illusion-brain/training-data/benchmarks/seizure-detection - Storage Bucket

CHB-MIT Preprocessed EEG Dataset

Feature-Engineered EEG Data for Seizure Forecasting (Pre-Ictal Detection)

347,075 labeled epochs | 24 cases / 22 subjects | 590 features per epoch | ~3.9 GB

No raw signal processing needed. Download, load, and train.

Research use only. Not for clinical decisions. Code: https://github.com/JayD108/Neuro_Pulse

---

Table of Contents

• Overview
• Dataset at a Glance
• Per-Case Statistics
• Column Schema
• Feature Reference
• Download and Usage
• Quick-Start Example
• Preprocessing Pipeline
• Important Notes
• Citation
• Related Resources
• License / Data Provenance

---

Overview

This dataset provides ready-to-use, feature-engineered EEG data derived from the CHB-MIT Scalp EEG Database v1.0.0 — one of the most widely cited benchmarks in epileptic seizure forecasting research.

The original CHB-MIT database contains continuous scalp EEG recordings from 22 pediatric subjects (organized into 24 cases) with medically intractable epilepsy, collected at Boston Children's Hospital (Shoeb, 2009). These raw EDF recordings have been fully preprocessed through a multi-stage feature extraction pipeline, producing a tabular dataset that is immediately usable for training classical ML and deep learning models.

This repository contains a derived/feature-engineered dataset, not the original EEG recordings. The original CHB-MIT EEG recordings are distributed by PhysioNet under the Open Data Commons Attribution License v1.0. See the License / Data Provenance section for details.

This dataset was produced as part of the Neuro_Pulse project — an end-to-end seizure forecasting system combining classical ML, deep learning, and a real-time monitoring interface. The binary task is interictal vs. pre-ictal classification: predicting that a seizure will begin within the next 30 minutes.

What makes this dataset useful?

• No signal processing required — features are pre-computed from raw EEG
• Comprehensive feature set — 590 features spanning frequency-domain, time-domain, wavelet, non-linear, and cross-channel coherence measures
• Seizure annotations — seizure onset/offset annotations provided with the CHB-MIT database
• Full coverage — all 24 CHB-MIT cases processed
• Traceable — 100% file source traceability to original EDF files. Note: some features contain NaN or ±Inf (e.g. spectral ratios with near-zero denominators, or channels absent from a recording). Replace Inf with NaN and use a model that handles missing values natively, or impute before use.

---

Dataset at a Glance

Property	Value
Cases / Subjects	24 cases from 22 pediatric subjects (chb01–chb24; note: chb21 is an additional case from the same subject as chb01)
Total Epochs	347,075
Pre-ictal windows (label 1)	22,312 (6.43%)
Interictal windows (label 0)	324,763 (93.57%)
Columns (standard)	595 (5 metadata + 590 features)
Epoch Length	10 seconds, non-overlapping
EEG Channels	18 bipolar pairs (International 10-20 system)
Bandpass Filter	0.5 – 50 Hz
Notch Filter	60 Hz (US mains)
Sampling Rate	256 Hz (native CHB-MIT)
Total Size	~3.9 GB (24 CSV files)
Task	Binary classification: interictal vs. pre-ictal (seizure forecasting)
Label Column	target_label — 0 = interictal (≥ 30 min from any seizure), 1 = pre-ictal (within 30 min before seizure onset), 2 = ictal or post-ictal (during a seizure or within 5 min after; excluded from training, so absent from the CSVs)

