Bonn University epilepsy EEG

Description

The Bonn dataset is the classic single-channel EEG benchmark for nonlinear time-series analysis and seizure detection, assembled at the Department of Epileptology, University of Bonn, and published by Andrzejak et al. (2001). It was created to test whether measures of nonlinear deterministic and finite-dimensional structure in brain electrical activity depend on recording region (scalp surface vs intracranial) and brain state (interictal vs ictal). The corpus contains 500 carefully selected, artifact-free 23.6-second segments drawn from a small pool of 10 individuals — 5 healthy volunteers (scalp recordings) and 5 presurgically-monitored epilepsy patients (intracranial recordings) — partitioned into five 100-segment sets (A—E, distributed as files Z, O, N, F, S). Despite its modest subject count, it remains one of the most widely used reference datasets in computational epilepsy research because of its clean, balanced design.

Citation

• Paper DOI: 10.1103/PhysRevE.64.061907
• Dataset DOI: TODO(confirm)
• Reference: Andrzejak, R. G., Lehnertz, K., Mormann, F., Rieke, C., David, P., & Elger, C. E. (2001). Indications of nonlinear deterministic and finite-dimensional structures in time series of brain electrical activity: Dependence on recording region and brain state. Physical Review E 64(6), 061907.

Catalog citation block

Andrzejak, R. G., Lehnertz, K., Mormann, F., Rieke, C., David, P., & Elger, C. E. (2001). Indications of nonlinear deterministic and finite-dimensional structures in time series of brain electrical activity: Dependence on recording region and brain state. Physical Review E, 64(6), 061907. DOI: 10.1103/PhysRevE.64.061907. (PDF included with the source.)

Download and access

Field	Value
Official source	https://www.upf.edu/web/ntsa/downloads
Access class	open
Size	five ZIP archives (one per set), each holding 100 plain-text files
Format	txt
BIDS	false
Mirrors	none recorded

Source & access

The dataset originates from the Andrzejak (2001) Physical Review E study and was distributed for years from the University of Bonn Department of Epileptology; the canonical mirror is now hosted by the Nonlinear Time Series Analysis group at Universitat Pompeu Fabra (UPF), Barcelona, at https://www.upf.edu/web/ntsa/downloads (the “EEG time series download” page). The accompanying paper is indexed under DOI 10.1103/PhysRevE.64.061907. Distribution is as five ZIP archives, one per set, each holding 100 plain-text files; the data are released for non-commercial/academic use (the original site requested citation of the 2001 paper). No OpenNeuro/PhysioNet/Zenodo accession exists for the original release; mirrors on Kaggle and IEEE DataPort are unofficial re-uploads.

Data structure

Each segment is a single-column ASCII text file (one integer amplitude sample per line, 4,097 lines), with no header and no per-channel label. Files are named by set prefix and index — e.g. Z001.txt—Z100.txt in Z.zip (set A), and similarly O###, N###, F###, S### for sets B—E. There is no BIDS layout, no events/sidecar metadata, and no DataLad/git-annex packaging; the release is flat ZIP archives of .txt files. Sampling rate (173.61 Hz) and filtering are documented only in the paper, not embedded in the files.

License

• Name: informal-academic
• Note (verbatim from the directory): non-commercial / academic use with citation requested; no formal license text (TODO(confirm), §13 item 19)
• Snippets on this site: no — the site ships no snippet or precomputed product from this dataset; lessons link the official access page.

Acquisition

Field	Value
Device	128-channel amplifier system; scalp 10-20 electrodes (sets A/B) and intracranial depth/strip electrodes (sets C–E)
Device class	TODO(confirm)
Channels	1
Sampling rate (Hz)	173.61
Online filters	0.53–40 Hz band-pass; 12-bit
Reference	average common reference
Mains frequency (Hz)	TODO(confirm)
Paradigms	rest-eo, rest-ec, interictal, ictal
Participants	500 (500 single-channel 23.6-s segments from ~10 individuals (5 healthy + 5 epilepsy); registry counts segments)
Population	10 individuals: 5 healthy volunteers (scalp), 5 epilepsy patients (intracranial)
Clinical groups	epilepsy, HC
Sessions	TODO(confirm)
Durations	500 artifact-free 23.6-s single-channel segments (4,097 samples each) in five 100-segment sets A–E (files Z/O/N/F/S)

