EEG Spikes

Pennsieve (University of Pennsylvania)

Detects interictal epileptiform discharges (spikes) on clinical scalp EEG with the Janca et al. 2015 (or Barkmeier et al. 2012) detector from epycom, both developed and validated on intracranial EEG, on the longitudinal bipolar montage; writes spikes, per-channel rates, multi-channel spike events with their leading channel, lateralisation and rates per sleep stage, leaving seizures and bad segments out.

GitHub Release Tag: v0.1.1
GitHub Commit Hash: 9a567bc
DOI: 10.26275/vxgc-rtud
Updated on November 13, 2026 (Version 2)

Corresponding Contributor:

Joost Wagenaar
Dataset Banner Image

Overview

This app packages the interictal spike detectors of epycom (ElectroPhYsiology COmputational Module), developed by the Biomedical Engineering Department of the International Clinical Research Center at St. Anne's University Hospital in Brno (FNUSA-ICRC) with the Institute of Scientific Instruments of the Czech Academy of Sciences (both copyright holders per its source files; https://github.com/ICRC-BME/epycom, Apache-2.0). The default detector is the envelope-distribution method of Radek Janca and colleagues (Janca et al. 2015, Brain Topography, doi:10.1007/s10548-014-0379-1), reprogrammed in Python for epycom by Daniel Uher and Jan Cimbalnik (per epycom's source); the alternative is the slope-and-amplitude method of Daniel Barkmeier, Jeffrey Loeb and colleagues (Barkmeier et al. 2012, Clinical Neurophysiology, doi:10.1016/j.clinph.2011.09.023). Recordings are read with MNE-Python (Alexandre Gramfort and the MNE community; https://mne.tools; Gramfort et al. 2013). Pennsieve packages these methods for scalp EEG and validated them on scalp data (below); it did not develop them. The demo and validation data are the Siena Scalp EEG Database (Paolo Detti and colleagues, University of Siena; BIDS conversion by Jonathan Dan, EPFL, with Paolo Detti, published on Zenodo by EPFL), CHB-MIT (Ali Shoeb, MIT, with Children's Hospital Boston) and vEpiSet (Nan Lin, Qiang Lu and colleagues, Peking Union Medical College Hospital) and OpenNeuro ds003555 (Dorottya Cserpan, Johannes Sarnthein and colleagues, University Hospital Zurich).

Both detectors were developed and validated on intracranial EEG, not on scalp EEG (Janca et al. 2015: "epileptic and non-epileptic intracranial recordings"; Barkmeier et al. 2012: "intracranial EEG"). Our own scalp validation, below, shows they find most expert-marked scalp spikes but also many non-epileptiform transients. Every rate, event and laterality index this app writes is a research output to be reviewed against the EEG, not a clinical finding.

A step of both scalp EEG workflows (workflow-scalp-eeg):

  • Scalp EEG Interictal Biomarkers: eeg-prep → {eeg-spectral, eeg-connectivity, eeg-sleep → eeg-spikes} → eeg-report
  • Scalp EEG Seizure Detection & Monitoring: eeg-prep → {eeg-seizures, eeg-sleep → eeg-spikes} → eeg-report

For the epileptologist. The app marks candidate interictal epileptiform discharges (spikes and sharp waves) on the longitudinal bipolar montage ("double banana") it derives from the referential recording, as you would read it. Spikes on different channels within 50 ms are merged into one spike event with its leading channel (the one that peaks first) and that channel's hemisphere. Per recording you get the spike rate of every channel, a laterality index of the events (left vs right leading channel) and the most frequent region of the leading channels: these say where the detector fired, not a lateralising or localising finding (on Siena they agreed with the seizure side no better than always answering "left"). With eeg-sleep's hypnogram linked, every spike gets the sleep stage of its epoch and rates are given per stage: interictal spiking increases with sleep depth in temporal lobe epilepsy (Malow et al. 1998), and spike rates around sleep help localise the seizure-onset zone (Conrad et al. 2023, an intracranial result). Seizures (with 60 s margins) and eeg-prep's bad segments are left out. Look at the example-spike figure first: if the examples are not epileptiform, neither are the rates.