---

Per-Case Statistics

Case	File Size	Epochs	Pre-ictal	Interictal	Pre-ictal %	EDF Files	Columns
chb01	160.8 MB	14,338	982	13,356	6.85%	42	595
chb02	141.4 MB	12,593	371	12,222	2.95%	36	595
chb03	150.9 MB	13,425	861	12,564	6.41%	38	595
chb04	629.4 MB	56,014	694	55,320	1.24%	42	595
chb05	155.4 MB	13,829	686	13,143	4.96%	39	595
chb06	266.0 MB	23,719	1,508	22,211	6.36%	18	595
chb07	269.3 MB	24,012	538	23,474	2.24%	19	595
chb08	78.1 MB	6,955	896	6,059	12.88%	20	595
chb09	272.1 MB	24,279	716	23,563	2.95%	19	595
chb10	200.2 MB	17,753	1,212	16,541	6.83%	25	595
chb11	139.9 MB	12,351	353	11,998	2.86%	35	719 *
chb12	82.2 MB	7,422	1,814	5,608	24.44%	24	1,091 *
chb13	127.4 MB	11,464	1,113	10,351	9.71%	33	595
chb14	101.2 MB	9,096	1,246	7,850	13.70%	26	595
chb15	152.2 MB	13,630	1,750	11,880	12.84%	40	626 *
chb16	72.5 MB	6,526	985	5,541	15.10%	19	595
chb17	82.9 MB	7,440	537	6,903	7.22%	21	595
chb18	143.4 MB	12,652	816	11,836	6.45%	36	719 *
chb19	118.9 MB	10,678	387	10,291	3.62%	30	595
chb20	108.0 MB	9,657	856	8,801	8.86%	29	595
chb21	129.8 MB	11,674	665	11,009	5.70%	33	595
chb22	102.5 MB	11,060	484	10,576	4.38%	31	595
chb23	104.4 MB	9,299	947	8,352	10.18%	9	595
chb24	80.9 MB	7,209	1,895	5,314	26.29%	22	595

* Subjects chb11, chb12, chb15, and chb18 have additional columns because the original CHB-MIT recordings for these patients include extra EEG channels beyond the standard 18-channel montage. Empty values in extra columns indicate features from channels not present in a given recording. When loading all subjects together, reindex every file to a fixed feature list (see Important Notes).

---

Column Schema

Each CSV file begins with 5 metadata columns, followed by 590 feature columns (or more for subjects with additional EEG channels):

Metadata Columns

#	Column	Type	Description
1	epoch_id	int	Unique sequential index for each time window within the file
2	window_start	float	Start time of the epoch in seconds from recording onset
3	window_end	float	End time of the epoch in seconds from recording onset
4	file_source	str	Source EDF filename from the original CHB-MIT dataset (e.g., chb01_03.edf)
5	target_label	int	Ground truth label: 0 = interictal (≥ 30 min from any seizure), 1 = pre-ictal (within 30 min before seizure onset), 2 = ictal/post-ictal (excluded; absent from these CSVs)

Feature Columns

Feature columns follow the naming convention: <Channel>_<FeatureName>

Example: FP1-F7_Delta_power, C3-P3_hjorth_activity, coh_FP1_F7_T7_P7_delta

---

Feature Reference

Feature Count Breakdown

The 590 features per standard epoch are composed as follows:

Feature Group	Count	Calculation
Per-channel signal features	414	23 features/channel × 18 channels
Spectral ratio features	144	8 ratios/channel × 18 channels
Cross-channel coherence	32	8 electrode pairs × 4 frequency bands
Total	590	414 + 144 + 32

The 23 per-channel signal features break down as:

Feature Type	Per Channel
Frequency-domain (Welch PSD)	10
Time-domain (Hjorth + RMS + ZC)	5
Wavelet decomposition (DWT)	6
Non-linear (entropy + fractal)	2
Subtotal	23

Note: The preprocessing pipeline diagram lists 23 features × 18 channels = 414 — this refers to the per-channel signal features only. Spectral ratio features (8 × 18 = 144) and cross-channel coherence (32) are computed separately and added to reach the total of 590.

1. Frequency-Domain Features (10 per channel)

Computed via Welch's method across five standard EEG sub-bands:

Band	Frequency Range
Delta	0.5 – 4 Hz
Theta	4 – 8 Hz
Alpha	8 – 13 Hz
Beta	13 – 30 Hz
Gamma	30 – 50 Hz

Feature Suffix	Description
*_power	Absolute band power (V^2/Hz, integrated via Simpson's rule)
*_relpow	Relative band power (fraction of total power in 0.5–50 Hz)

2. Time-Domain Features (5 per channel)

Feature Suffix	Description
*_hjorth_activity	Signal variance — Hjorth Activity parameter
*_hjorth_mobility	Normalized standard deviation of the first derivative
*_hjorth_complexity	Ratio of mobility of the derivative to mobility of the signal
*_rms	Root mean square amplitude
*_zero_crossings	Number of zero-axis crossings in the 10-second window

