guidance on the three meditation training protocols. We are grateful to the undergraduate and graduate student research assistants who collected and curated the EEG sessions: Ab Basit Rafi Syed, Pratham Pradhan, Annie Wozniak, Vu Song Thuy Nguyen, Genevieve Orlewicz, and Alisia Coipel. Most of all, we thank the 74 participants who completed the demanding longitudinal protocol.

Changelog

• —v2.0 — 2026-06-11 —
• —Added per-participant psychometric data (questionnaire_summary.xlsx: PSS, FFMQ-SF, MAIA-2 across pre/post sessions) and a new Psychometric assessments section summarising the longitudinal findings.
• —Replaced the placeholder citation with the arXiv preprint (DOI 2605.22893).
• —Reworded Task 1 description ("focused meditation" → "silent meditation") so "focused" no longer ambiguously refers only to the silent block.
• —Added Michigan State University IRB ethics statement.
• —De-anonymised the README: filled in the real author block and replaced the anonymous.4open.science benchmark code link with <https://github.com/Angqi-Li/L-FAME-Benchmark>; removed all NeurIPS / double-blind / "upon acceptance" framing.
• —v1.0 — 2026-05 — Initial public release: 74-subject pre-intervention cohort, 44-subject post-intervention cohort, three derivative tiers, three benchmark task definitions.

<!-- TODOs to address in follow-up commits (not blocking this README):

1. 1.Add a ml_continuous_tensors FileSet to croissant_metadata.json distribution[]. -->