Data

LEMON EEG preprocessed data.

This page serves as the stable hub for the preprocessed EEG data behind the Conv-VaDE microstate study (Brain Informatics, 2026). Links here are kept current; if a mirror moves, this page is updated, so cite this page rather than any single download link.

203

participants, healthy adults

18

cluster counts per participant, K = 3 to 20

3654

trained Conv-VaDE models

22

verified cohort tables

Sources

Primary source (canonical). Recordings originate from the MPI-Leipzig Mind-Brain-Body LEMON release (Babayan et al., Scientific Data, 2019). The exact archive used in the study, with paired eyes-closed and eyes-open EEGLAB files for the 203 analysed participants, is hosted by GWDG: EEG_Preprocessed_BIDS_ID/EEG_Preprocessed .

Processed mirror. Derived materials (topographic-map extracts, cohort metric tables, and the paper's regenerated figures) are mirrored on Google Drive: Conv-VaDE LEMON mirror . Drive folder locations can change; this page always carries the current link.

Repository

The cnn_vae_microstate repository contains all necessary components for auditing the study: the analysis code, 22 cohort tables of per-participant metrics across K = 3 to 20 for all 203 participants, and the figure pipeline that rebuilds the five data figures.

images/ carries the seven paper figures. cohort_tables/ holds 22 files of per-participant metrics: cluster-validity indices for Conv-VaDE and ModKMeans across K = 3 to 20 for all 203 participants, plus the derived summary tables behind every number quoted in the paper. analysis/fig_pipeline/ contains the scripts that regenerate the five data figures from those tables; build_verified.py constructs the tidy frames with completeness asserts, then the per-figure scripts run unchanged.

To reproduce the figures:

git clone https://github.com/fayisode/cnn_vae_microstate.git
cd cnn_vae_microstate/analysis/fig_pipeline
python3 build_verified.py          # asserts 203 x 18, no NaN
python3 distributions/raincloud_plots.py
python3 baseline_comparison/vae_vs_baseline.py
python3 k_selection_stats/k_selection.py
python3 run_optimal_k_primary.py   # K* distribution: K=4 76/203, K=3 60/203

Final results

The final outputs consist of verified cohort tables and five regenerated data figures. Every number quoted in the study traces to the cohort tables in cohort_tables/, and the derived tables and figures are mirrored in the Conv-VaDE LEMON Drive folder linked above.

Companion analysis: transition dynamics

A companion analysis examined whether the segmentations reproduce the transition dynamics established in the classical microstate literature. State sequences from both Conv-VaDE and ModKMeans were tested for Markov structure, recurrence properties, entropy, and long-range temporal correlations across all 203 participants.

A second-order G-squared test against a first-order Markov chain was insignificant for a mean of 76.6% of runs, so a first-order chain describes most of the cohort. Against first-order Markov surrogates, both methods depart significantly from chance (z = 7.5 for Conv-VaDE, z = 8.2 for ModKMeans). Conv-VaDE departs further in absolute magnitude: the mean absolute transition-matrix difference is 35.6 ± 15.5%, versus 23.6 ± 9.1% for ModKMeans, and it ranks higher in all 203 participants (paired p = 4.6 × 10⁻³⁵, d = 1.38).

Recurrence quantification puts determinism at 0.998 with dwell times retained, dropping to 0.341 with dwell removed. The apparent structure is carried by how long states persist, while the transition rule itself is near-memoryless. Entropy rate rises with the number of states, from 0.181 at K = 3 to 0.381 at K = 20, reflecting the cost of a finer state alphabet.

Occupancy is persistent: the Hurst exponent is 0.833, above 0.5 in every participant. The K trade-off is explicit. Decoder reconstruction stays flat across K (0.910 at K = 3 to 0.905 at K = 20), silhouette decreases from 0.184 to 0.103, and GEV rises from 0.495 to 0.702. Per-state variance gain drops below 0.01 at K = 8, indicating a data-driven balance at K = 8 to 9, while K = 4 remains preferred where alignment with the canonical four-microstate literature matters. Taken together, the sequences are persistent in occupancy, memoryless in transitions, and more structured under the generative model.

Preprocessing

What the preprocessing does, in order.

  • 61 scalp electrodes plus one VEOG channel, 10–10 montage, 250 Hz sampling, eyes-closed resting condition.
  • Common-average reference; zero-phase FIR bandpass, 2–20 Hz, to keep microstate boundaries temporally aligned.
  • Topographic maps extracted at local maxima of the Global Field Power signal with pycrostates, minimum inter-peak distance of three samples (12 ms).
  • Azimuthal-equidistant projection of electrode positions, cubic-spline interpolation to a 40 × 40 topographic image; points outside the electrode convex hull set to zero.
  • Per-participant z-score normalisation (clipped at ±5σ) with statistics estimated on the training split only, so no evaluation information leaks into the scaling.
  • Polarity invariance: each map and its sign-inverted counterpart are treated as the same microstate state throughout training and evaluation.

Participant identifiers and the full pipeline are in the cnn_vae_microstate repository . Questions and correction requests via the contact page .

EEG LEMON microstates open data

Please cite Babayan et al. (2019, Scientific Data 6:180308) for the LEMON dataset and the Brain Informatics paper for the processing pipeline.