Yvhang009/eeg-fmri-ds006040-derived

EEG-fMRI ds006040 standardized derivatives

This public development dataset is the persistent, versioned data layer for an external NeuroBOLT replication on OpenNeuro ds006040.

The repository does not mirror the complete OpenNeuro dataset. OpenNeuro is the authoritative raw-data source. This repository records immutable source identifiers and checksums and will contain only standardized run-level EEG, DiFuMo BOLD features, synchronization evidence, QC records and preprocessing provenance.

Current status

The faithful v0.2 multistate pilot and its paired 15 Hz artifact-guard v0.3 branch are complete: 48 deterministic packages represent 24 runs from four participants and six conditions. The frozen split contains 12 train runs (sub-001, sub-005), six validation runs (sub-010) and six test runs (sub-019). Together they provide 4,502 eligible BOLD target volumes. Each faithful run has one guard companion with identical BOLD targets, timebase and target indices. These are preprocessing-validation samples, not a model-performance or clinical result.

All 24 paired audits passed. Relative to faithful EEG, the guard branch changed 1-10 Hz power by only -0.0220 to -0.0153 dB while attenuating 20-90 Hz power by 130.41 to 142.65 dB. This confirms construction of the sensitivity branch; it does not yet show whether NeuroBOLT performance depends on those frequencies.

The original v0.1 checkerboard manifest is preserved under manifests/legacy_v0-1/; its old shard and index remain for audit but are not training-eligible. The canonical checkerboard manifest now describes v0.2.

Each run package contains 200 Hz scalp EEG, causal 16 s EEG window indices, DiFuMo-64 BOLD targets with and without global-signal regression, a NeuroBOLT-faithful target, atlas labels, events, synchronization evidence, source hashes and complete preprocessing provenance.

Version 0.2 and v0.3 store original target-volume indices, hash all nine BIDS sources including the EEGLAB .fdt payload, excludes explicitly marked non-steady-state targets without motion scrubbing, and drops rather than pads targets that lack a complete 16-second causal EEG history. All 48 packages were validated locally and then anonymously re-downloaded at their upload revision.

Intended schema

• —index/*.parquet: faithful run indices;
• —index/artifact_guard_15hz/runs/*.parquet: paired guard indices;
• —shards/{train,validation,test}/*.tar: faithful shards;
• —shards/artifact_guard_15hz/*/*.tar: paired guard shards;
• —manifests/openneuro/*.json: versioned raw-source paths and hashes;
• —manifests/validation/*.json: modality and synchronization QC;
• —docs/storage_contract.md: lifecycle, versioning and privacy rules.
• —docs/preprocessing_and_sync_protocol.md: frozen signal and alignment rules.

Overlapping EEG contexts are constructed by the training loader after complete participant/run splits are assigned. They are not stored as duplicated samples.

Source and license

• —Source dataset: OpenNeuro ds006040, snapshot 1.0.1
• —Pinned source commit: fcba697fd4225ffbd735d593d1b7946663e178b4
• —Source license: CC0

Future clinical or controlled-access data are outside this license and must not be uploaded without explicit consent, IRB and data-use authorization.