Skip to content

DATA_SOURCE.md

Latest commit

 

History

History
19 lines (15 loc) · 1.52 KB

DATA_SOURCE.md

File metadata and controls

19 lines (15 loc) · 1.52 KB
Raw

Data source & provenance — SpikeNet2 (Li et al. 2025)

Raw / source data + weights (canonical home)

  • bdsp.io project: spikenet2 — SpikeNet 2.0 → https://bdsp.io/content/spikenet/2.0/ · DOI 10.60508/mbxb-hn49
  • S3 (credentialed / restricted): s3://bdsp-opendata-restricted/spikenet2/
    • EEG/ — continuous EEG (~260 GB) · Events/ — event .mat files (~35 GB)
    • Models/new_weights.ckpt — the trained model (369 MB); Models/1s-round11-hardmine-chan_weights-v1.ckpt — earlier round
  • Dataset: 17,524 EEGs (MGH/BWH) + 188 (Human Epilepsy Project) + 100 (SCORE-AI).

Proximal artifact committed in this repo (de-identified)

conbine_localization_predictions.csv — 848 events, per-channel model predictions + labels, subjects as surrogate Bonobo#### IDs (no PHI). Feeds 2_localization.ipynb (the localization figure regenerates from it with no download).

Raw → derived lineage

  1. Raw EEG + expert spike annotations (the spikenet2 dataset) →
  2. SpikeNet2 model (train_model.py → continurous.py → hard-mining → train_hard_mining.py) → new_weights.ckpt (on S3).
  3. Inference (1_calculate_local_predictions.ipynb / prediction.ipynb) on EEG + weights → predictions.csv / conbine_localization_predictions.csv.
  4. Figures: localization figure from the committed CSV (no data needed); detection/ROC figures from predictions.csv.

Weights + raw EEG are not committed (they live on S3); the small localization CSV is.