For the researcher. epycom 0.3's detect_spikes_janca (10–60 Hz envelope, log-normal background fitted in 5 s windows, threshold k1 = 3.65) or detect_spikes_barkmeier runs per channel on the signal resampled to 200 Hz (polyphase FIR), in 5-minute windows with 10 s context, channels in parallel. Each detection is refined to the largest deflection of the 1–70 Hz signal within ±40 ms (amplitude, sign) and given a base-to-base duration. Outputs are long-format CSV keyed by recording_id and channel (workflow-scalp-eeg docs/conventions.md), shaped like processor-ieeg-spikes' tables so scalp and intracranial results sit side by side, plus features.csv (spikerate_<detector>) for normative comparison.

Ports
Port Direction Placement What
eeg input INPUT_DIR/eeg/ eeg-prep's BIDS-EEG derivative (recordings.csv, *_desc-preproc_eeg.edf, _channels.tsv, _events.tsv). Also any BIDS-EEG dataset (BIDS Siena's Fp1-Avg names, old T3–T6 names and HED-SCORE eventType seizure rows are understood) and loose EDF/BDF/BrainVision/FIF/EEGLAB files with 10-20 names, referential or bipolar (FP1-F7; a CHB-MIT chbNN-summary.txt gives their seizures). Found recursively, symlinks followed.
sleep input, optional INPUT_DIR/sleep/ eeg-sleep's hypnogram.csv (30 s epochs, stages W/N1/N2/N3/R), matched by recording_id. Without it, annotated sleep/wake events of the recording are used when present.
spikes output OUTPUT_DIR/eeg-spikes/ (or eeg-spikes/<LABEL>/) tables, figures/, summary.json, provenance.json (below)

Files are found by name and content anywhere under INPUT_DIR, so the app also works when everything is merged at its root.

Parameters
Name Default What it does
DETECTOR janca janca (Janca et al. 2015) or barkmeier (Barkmeier et al. 2012). Both intracranial-validated; see Validation.
JANCA_K1 3.65 Janca's threshold factor (the published default). Higher: fewer, more obvious spikes. On vEpiSet, 5 cuts false detections in patients without IEDs from 18.7 to 8.0/min but finds 51 % instead of 75 % of the expert-marked spikes.
MONTAGE bipolar bipolar: the 18-pair longitudinal bipolar montage (Fp1-F7, F7-T7, T7-P7, P7-O1; Fp1-F3, F3-C3, C3-P3, P3-O1; Fz-Cz, Cz-Pz; and the right side), from pairs whose two electrodes are good. referential: the channels as stored (eeg-prep's average reference). Bipolar-only recordings are always analysed on their own pairs (with a note).
EXCLUDE_SEIZURE true Leave annotated seizures out, SEIZURE_MARGIN_S on both sides; a seizure without an end is cut to the end of the recording. bad_segment and artifact events are always masked: the detector runs through them (its background stays continuous) but what it finds there is dropped and their time is not counted.
SEIZURE_MARGIN_S 60 Seconds left out before each seizure and after it.
WINDOW_S 300 Windows the recording is read and analysed in (plus 10 s context each side). Bounds memory. 60–570: epycom 0.3's Janca detector fails on reads of 600 s or more.
GROUP_WINDOW_MS 50 Spikes within this many ms of an event's first spike belong to that event.
MIN_EVENT_CHANNELS 1 Events spanning fewer channels stay in spike_events.csv but are not counted in lateralisation.csv, the event rates in summary.json, the rates per stage and the example figure. 2–3 asks for a field; on vEpiSet it trades sensitivity for specificity without improving their balance (Validation).
THREADS 0 Channels analysed in parallel (processes); 0 = one per vCPU of the container (cgroup CPU quota, else CPU affinity, else CPU count), capped so ~250 MB per worker + ~600 MB fit the memory limit. Logged and kept in provenance.json.
LABEL (empty) Write to eeg-spikes/<LABEL>/ (a workflow running this step twice).
Outputs