3. Wavelet Decomposition Features (6 per channel)

Discrete Wavelet Transform (Daubechies-4, 5-level decomposition):

Feature Suffix	Description
*_wt_A5_energy	Approximation coefficients energy (level 5)
*_wt_D5_energy	Detail coefficients energy (level 5, ~0.5–2 Hz content)
*_wt_D4_energy	Detail coefficients energy (level 4, ~2–4 Hz content)
*_wt_D3_energy	Detail coefficients energy (level 3, ~4–8 Hz content)
*_wt_D2_energy	Detail coefficients energy (level 2, ~8–16 Hz content)
*_wt_D1_energy	Detail coefficients energy (level 1, ~16–32 Hz content)

4. Non-Linear Features (2 per channel)

Feature Suffix	Description
*_sample_entropy	Regularity/complexity measure (higher = more irregular)
*_higuchi_fd	Higuchi Fractal Dimension (self-similarity, range ~1.0–2.0)

5. Spectral Ratio Features (8 per channel)

Feature Suffix	Description
*_Delta_relpow	Relative delta power
*_Theta_relpow	Relative theta power
*_Alpha_relpow	Relative alpha power
*_Beta_relpow	Relative beta power
*_Gamma_relpow	Relative gamma power
*_theta_alpha_ratio	Theta/Alpha power ratio (drowsiness marker)
*_slow_fast_ratio	(Delta+Theta) / (Alpha+Beta) ratio
*_beta_alpha_ratio	Beta/Alpha power ratio (alertness marker)

6. Cross-Channel Coherence Features (32 total)

Spectral coherence computed between 8 clinically relevant electrode pairs across 4 frequency bands (Delta, Theta, Alpha, Beta):

Electrode Pairs:
  FP1-F7 <-> T7-P7    FP2-F8 <-> T8-P8    T7-P7 <-> P7-O1    T8-P8 <-> P8-O2
  FP1-F7 <-> F7-T7    FP2-F8 <-> F8-T8    FP1-F7 <-> FP2-F8  T7-P7 <-> T8-P8

EEG Channel Montage (18 Bipolar Pairs)

Left Temporal:   FP1-F7  F7-T7   T7-P7   P7-O1
Left Parasag.:   FP1-F3  F3-C3   C3-P3   P3-O1
Right Parasag.:  FP2-F4  F4-C4   C4-P4   P4-O2
Right Temporal:  FP2-F8  F8-T8   T8-P8   P8-O2
Midline:         FZ-CZ   CZ-PZ

---

Download and Usage

Option 1 — Hugging Face datasets library (Recommended)

from datasets import load_dataset

# Load a single subject
ds = load_dataset(
    "JayF14/chb-mit-preprocessed-eeg",
    data_files="chb01_labeled_features.csv"
)
df = ds["train"].to_pandas()

print(f"Shape: {df.shape}")
print(f"Pre-ictal ratio: {df['target_label'].mean():.2%}")

Option 2 — Download specific files with huggingface_hub

from huggingface_hub import hf_hub_download
import pandas as pd

path = hf_hub_download(
    repo_id="JayF14/chb-mit-preprocessed-eeg",
    filename="chb01_labeled_features.csv",
    repo_type="dataset"
)
df = pd.read_csv(path)

Option 3 — Download all 24 files at once

from huggingface_hub import snapshot_download
import pandas as pd
import os, numpy as np

local_dir = snapshot_download(
    repo_id="JayF14/chb-mit-preprocessed-eeg",
    repo_type="dataset",
    local_dir="./chb_mit_data"
)

META = ["epoch_id", "window_start", "window_end", "file_source", "target_label"]

# Build the reference feature list from chb01 (all 590 standard features)
chb01 = pd.read_csv(os.path.join(local_dir, "chb01_labeled_features.csv"), nrows=0)
feature_list = [c for c in chb01.columns if c not in META]

# Load all patients, reindex to the same 590 columns (absent columns become NaN)
dfs = []
for i in range(1, 25):
    fpath = os.path.join(local_dir, f"chb{i:02d}_labeled_features.csv")
    if os.path.exists(fpath):
        df = pd.read_csv(fpath)
        df["subject_id"] = f"chb{i:02d}"
        df = df.reindex(columns=META + ["subject_id"] + feature_list)
        dfs.append(df)

full_df = pd.concat(dfs, ignore_index=True)
print(f"Combined shape: {full_df.shape}")
print(f"Label distribution:\n{full_df['target_label'].value_counts()}")

Note: reindex is used (not join="inner") so all 590 standard columns are preserved. Patients chb12–chb22, which lack the 4 right-temporal channels, will have NaN in those 40 columns. join="inner" would silently drop those columns for all patients and produce only 550 features.

