Datasets:
File size: 5,592 Bytes
5b4daa9 10d487d 477ef2e 5b4daa9 09bdecd 5b4daa9 bc5cd49 5b4daa9 bf17407 5b4daa9 8866d28 5b4daa9 09bdecd 5b4daa9 09bdecd |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 |
---
license: other
license_name: commercial
license_link: LICENSE
tags:
- eeg
- synthetic-data
- biosignals
- healthcare
- neuroscience
- machine-learning
- deep-learning
- time-series
- seizure-detection
- real-world-evidence
- data-generation
- biomedical
- brain-signals
- artificial-data
- education
size_categories:
- 1M<n<10M
task_categories:
- time-series-forecasting
language:
- en
pretty_name: DBbun Synthetic EEG (10–20 @ 250 Hz)
---
# DBbun — Synthetic Scalp EEG (10–20 @ 250 Hz)
## Overview
DBbun EEG is a generated collection of synthetic scalp EEG recordings that replicate the structure and statistical properties of real clinical EEG while remaining fully privacy-safe.
Each file contains multi-channel 10–20 EEG sampled at 250 Hz, with per-second seizure labels and optional TCP (bipolar) differential channels.
The dataset enables research and commercial development in seizure detection, artifact rejection, biosignal modeling, and algorithm benchmarking, without any regulatory or privacy restrictions.
All content is synthetically generated and contains no patient or institutional data.
This work was inspired by [the following study](https://pubmed.ncbi.nlm.nih.gov/33745882/): Roy S, Kiral I, Mirmomeni M, Mummert T, Braz A, Tsay J, Tang J, Asif U, Schaffter T, Ahsen ME, Iwamori T, Yanagisawa H, Poonawala H, Madan P, Qin Y, Picone J, Obeid I, Marques BA, Maetschke S, Khalaf R, Rosen-Zvi M, Stolovitzky G, Harrer S; IBM Epilepsy Consortium. Evaluation of artificial intelligence systems for assisting neurologists with fast and accurate annotations of scalp electroencephalography data. EBioMedicine. 2021 Apr;66:103275. doi: 10.1016/j.ebiom.2021.103275. Epub 2021 Mar 18. PMID: 33745882; PMCID: PMC8105505.
---
## Dataset Composition
### Splits
| Split | Records | Duration | Purpose |
|:------|:--------:|:---------:|:--------|
| train/ | 200 | approximately 200 hours | Model training |
| valid/ | 25 | approximately 25 hours | Hyperparameter tuning |
| test/ | 25 | approximately 25 hours | Evaluation |
Each split includes:
- `patient_XXXXXX.npz` — EEG waveform and label arrays
- `patient_XXXXXX_events.json` — seizure interval metadata
- `metadata.csv` — session-level descriptors
### File contents
| Key | Type | Shape | Description |
|:----|:-----|:------|:------------|
| `eeg` | float32 | [n_channels, n_samples] | Scalp EEG waveforms (10–20 layout ± TCP montage) |
| `sr` | int | — | Sampling rate (250 Hz) |
| `channels` | list[str] | [n_channels] | Channel names |
| `labels_sec` | uint8 | [seconds] | 0 = non-ictal, 1 = ictal |
---
## Validation and Quality Assurance
### Quantitative validation
Descriptive realism metrics were computed using the included notebook `DBbun_EEG_Validation.ipynb`, which summarizes power ratios, durations, and seizure prevalence.
Results are saved in `validation/realism_report.md`.
Key findings:
| Metric | Mean | Std | Description |
|:--------|-----:|----:|:------------|
| Alpha ratio | 0.496 | 0.020 | Balanced 8–13 Hz band typical of resting EEG |
| Delta ratio | 0.428 | 0.021 | Realistic low-frequency power |
| Channels | 38 | 0 | 20 raw + 18 TCP differential channels |
| Duration (s) | 3469 | 514 | Approximately 58 minutes per record |
| Seizure fraction | 0.0126 | 0.012 | Approximately 1.3% ictal content |
Validation confirmed:
- Stable 10–20 montage structure
- Physiologically plausible alpha/delta ratios and amplitudes
- Realistic seizure sparsity and variability
- Clean baselines without abnormal drift
- Consistent metadata across all splits
### Visual previews
Representative 10-second EEG windows were exported as `previews/*.png`.
Each shows realistic amplitude, rhythmic composition, and spectral balance.
---
## Comparison with Other Synthetic EEG Datasets
A variety of research and open projects have proposed methods for generating synthetic EEG, but very few provide long, continuous, multi-channel signals with structured metadata.
The table below summarizes key characteristics of known efforts compared to DBbun EEG.
| Feature | Existing Synthetic EEG Datasets | DBbun EEG |
|:--------|:-------------------------------|:-----------|
| Continuous multichannel raw EEG signals | Rare – most provide only extracted features or short samples | Yes – full 10–20 recordings (≈1 hour each) |
| Per-second or sample-level annotations | Usually absent or limited to class-level labels | Yes – one label per second (ictal / non-ictal) |
| Number and duration of sessions | Typically fewer than 10 sessions, often only a few minutes each | 250 sessions totaling over 250 hours |
| Validation and realism metrics | Often qualitative or absent | Full quantitative and visual validation |
| Montage and channel structure | Often single-channel or unspecified | Standard 10–20 + TCP differential montage |
| Licensing and accessibility | Often research-only or restricted | Commercial license with clear permitted use |
| Intended applications | Algorithm demos or augmentation | Training, validation, and commercial prototyping |
This comparison highlights DBbun EEG’s unique position as a large-scale, validated, and legally distributable synthetic EEG dataset.
---
## How to Use
### Load one record
```python
import numpy as np
z = np.load("data/train/patient_000001.npz")
eeg = z["eeg"]
sr = int(z["sr"])
channels = [c.decode() if isinstance(c, bytes) else c for c in z["channels"]]
labels = z["labels_sec"]
print("Channels:", len(channels))
print("Shape:", eeg.shape)
print("Duration (s):", eeg.shape[1] / sr)
|