Everything is under OUTPUT_DIR/eeg-spikes/. Times are seconds from the recording start.

spikes.csv: one row per detected spike: recording_id, channel (F7-T7, or T7 when referential), time_s (peak), amplitude_uv (signed peak of the 1–70 Hz signal; on bipolar pairs the sign depends on which electrode of the pair is more negative), duration_ms (base to base: the turning points on either side of the peak, within 150 ms; 5 ms resolution), detector, sleep_stage (empty when no stages are known), event_id.

spike_rates.csv: one row per analysed channel, zeros included: recording_id, channel, n_spikes, minutes, rate_per_min, detector. minutes is the time actually analysed (seizures with margins, bad segments, and the first and last second of each analysed stretch excluded).

spike_events.csv: one row per multi-channel spike event: recording_id, event_id, time_s (leading channel's peak), n_channels, channels (|-separated, firing order, leader first), leading_channel (earliest peak; ties within one 5 ms sample go to the larger amplitude), hemisphere (L/R/Z), region, detector, sleep_stage. Isolated spikes are events with n_channels = 1.

lateralisation.csv: per recording recording_id, n_left, n_right, n_midline (counted events by the hemisphere of their leading channel), laterality_index = (L − R)/(L + R) (empty with no lateral events), dominant_region (most frequent side and region of the leaders, e.g. left temporal; where the detector fired most, not a localising finding), and n_events, minutes, rate_left_per_min, rate_right_per_min (mean per-channel rate over left/right channels), rate_laterality_index, dominant_channel, montage, detector.

rates_by_stage.csv (when stages are known): recording_id, channel, stage, n_spikes, minutes, rate_per_min, stages W, N1, N2, N3, R (or W/sleep from annotations).

features.csv: recording_id, channel, feature, value with spikerate_janca or spikerate_barkmeier.

figures/ (150 dpi PNG, captions in summary.json): <recording_id>_spike_rates.png (topomap of the rate per electrode, for bipolar pairs the mean of the pairs an electrode belongs to, next to every channel's rate), <recording_id>_spikes_by_stage.png (events/min per stage), <recording_id>_example_spikes.png (three events of the most frequent leader, at the median, 75th and 90th percentile of its amplitude, all channels stacked), laterality_index.png (all recordings).

summary.json: headline numbers per recording (minutes analysed, spikes, events, events/min, top channel, laterality index, most_frequent_leader_region and most_frequent_leading_channel, events/min per stage, NREM/wake ratio), the figures in report order with captions, the tables, warnings, and a notice that the detectors are intracranial-validated and that laterality and regions are not lateralising or localising findings. Detector failures are never silent: a channel on which the detector fails in some windows is listed in provenance.json (channel_errors: windows failed, first error), the log and summary.json warnings; a channel that fails in every window is named in the recording's notes; a recording on which every channel fails is failed, and the step exits 1 when every recording failed. provenance.json: the keys of conventions.md (tools and versions, parameters, inputs, times, peak RSS, threads, per recording status/seconds/montage plus analysed and excluded intervals, channels not analysed and why, notes), citations (DOIs for the report).

Validation

Run by us in November 2026; scripts in validation/. Numbers are for the default parameters unless stated.

