---
license: cc-by-nc-sa-4.0
tags:
  - sleep
  - eeg
  - polysomnography
  - sleep-spindle
  - k-complex
  - rem
  - sleep-staging
language:
  - en
pretty_name: DREAMS Sleep Databases
size_categories:
  - n<1K
---

# DREAMS Sleep Databases (Mirror of Zenodo 2650142)

> **Source**: [Zenodo · DREAMS Databases](https://zenodo.org/records/2650142)
> **License**: CC BY-NC-SA 4.0 (academic / non-commercial use)
> **Mirror maintained by**: `lemonLHC` for internal sleep-event model training (EEG2LLM project)

This is a **private mirror** of the original DREAMS databases for convenience of training pipeline access. All credit goes to the original authors at the University of MONS — TCTS Laboratory and Université Libre de Bruxelles — CHU de Charleroi Sleep Laboratory.

## Contents (A-strategy: original `.rar` + docs)

| File | Size | Content after extraction |
|---|---|---|
| DatabaseSpindles.rar | 45 MB | 8 × 30-min excerpts, expert-annotated sleep spindles |
| DatabaseKcomplexes.rar | 44 MB | 10 × 30-min excerpts, K-complex annotations |
| DatabaseREMs.rar | 36 MB | 9 × 30-min excerpts, rapid eye movement annotations |
| DatabaseArtifacts.rar | 49 MB | 20 × 15-min excerpts, multi-class artifact annotations |
| DatabasePLMs.rar | 536 MB | 10 whole-night recordings, periodic limb movement annotations |
| DatabaseApnea.rar | 719 MB | 12 whole-night recordings, hypopnea annotations |
| DatabaseSubjects.rar | 1.0 GB | 20 healthy whole-night recordings (R&K + AASM hypnograms) |
| DatabasePatients.rar | 1.5 GB | 27 patient whole-night recordings (R&K + AASM hypnograms) |
| `*.m` / `*.txt` / `*.docx` / `*.pdf` | small | Original assessment algorithm + license + documentation |

## How to use

```bash
# 1. Clone (requires HF SSH key + git-lfs)
GIT_LFS_SKIP_SMUDGE=0 git clone git@hf.co:datasets/lemonLHC/DREAMS

# 2. Extract on macOS
brew install unar
cd DREAMS && for f in *.rar; do unar -f "$f"; done

# 3. Extract on Linux
sudo apt-get install unar  # or: bsdtar -xf <file>.rar
cd DREAMS && for f in *.rar; do unar -f "$f"; done
```

## Annotation channels & event types (실측 / measured)

| Sub-database | Channels (typical) | Sampling | Expert events |
|---|---|---|---|
| Spindles | C3-A1 / FP1-A1 / O1-A1 / EOG | 100 / 200 Hz | 355 spindles |
| Kcomplexes | CZ-A1 / FP1-A2 / O1-A2 | 200 Hz | 272 K-complexes |
| REMs | CZ-A1 / EOG1 / EOG2 | 200 Hz | 847 REM eye movements |
| Artifacts | CZ-A1 / EOG / EMG | 50 Hz | 2000 artifact events (10 sub-types) |
| PLMs | CZ-A1 / EMG1-3 | 200 Hz | 4421 PLM events |
| Apnea | CZ-A1 / NAF / VTH / VAB / SAO2 | 200 Hz | 4058 hypopnea events |
| Subjects | CZ-A1 / FP1-A2 / O1-A2 / EOG | 200 Hz | hypnogram only (5 s epoch) |
| Patients | CZ-A1 / FP1-A2 / O1-A2 / EOG | 200 Hz | hypnogram only (5 s epoch) |

## Hypnogram conventions

DREAMS uses a **5-second epoch** hypnogram (finer than AASM 30-second standard) to align with event-level annotations.

R&K coding:
```
5 = Wake     4 = REM     3 = N1     2 = N2     1 = N3     0 = N3 (stage 4 merged)     -1 = MT/Unscored
```

`Subjects/Patients` additionally provide AASM-coded hypnograms in `HypnogramAASM_*.txt`.

## Training plan reference

See [`sleep_event_training_plan.md`](./sleep_event_training_plan.md) for the full 5-class multi-label sleep event classifier training plan combining DREAMS + MASS data.

## Citation

If you use this dataset, please cite the original DREAMS publications:

```bibtex
@dataset{devuyst_dreams_2005,
    author = {Devuyst, Stephanie},
    title  = {The DREAMS Databases and Assessment Algorithm},
    year   = {2005},
    publisher = {Zenodo},
    doi    = {10.5281/zenodo.2650142}
}
```