Skip to content

DATA_SOURCE.md

Latest commit

 

History

History
32 lines (27 loc) · 1.94 KB

DATA_SOURCE.md

File metadata and controls

32 lines (27 loc) · 1.94 KB
Raw

Data source & provenance — SPaRCNet

Raw / source data (the canonical home)

The raw EEG segments, expert labels, and model artifacts underlying this paper are published as the credentialed BDSP dataset:

  • bdsp.io project: bdsp-sparcnet — "SPaRCNet data: Seizures, Rhythmic and Periodic Patterns in ICU Electroencephalography" → https://bdsp.io/content/bdsp-sparcnet/
  • DOI: 10.60508/cw6j-s785
  • Access: credentialed (BDSP DUA); data served via the BDSP restricted S3 access point (bdsp-restricted-access-point). Apply on bdsp.io, then aws s3 sync per the dataset's usage notes.

Proximal artifacts in this repo (committed, de-identified)

code_for_figures/Data/ holds the small "last-mile" files each figure/table is computed from directly — model probabilities, expert consensus labels, precomputed UMAP coordinates, and cohort tables. These are de-identified: subjects are surrogate IDs (sparcnet_subject<NNNN>), no names/MRNs/dates. Total ~60 MB across 43 .mat files. See REPRODUCE.md for which file feeds which figure.

Raw → derived lineage

  1. Raw EEG (10 s, 128 Hz, 19-channel bipolar segments) + expert IIIC labels — in the bdsp-sparcnet dataset.
  2. SPaRCNet model (IIIC_train_0502.py, code/) is trained on the raw data → per-segment class probabilities.
  3. Proximal artifacts (code_for_figures/Data/*.mat): the model probabilities + expert operating points (Fig 1/2), the 2-D UMAP embedding + selected samples (Fig 3–6), spread/IRR summaries (Fig S3/S5/S8), and cohort split tables (Table 1).
  4. Figures/tables: the code_for_figures/*.m scripts turn (3) into the published figures — no raw data needed at this step.

To regenerate the proximal artifacts from raw data (rather than reproduce the figures from the committed artifacts), you need credentialed access to bdsp-sparcnet plus the training/inference code in code/ and a GPU.