(a) Synthetic fixture (tests/make_fixture.py, 19-channel 10-20 EEG at 256 Hz in average reference with posterior alpha, a 3 Hz spike-and-wave seizure, an artefact marked bad_segment, sleep-like slow waves, and planted surface-negative spikes with a temporal field, 80 % left (T7), 20 % right (T8), twice as frequent in sleep; plus a bipolar-only copy). Event-level recovery at ±50 ms: sensitivity 1.00, precision 1.00 on both the referential and the bipolar-only recording (91 and 39 planted spikes), median timing error 1.8 ms, leading-channel hemisphere correct for every event, laterality index 0.54 for a planted 0.54, no detections inside the seizure ± 60 s or the bad segment, rates per stage N2/N3 ≈ 2× wake as planted. The same signal given as plain BIDS without the bad_segment annotation shows why the mask matters: the artefact's sharp transients come out as 34 false events. pytest asserts all of this (38 tests). A clean synthetic fixture is a test of the plumbing, not of scalp performance.

(b) vEpiSet (Lin et al. 2025; figshare doi:10.6084/m9.figshare.28069568.v2, CC BY 4.0): 30 of the 84 continuous 20-minute recordings (20 randomly chosen of the patients with IEDs, 10 of those without), 19 scalp electrodes at 500 Hz in average reference, 1,477 single IED marks by consensus of at least three experts. Epochs cut as the dataset's own code cuts them (4 s; 900 IED epochs, 8,135 non-IED epochs). Epoch score = channel detections in the epoch; an epoch is "detected" when it holds an event spanning ≥ m channels.

Setting AUROC Expert marks with an event within ±0.25 s Events/min in patients without IEDs m=1: sens / spec / precision m=2 m=3
janca, k1 3.65, bipolar (default) 0.79 0.75 18.7 0.87 / 0.57 / 0.18 0.71 / 0.73 / 0.23 0.58 / 0.84 / 0.29
janca, k1 5 0.73 0.51 8.0 0.64 / 0.80 / 0.26 0.50 / 0.88 / 0.31 0.38 / 0.93 / 0.37
janca, k1 6 0.70 0.39 5.4 0.52 / 0.86 / 0.29 0.40 / 0.91 / 0.33 0.27 / 0.95 / 0.37
janca, k1 8 0.63 0.22 3.0 0.35 / 0.91 / 0.30 0.23 / 0.95 / 0.33 0.13 / 0.97 / 0.33
janca, k1 3.65, referential 0.82 0.78 18.2 0.91 / 0.58 / 0.19 0.76 / 0.78 / 0.27 0.56 / 0.88 / 0.34
barkmeier, bipolar 0.73 0.52 22.2 0.70 / 0.70 / 0.20 0.49 / 0.87 / 0.30 0.35 / 0.93 / 0.36

For comparison, the dataset authors' deep-learning models reach 97 % specificity and 74 % precision at 80 % sensitivity on held-out epochs (VGG; Lin et al. 2025, Table 2). The classical detectors find most expert-marked spikes but at that sensitivity flag about 40 % of non-IED epochs: most detections on scalp EEG are not epileptiform discharges (muscle, eye movements, vertex waves, sharp background). Simple amplitude, duration and channel-count criteria did not improve the balance (we tried them on these data and left them out rather than tune them on 30 recordings). The referential montage did slightly better than bipolar here; we keep bipolar as the default because clinicians read spikes on it and the two are within noise of each other. vEpiSet's wake/sleep annotations (used through the annotation fallback): the experts' IED rate was higher in sleep than in wake in 8 of 9 patients with both states, the detector's in 6 of 9 (in wake, artefact detections dominate). vEpiSet is not staged by the workflow's fetch script (16.9 GB); the 30 recordings were read out of the figshare zip with HTTP range requests (validation/vepiset_to_bids.py, validation/vepiset_eval.py). The 30 recordings are listed in validation/data/vepiset_recordings.txt (the first 20 with IEDs, then 10 without); the scores of each setting in the table are validation/data/vepiset_*.json.