Option 4 — Direct download via CLI

# Single file
curl -L -o chb01_labeled_features.csv \
  "https://huggingface.co/datasets/JayF14/chb-mit-preprocessed-eeg/resolve/main/chb01_labeled_features.csv"

# All files (requires git-lfs)
git lfs install
git clone https://huggingface.co/datasets/JayF14/chb-mit-preprocessed-eeg

---

Quick-Start Example

⚠️ Evaluation warning: A random window-level split (e.g. train_test_split) leaks information between neighbouring windows from the same patient and yields inflated scores (often > 99% ROC-AUC). Always use a patient-wise split for meaningful evaluation. See the patient-wise example below.

The example below uses a leave-one-patient-out split for clinically meaningful evaluation. Tested on chb01, chb08, chb12 with n_splits=3 — runs without errors.

from huggingface_hub import snapshot_download
import pandas as pd, numpy as np
from sklearn.model_selection import GroupKFold
import xgboost as xgb
from sklearn.metrics import roc_auc_score
import os

# 1. Download and load all patients
local_dir = snapshot_download("JayF14/chb-mit-preprocessed-eeg", repo_type="dataset",
                               local_dir="./chb_mit_data")

META = ["epoch_id", "window_start", "window_end", "file_source", "target_label"]

# Build the reference feature list from chb01 (all 590 standard features)
chb01 = pd.read_csv(os.path.join(local_dir, "chb01_labeled_features.csv"), nrows=0)
feature_list = [c for c in chb01.columns if c not in META]

dfs = []
for i in range(1, 25):
    fpath = os.path.join(local_dir, f"chb{i:02d}_labeled_features.csv")
    if os.path.exists(fpath):
        df = pd.read_csv(fpath)
        df["patient"] = f"chb{i:02d}"
        # Reindex to fixed feature list — absent columns (e.g. right-temporal
        # channels for chb12-chb22) become NaN; XGBoost handles them natively.
        df = df.reindex(columns=META + ["patient"] + feature_list)
        dfs.append(df)

full = pd.concat(dfs, ignore_index=True)
# Keep only pre-ictal (1) and interictal (0) rows
full = full[full["target_label"].isin([0, 1])]

# 2. Prepare — handle Inf/NaN
X = full[feature_list].replace([np.inf, -np.inf], np.nan).values.astype(np.float32)
y = full["target_label"].values
groups = full["patient"].values

# 3. Patient-wise leave-one-out evaluation
gkf = GroupKFold(n_splits=len(full["patient"].unique()))
aucs = []
for train_idx, test_idx in gkf.split(X, y, groups=groups):
    X_tr, X_te = X[train_idx], X[test_idx]
    y_tr, y_te = y[train_idx], y[test_idx]
    spw = float((y_tr == 0).sum()) / max(float((y_tr == 1).sum()), 1)
    clf = xgb.XGBClassifier(n_estimators=500, max_depth=6, learning_rate=0.05,
                             subsample=0.8, colsample_bytree=0.8,
                             tree_method="hist", scale_pos_weight=spw, random_state=42)
    clf.fit(X_tr, y_tr)
    aucs.append(roc_auc_score(y_te, clf.predict_proba(X_te)[:, 1]))
    print(f"Patient {groups[test_idx[0]]}: ROC-AUC = {aucs[-1]:.4f}")

print(f"\nMean ROC-AUC: {np.mean(aucs):.4f}  Median: {np.median(aucs):.4f}")

Reference result (XGBoost, leave-one-patient-out, 24 patients, 500 trees, 590 features):

