IESSEEG

Code · Experiment guides · Benchmark results · Questions

IESSEEG is an EEG dataset for infantile epileptic spasms syndrome (IESS). It contains approximately 3,172 hours of original EEG from 50 cases and 50 controls, together with clinician-selected EEG segments, simulated routine EEG segments, clinical annotations, and participant metadata. The accompanying benchmark evaluates IESS diagnosis and treatment-response prediction before and after treatment.

This is the v1.1 release. EEG signals are provided as EDF files in an EEG-BIDS directory structure, with annotations and metadata in TSV and JSON files. The preview above shows the recording index, participant metadata, and BASED ratings. Its records split is a table of available data; the benchmark's five-fold assignments are provided separately in splits/.

Dataset contents

EEG subset	Files	Hours	Recordings
Original long EEG recordings	150	3,172.47	50 case pre-treatment, 50 case post-treatment, and 50 control recordings
Clinician-selected EEG segments	600	250.53	Two sleep and two awake segments from each original long recording
Simulated routine EEG segments	200	100.06	Two sampled 30-minute segments from each case pre-treatment and control recording

The segments are extracted from the original long recordings, so their hours are included in the 3,172.47-hour total. No video is included.

Read an EEG recording

You can inspect one recording without downloading the full dataset:

pip install huggingface_hub pandas mne

from pathlib import Path

import mne
import pandas as pd
from huggingface_hub import hf_hub_download

repo_id = "Capur/IESSEEG"
root = Path("./data/IESSEEG")
index_path = hf_hub_download(repo_id, "recordings.tsv", repo_type="dataset", local_dir=root)
recordings = pd.read_csv(index_path, sep="\t", na_values=["n/a"])
row = recordings.loc[recordings["subset"].eq("clinician_selected")].iloc[0]
eeg_path = hf_hub_download(repo_id, row["path"], repo_type="dataset", local_dir=root)

raw = mne.io.read_raw_edf(eeg_path, preload=False)
signal = raw.get_data(start=0, stop=round(10 * raw.info["sfreq"]))
print(row["recording_id"], row["sleep_awake_label"])
print(signal.shape, raw.info["sfreq"])  # channels × samples, sampling rate in Hz

This reads the first ten seconds. MNE returns signal amplitudes in volts. The downloaded EDF retains its original channels and sampling rate; the benchmark code applies each model's montage, filtering, resampling, and scaling.

Download the dataset

The complete release contains approximately 119 GB of EDF files.

hf download Capur/IESSEEG --repo-type dataset --local-dir /data/IESSEEG

For a subset, download the metadata together with the matching EEG files:

# Metadata and clinician-selected EEG segments
hf download Capur/IESSEEG --repo-type dataset --local-dir /data/IESSEEG \
  --include '*.tsv' '*.json' '*.csv' 'README.md' 'LICENSE' '*acq-selected*'

Use *acq-long* for original long recordings or *acq-routine* for simulated routine EEG segments. The complete benchmark expects all three subsets. Downloading one subset is sufficient for inspecting its signals and annotations.

Annotations and metadata

Annotation	Available for	Meaning
Patient diagnosis label	All recordings and segments	Whether the source participant is a case (1) or control (0)
Clinician diagnosis label	200 simulated routine EEG segments	Whether a clinician judged the segment consistent with IESS, without access to the patient diagnosis
Treatment-response label	50 cases, shared by their pre-treatment and post-treatment recordings	Sustained response assessed at approximately six weeks; 28 responders and 22 non-responders
Sleep/awake label	600 clinician-selected EEG segments	The segment's annotated state
BASED score	20 original long recordings	Seven clinicians' independent ratings, giving 140 ratings across 10 pre-treatment and 10 post-treatment recordings

Clinical metadata include age at EEG, age at onset, and lead time from disease onset to the first EEG. Field definitions and units are provided in participants.json and recordings.json. Missing values are stored as n/a; metadata availability differs between cases and controls.

The additional first-follow-up response field is distinct from the sustained response label used in the benchmark. BASED scores retain their clinician identifiers; representative interval coordinates and a consensus score are not provided.

Files and identifiers

File or directory	Contents
participants.tsv	Participant demographics and clinical metadata
recordings.tsv	One row per EDF, including its relative path, labels, and source recording
sub-*/ses-*/eeg/	EDF signals, channel tables, and EEG-BIDS JSON sidecars
annotations/	Diagnosis, response, sleep/awake, and BASED annotation tables
metadata/segment_mapping.tsv	Links from derived segments to their source long recordings
splits/	Patient fold assignments and ordered diagnosis recording assignments
sampling/	Segment coordinates used for response training and the Section 6 analysis

Join participant information using participant_id or patient_id, and recording information using recording_id. The integer patient_id is a stable benchmark identifier. source_recording_id links a segment to its original long EEG; for a long recording, it refers to that recording itself.