(c) Siena, per-patient laterality of spike events against the seizure lateralisation in PhysioNet's subject_info.csv (13 patients with a left or right temporal focus, 9 left and 4 right; PN10, bilateral frontal, left out), all runs of a patient pooled (validation/siena_lateralisation.py; subject_info.csv is fetched from PhysioNet, https://physionet.org/files/siena-scalp-eeg/1.0.0/subject_info.csv):

Input (41 recordings, 141 h) Events/min per recording, median (IQR) Laterality agrees with the seizure side Patients with a left-sided index
eeg-prep derivative (ICA + ICLabel; bad segments for technical artefacts only, 4.6 % of the time; data/work/siena/eeg-prep/eeg) 21 (14–35) 9 of 13 (sign test p = 0.27); also 9 of 13 with MIN_EVENT_CHANNELS 2–3 11 of 13
plain BIDS Siena (no artefact handling; output in data/work/siena/eeg-spikes-raw-bids/) 40 (31–65) 4 of 13 (p = 0.27); 6–8 of 13 with MIN_EVENT_CHANNELS 2–6 or the referential montage 7 of 13

eeg-prep's artefact handling halves the detections and moves the agreement from below to above chance, but 9 of 13 is what always answering "left" scores on this cohort (9 of the 13 foci are left): the index came out left-sided in 11 of 13 patients, including 3 of the 4 right-sided ones. So on Siena the app's lateralisation does not reliably recover the seizure side, and a left bias (artefact or montage asymmetry) cannot be excluded. Rates of 21 events/min are implausibly high for interictal spiking on scalp EEG, which again says most detections are not IEDs. The seizure side is also only a proxy: interictal spikes need not be on the seizure side. Siena is low-passed at 30 Hz by its source.

With eeg-sleep's hypnograms (35 of the 41 Siena recordings staged), the spike-event rate was higher in NREM than in wake in only 10 of 35 recordings (median NREM/wake ratio 0.62), the opposite of the sleep activation expected of true interictal spikes (Malow et al. 1998): in wake the detections are dominated by muscle and movement artefact. Read rates per stage from this app only together with the example-spike figure.

CHB-MIT (chb01, 6 one-hour bipolar-only recordings, 4 with a seizure, read directly; output in data/work/chbmit/eeg-spikes-raw-edf/): runs as recorded on its 22 distinct pairs, seizures from chb01-summary.txt left out with their margins; median 27 events/min. Through eeg-prep (all 42 chb01 recordings, kept bipolar-longitudinal, data/work/chbmit/eeg-spikes/): 42 ok, median 30 events/min, median laterality index −0.01. With eeg-sleep's chb01 hypnograms linked (39 of 42 recordings staged; 36 with both NREM and wake), the spike-event rate was higher in NREM than in wake in only 9 of 36 recordings (median NREM/wake ratio 0.79), as on Siena: artefact in wake dominates the detections. No spike annotations exist for it.

ds003555 (eeg-prep and eeg-sleep outputs, 3 children, 3 h of sleep EEG each; the workflow's own path, output in data/work/ds003555/eeg-spikes/): spike events per minute in W / N1 / N2 / N3 / R: sub-01 (bilateral ischaemic lesions) 92 / 94 / 83 / 103 / 84; sub-02 (genetic) 21 / 13 / 17 / – / 13 (no N3 scored); sub-03 (suspected FCD) 66 / 82 / 125 / 136 / 89, i.e. NREM/wake 1.01, 0.80 and 1.91. The dataset has no spike annotations to check these against; at these rates sleep transients (vertex waves, K-complexes) are certainly among the detections.

Runtime

computeTypes: [standard], 2 vCPU / 4 GiB (Fargate's smallest size for 2 vCPU; peak use is about 1.6 GB), 2 h timeout. One recording at a time, one 5-minute window of every channel at a time, channels in parallel; memory is bounded by the window, not by the recording length.