Acquisition

All 500 segments were recorded on the same 128-channel amplifier system, digitized at 173.61 Hz with 12-bit A/D resolution, using an average common reference (Andrzejak, 2001). A 0.53—40 Hz bandpass filter was applied to every signal. Surface (scalp) recordings in sets A/B used standardized 10—20-system electrode placement on the healthy volunteers. Intracranial recordings in sets C/D/E used intracranial depth and strip electrodes implanted for presurgical monitoring — set C/D from depth electrodes placed symmetrically in the hippocampal formations, and set E from strip/depth contacts in the seizure-generating area. The published single-channel segments are individual channels extracted from the multichannel continuous recordings. Each released segment is 4,097 samples = 23.6 s. The full corpus is 500 segments (5 sets x 100).

Participants

Ten individuals total: 5 healthy volunteers and 5 epilepsy patients (Andrzejak, 2001). The healthy volunteers were recorded awake and relaxed; the patients had undergone presurgical diagnosis with a definite epileptogenic focus identified, and their recordings come from the period of presurgical intracranial monitoring. The dataset is not segmented per subject — the 100 segments within each set are pooled across that set’s subjects rather than labeled by individual. Age, sex, and per-subject demographics are not documented in the dataset distribution.

Tasks / conditions

There is no behavioral task — segments are continuous resting/clinical EEG selected to be artifact-free. The five sets encode recording region and brain state. The canonical lettering convention (Andrzejak, 2001) maps the published file prefixes to the A—E sets as follows:

Set	File prefix	Source	Region	State
A	Z	Healthy volunteers (scalp)	Standardized 10—20 surface	Awake, eyes open
B	O	Healthy volunteers (scalp)	Standardized 10—20 surface	Awake, eyes closed
C	N	Epilepsy patients (intracranial)	Hippocampal formation of the opposite (contralateral) hemisphere	Seizure-free (interictal)
D	F	Epilepsy patients (intracranial)	Within the epileptogenic zone	Seizure-free (interictal)
E	S	Epilepsy patients (intracranial)	Seizure-generating area	Seizure (ictal) activity

Sets A and B are surface EEG from the healthy controls; sets C, D, E are intracranial EEG from the patients. Sets C and D are both interictal (seizure-free) — they differ by recording site, not by state — while set E is the only ictal set.

Used on this site

L0.1, L1.3, L6.3

Loader

Fetches only the stated subset. Recorded loader: TODO(confirm).

TODO(confirm) — the catalog records no loader for this dataset.

Caveats worth knowing

• 500 segments from 10 people (pseudoreplication).
• Two lettering conventions (A–E vs Z/O/N/F/S) that some secondary sources swap.
• Scalp and intracranial sets differ by orders of magnitude in amplitude.

Notable caveats

• Mixed recording modalities. Sets A/B are scalp surface EEG while C/D/E are intracranial — spectra and amplitudes differ by orders of magnitude, so pooling them in a single analysis crosses recording modalities and should be done deliberately, not by accident.
• No notch needed. The 0.53—40 Hz hardware bandpass removes 50/60 Hz line frequencies, so there is no notch-hole in the spectrum.
• Limited usable band. The 40 Hz low-pass means nothing above ~40 Hz is physiological signal.
• Tiny, non-independent subject pool. Only 10 individuals (5 + 5); the 100 segments per set are not independent subjects, which inflates apparent sample size and limits generalization — a frequently-noted limitation when the dataset is used for machine-learning benchmarks.
• Lettering ambiguity. Two conventions circulate: the original A—E labels and the file-prefix Z/O/N/F/S labels. Some secondary sources swap A↔B (eyes open vs closed) or C↔D (contralateral hippocampus vs epileptogenic zone); the mapping in the table above follows the original Andrzejak (2001) definitions.
• Selection bias. Segments were hand-selected to be artifact-free and to satisfy a stationarity/quality criterion, so the corpus is not a representative continuous recording — it is a curated benchmark.

Related datasets

• ERP CORE (Compendium of Open Resources and Experiments)
• EEG Motor Movement/Imagery Dataset (EEGMMIDB)
• LEMON (MPI-Leipzig Mind-Brain-Body)
• Dortmund Vital Study resting EEG
• SRM Resting-state EEG

# Loader: TODO(confirm) — the catalog records no loader for this dataset.
# Download a small subset from the official source: https://www.upf.edu/web/ntsa/downloads