For example, attach diagnosis fold assignments to the recording index:

folds = pd.read_csv(root / "splits/diagnosis_folds.csv")
recordings_with_folds = recordings.merge(folds, on="patient_id", validate="many_to_one")

This example requires the metadata download above. Fold assignments must be joined by identifier, rather than matched by row position.

PRE and POST refer to the treatment evaluated in the source study. Cases are not uniformly treatment-naive. Control recordings use visit=CONTROL. Source recording codes are de-identified dataset identifiers.

Run the benchmark

The GitHub repository provides installation, model downloads, and commands for each experiment. Use the same downloaded dataset directory for all tasks.

Experiment	EEG inputs	Instructions
IESS diagnosis	Clinician-selected case pre-treatment and control segments for fitting; simulated routine EEG segments for testing	Table 3
Pre-treatment and post-treatment response prediction	Original case long recordings, evaluated separately at each visit	Table 4
Recording context and lead time	Sampled 30-minute embeddings, whole-recording qEEG, and clinical lead time	Table 5 and Appendix C.2

After installing the code, run these commands from the code repository:

export DATASET_ROOT=/data/IESSEEG
export BENCHMARK_WORK=/scratch/iesseeg
python -m iesseeg validate --data "$DATASET_ROOT" --work "$BENCHMARK_WORK"
python -m iesseeg prepare --data "$DATASET_ROOT" --work "$BENCHMARK_WORK"

prepare creates manifests and links to the downloaded EDF files. Experiment outputs are saved under BENCHMARK_WORK. The data directory is not modified. The evaluation guide explains the metrics and links to the reported results.

All recordings and segments from a participant remain in the same outer test fold. Both response tasks share the same response fold assignments and labels. Training and validation procedures are documented with the code.

Segment timing and sampling

Saved sampling tables contain the segment starts used in the experiments. Sampled segments may overlap. Their coordinates describe experimental inputs, not annotated clinical events. Exported simulated routine EEG segments contain 1,801 seconds of signal for a nominal 30-minute sampling interval.

The source-recording mappings for clinician-selected EEG segments do not include their within-recording start and end positions. Sleep/awake labels apply to the selected segments, not to timed intervals in the original long EEG.

Signal format and de-identification

Files retain their sampling rates, digital signal samples, physical calibration, and continuous/discontinuous recording structure. Standard EDF patient/recording identification fields are replaced with pseudonymous identifiers. Calendar dates and wall-clock start times are replaced with a fixed placeholder, and free-text EDF annotations and acquisition descriptions are removed. Relative timekeeping annotations needed to read EDF+D are preserved. Duplicate auxiliary POL RESP labels are disambiguated as POL RESP-0 and POL RESP-1 in the EDF header and channel table. Original labels are retained in the source_name column of affected channel tables. Scalp EEG channel labels and all signal samples are unchanged. Absolute recording timestamps must not be used as clinical data. Incomplete trailing bytes beyond complete EDF data records are excluded.

The model code performs its own resampling, montage construction, and filtering; the EDF signals have not been converted to one common model input. Missing acquisition information is marked n/a, rather than inferred. EEG locations are described by the source study as the international 10–20 system; subject- specific digitized electrode positions are not provided.

BIDS validation

The release passes BIDS Validator 3.0.2 with zero errors. bids_validation.json records the validator version and remaining warning categories. Warnings concern unavailable recommended acquisition metadata, anonymized author attribution, and absent event tables for clinician-selected sleep/awake segments. No timed clinical events are inferred from segment-level state labels. Benchmark-specific tables under metadata, annotations, splits, and sampling extend the BIDS EEG organization and are checked separately by the accompanying code.

deno run -ERWN jsr:@bids/validator@3.0.2 /path/to/IESSEEG

Research use

This retrospective single-institution cohort supports research on IESS diagnosis and treatment-response prediction. Controls are not fully age matched. Segments from the same participant are dependent observations. The benchmark is not a prospectively validated clinical diagnostic or treatment-selection system.

The retrospective study received institutional review board approval under protocol 17-000026, with waivers of informed consent and HIPAA research authorization. Do not attempt to re-identify participants.

License and citation

The dataset and annotations are released under CC BY 4.0. Benchmark code and pretrained model weights have their respective licenses; see code attribution.

Please cite IESSEEG: An Open EEG Dataset Towards Better Understanding of Infantile Epileptic Spasms Syndrome when using this release. The 400 case clinician-selected EEG segments were previously analyzed by Rajaraman et al. (2024), Computational EEG attributes predict response to therapy for epileptic spasms.

For questions about the data, open a dataset discussion. For code issues, use GitHub Issues.