Measured in Docker (linux/amd64 emulated on an M-series Mac, --cpus 2 --memory 3g):

Input Wall time Peak memory (cgroup memory.peak, page cache included)
Siena sub-01 run-00 (plain BIDS): 13.5 h, 19 channels, 256 Hz (803 min analysed, 18 pairs) 105 s 1.23 GB (main process 0.70 GB RSS)
ds003555 eeg-prep derivative + eeg-sleep hypnograms: 3 × 3 h, 23 channels, 1024 Hz 87 s 1.60 GB (main process 0.61 GB RSS)
synthetic fixture (2 recordings, 9 min), --read-only --user 65534 Lambda-style 8 s 0.85 GB

About 8 s per recording-hour on 2 emulated vCPU (native arm64: all 141 h of Siena in 5.3 min on 4 workers). A 24 h recording takes ~3–4 min; the 2 h timeout leaves room for a few days of recordings in one run. Image: 212 MB compressed.

Running it locally
python3.12 -m venv .venv && .venv/bin/pip install --no-deps --only-binary=:all: -r requirements-dev.txt
make test PYTHON=.venv/bin/python     # pytest on the synthetic fixture
make run PYTHON=.venv/bin/python      # Docker over the fixture -> data/output/
make run-demo DEMO=../workflow-scalp-eeg/data/work/siena/eeg-prep/eeg   # an eeg-prep output
Limitations
  • Intracranial methods on scalp EEG. Janca's and Barkmeier's detectors and their thresholds were developed on intracranial recordings, where spikes are large relative to the background and there is no scalp muscle or eye artefact. On scalp EEG they are sensitive but not specific (Validation b): read rates as "candidate discharges", review the example figure, and compare recordings with the same montage and preprocessing, not across datasets.
  • No artefact classifier. Muscle, eye movements (frontopolar and F7/F8 channels), electrode pops and sleep transients (vertex waves, K-complexes, POSTS) are detected as spikes. eeg-prep's ICA and bad segments remove some of them; nothing in this app classifies morphology.
  • Lateralisation did not reliably agree with the seizure side on Siena (9 of 13 on eeg-prep output, no better than always answering "left"; Validation c). Do not read a laterality index as a lateralising finding.
  • Rates per sleep stage inherit the detector's false positives (more artefact in wake) and eeg-sleep's staging (EEG-only). The increase of spiking with sleep depth is long described (Malow et al. 1998); the use of spike patterns around sleep for localisation (Conrad et al. 2023) is an intracranial result.
  • Timing and leaders. On a bipolar montage a discharge under one electrode appears on two pairs at the same time with opposite sign (phase reversal); the leading channel is then decided by amplitude. Leader and hemisphere are timing labels, not source localisation.
  • Janca's implementation notches 50 Hz internally; 60 Hz mains is notched by this app when the sidecar says 60 Hz (eeg-prep removes it anyway). Siena is low-passed at 30 Hz at source, which narrows Janca's 10–60 Hz band to 10–30 Hz.
  • Stretches shorter than 30 s between excluded intervals are not analysed (Janca needs ≥ 20 s).
  • vEpiSet validation used 30 of 84 recordings from one centre; no full scalp benchmark with a permissive licence exists for classical detectors.
Licences

Apache-2.0 for this app (LICENSE). The image bundles epycom (Apache-2.0), MNE-Python (BSD-3-Clause) and their dependencies (THIRD_PARTY.md); no model weights. The demo and validation data are not in the image or the repository.

Citations

Please cite these works if you use this app. Verified on Crossref/DataCite and PubMed; full author lists in CITATION.cff.

Spike detectors (both validated on intracranial EEG):

Janca R, Jezdik P, Cmejla R, Tomasek M, Worrell GA, Stead M, Wagenaar J, Jefferys JGR, Krsek P, Komarek V, Jiruska P, Marusic P. Detection of Interictal Epileptiform Discharges Using Signal Envelope Distribution Modelling: Application to Epileptic and Non-Epileptic Intracranial Recordings. Brain Topography 28(1):172–183 (2015). doi:10.1007/s10548-014-0379-1, PMID 24970691.

Barkmeier DT, Shah AK, Flanagan D, Atkinson MD, Agarwal R, Fuerst DR, Jafari-Khouzani K, Loeb JA. High inter-reviewer variability of spike detection on intracranial EEG addressed by an automated multi-channel algorithm. Clinical Neurophysiology 123(6):1088–1095 (2012). doi:10.1016/j.clinph.2011.09.023, PMID 22033028.

Implementation: epycom 0.3, Biomedical Engineering, FNUSA-ICRC Brno, https://github.com/ICRC-BME/epycom (Apache-2.0).

Reading the recordings:

Gramfort A, Luessi M, Larson E, Engemann DA, Strohmeier D, Brodbeck C, Goj R, Jas M, Brooks T, Parkkonen L, Hämäläinen M. MEG and EEG data analysis with MNE-Python. Frontiers in Neuroscience 7:267 (2013). doi:10.3389/fnins.2013.00267, PMID 24431986.

Sleep and spikes:

Malow BA, Lin X, Kushwaha R, Aldrich MS. Interictal Spiking Increases with Sleep Depth in Temporal Lobe Epilepsy. Epilepsia 39(12):1309–1316 (1998). doi:10.1111/j.1528-1157.1998.tb01329.x, PMID 9860066.

Conrad EC, Revell AY, Greenblatt AS, et al. Spike patterns surrounding sleep and seizures localize the seizure-onset zone in focal epilepsy. Epilepsia 64(3):754–768 (2023). doi:10.1111/epi.17482, PMID 36484572. (Intracranial EEG.)

Validation and demo data:

Lin N, Zheng M, Li L, et al. An EEG dataset for interictal epileptiform discharge with spatial distribution information. Scientific Data 12:229 (2025). doi:10.1038/s41597-025-04572-1, PMID 39920162. Data: figshare, doi:10.6084/m9.figshare.28069568.v2 (CC BY 4.0).

Detti P. Siena Scalp EEG Database (v1.0.0). PhysioNet (2020). doi:10.13026/5d4a-j060. And Detti P, Vatti G, Zabalo Manrique de Lara G. EEG Synchronization Analysis for Seizure Prediction: A Study on Data of Noninvasive Recordings. Processes 8(7):846 (2020). doi:10.3390/pr8070846. BIDS conversion: Dan J, Detti P. BIDS Siena Scalp EEG Database. EPFL, Zenodo (2024). doi:10.5281/zenodo.10640762 (CC BY 4.0).

Shoeb A. CHB-MIT Scalp EEG Database. PhysioNet (2010). doi:10.13026/C2K01R (ODC-By 1.0). And Shoeb AH. Application of machine learning to epileptic seizure onset detection and treatment. PhD thesis, MIT (2009). https://dspace.mit.edu/handle/1721.1/54669.

Cserpan D, Boran E, Rosch R, Lo Biundo SP, Ramantani G, Sarnthein J. Dataset of EEG recordings of pediatric patients with epilepsy based on the 10-20 system. OpenNeuro (2021). doi:10.18112/openneuro.ds003555.v1.0.1 (CC0). And Cserpan D et al. Scalp high-frequency oscillation rates are higher in younger children. Brain Communications 3(2):fcab052 (2021). doi:10.1093/braincomms/fcab052, PMID 33870193.

Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PCh, Mark RG, Mietus JE, Moody GB, Peng C-K, Stanley HE. PhysioBank, PhysioToolkit, and PhysioNet. Circulation 101(23):e215 (2000). doi:10.1161/01.CIR.101.23.e215, PMID 10851218.

Acknowledgements

epycom is developed by the Biomedical Engineering Department of the International Clinical Research Center, St. Anne's University Hospital Brno. MNE-Python is developed by the MNE community. The Siena Scalp EEG Database and CHB-MIT are distributed by PhysioNet, which asks that its databases be cited with Goldberger et al. 2000. vEpiSet was collected at the Epilepsy Center of Peking Union Medical College Hospital (funding listed in Lin et al. 2025).

Files

Root Directory

1 - 0 of 0 files

No files found.

About this repository

Publishing history

November 13, 2026
Originally Published
November 13, 2026 (Version 2)
Last Updated

Cite this dataset

References

References

Janca, R., Jezdik, P., Cmejla, R., Tomasek, M., Worrell, G. A., Stead, M., Wagenaar, J., Jefferys, J. G. R., Krsek, P., Komarek, V., Jiruska, P., & Marusic, P. (2014). Detection of Interictal Epileptiform Discharges Using Signal Envelope Distribution Modelling: Application to Epileptic and Non-Epileptic Intracranial Recordings. Brain Topography, 28(1), 172–183. https://doi.org/10.1007/s10548-014-0379-1

Barkmeier, D. T., Shah, A. K., Flanagan, D., Atkinson, M. D., Agarwal, R., Fuerst, D. R., Jafari-Khouzani, K., & Loeb, J. A. (2012). High inter-reviewer variability of spike detection on intracranial EEG addressed by an automated multi-channel algorithm. Clinical Neurophysiology, 123(6), 1088–1095. https://doi.org/10.1016/j.clinph.2011.09.023

Gramfort, A. (2013). MEG and EEG data analysis with MNE-Python. Frontiers in Neuroscience, 7. https://doi.org/10.3389/fnins.2013.00267

Malow, B. A., Lin, X., Kushwaha, R., & Aldrich, M. S. (1998). Interictal Spiking Increases with Sleep Depth in Temporal Lobe Epilepsy. Epilepsia, 39(12), 1309–1316. Portico. https://doi.org/10.1111/j.1528-1157.1998.tb01329.x

Conrad, E. C., Revell, A. Y., Greenblatt, A. S., Gallagher, R. S., Pattnaik, A. R., Hartmann, N., Gugger, J. J., Shinohara, R. T., Litt, B., Marsh, E. D., & Davis, K. A. (2023). Spike patterns surrounding sleep and seizures localize the seizure‐onset zone in focal epilepsy. Epilepsia, 64(3), 754–768. Portico. https://doi.org/10.1111/epi.17482

Lin, N., zheng, M., Li, L., Hu, P., Gao, W., Sun, H., Xu, C., Yuan, G., Liang, Z., Dong, Y., He, H., Cui, L., & Lu, Q. (2025). An EEG dataset for interictal epileptiform discharge with spatial distribution information [Dataset]. figshare. https://doi.org/10.6084/M9.FIGSHARE.28069568.V2

Detti, P. (2020). Siena Scalp EEG Database (Version 1.0.0) [Dataset]. PhysioNet. https://doi.org/10.13026/5D4A-J060

Dan, J., & Detti, P. (2024). BIDS Siena Scalp EEG Database (Version v1.0.0) [Dataset]. EPFL. https://doi.org/10.5281/ZENODO.10640762

Shoeb, A. (2010). CHB-MIT Scalp EEG Database [Dataset]. physionet.org. https://doi.org/10.13026/C2K01R

Goldberger, A. L., Amaral, L. A. N., Glass, L., Hausdorff, J. M., Ivanov, P. Ch., Mark, R. G., Mietus, J. E., Moody, G. B., Peng, C.-K., & Stanley, H. E. (2000). PhysioBank, PhysioToolkit, and PhysioNet. Circulation, 101(23). https://doi.org/10.1161/01.cir.101.23.e215

Dorottya Cserpan, Ece Boran, Richard Rosch, San Pietro Lo Biundo, Georgia Ramantani, & Johannes Sarnthein. (2021). Dataset of EEG recordings of pediatric patients with epilepsy based on the 10-20 system [Dataset]. OpenNeuro. https://doi.org/10.18112/OPENNEURO.DS003555.V1.0.1