• Mean ROC-AUC: 0.56, Median: 0.59, Range: 0.26–0.71
• Full per-patient breakdown: apala_xgb_v1_lopo.json

---

Preprocessing Pipeline

The raw .edf recordings were processed through the following pipeline:

Raw EDF  -->  Load (MNE-Python)  -->  Bandpass 0.5-50 Hz  -->  Notch 60 Hz
         -->  Segment into 10s non-overlapping epochs
         -->  Per-epoch, per-channel feature extraction:
              |-- Welch PSD (5 bands x 2 = 10 features)
              |-- Hjorth parameters (3 features)
              |-- RMS + Zero crossings (2 features)
              |-- DWT energy, db4, 5-level (6 features)
              |-- Sample Entropy (1 feature)
              |-- Higuchi Fractal Dimension (1 feature)
              = 23 features x 18 channels = 414 features
         -->  Cross-channel coherence (8 pairs x 4 bands = 32 features)
         -->  Spectral ratios (8 ratios x 18 channels = 144 features)
         -->  Total per epoch: 414 + 32 + 144 = 590 features
         -->  Label assignment from CHB-MIT seizure annotations
         -->  Export as per-subject CSV

Pipeline Parameter	Value
Signal loader	MNE-Python
Bandpass filter	4th-order Butterworth, 0.5–50 Hz (SOS)
Notch filter	60 Hz (US power line)
Epoch window	10 seconds, non-overlapping
PSD method	Welch's method
Wavelet	Daubechies-4 (db4), 5-level decomposition
Entropy	Sample entropy (m=2)
Fractal dimension	Higuchi FD
Coherence	Scipy signal.coherence
Label source	CHB-MIT chbXX-summary.txt seizure annotations

---

Important Notes

This is a Derived Dataset

This repository contains feature-engineered representations computed from the original CHB-MIT EEG recordings. It does not contain or redistribute the original .edf files. The original CHB-MIT EEG recordings are distributed by PhysioNet under the Open Data Commons Attribution License v1.0.

Users who need the raw EEG signals should obtain them directly from: https://physionet.org/content/chbmit/1.0.0/

Label Definitions

Label	Meaning	Count
0	Interictal — epoch starts ≥ 30 min from any seizure onset	324,763
1	Pre-ictal — epoch starts within 30 min before a seizure onset	22,312
2	Ictal / post-ictal — epoch overlaps a seizure or starts within 5 min after seizure end	excluded (not in CSVs)

Epochs labelled 2 are excluded from the CSV files; the binary task is interictal vs. pre-ictal.

Data Quality — NaN and Inf Values

Some features contain NaN or ±Inf values. Common causes:

• Spectral ratios (e.g. *_theta_alpha_ratio) can produce ±Inf when the denominator band power is near zero.
• Missing channels: patients chb12–chb22 lack the four right-temporal bipolar channels FP2-F8, F8-T8, T8-P8 and P8-O2 in most recordings (40 of the 590 standard features are absent for these 11 patients). Verified from CSV headers.
• Non-linear features (sample entropy, Higuchi FD) may fail on near-flat segments.

Recommended handling:

import numpy as np
X = X.replace([np.inf, -np.inf], np.nan)
# Then either:
#   A) Use a model that handles NaN natively (e.g. XGBoost with tree_method="hist")
#   B) Impute: X = X.fillna(X.mean())

Montage Differences Across Patients

• Patients chb12–chb22 (verified from CSV headers) are missing the right-temporal bipolar channels FP2-F8, F8-T8, T8-P8 and P8-O2 (40 features out of 590). These columns are simply absent from those CSVs.
• Patients chb11, chb12, chb15, and chb18 have extra non-standard channels beyond the 18-pair montage (see Columns column in Per-Case Statistics above).

When combining patients, reindex every file to a fixed 590-feature list and treat absent columns as missing (NaN):

FEATURE_COLS = [...]  # your reference 590-column list (e.g. from chb01)
X = df.reindex(columns=FEATURE_COLS)   # absent columns become NaN automatically

Class Imbalance

This dataset exhibits severe class imbalance (6.43% pre-ictal overall). Some subjects have as little as 1.24% pre-ictal windows (chb04), while others reach 26.29% (chb24). Recommended strategies:

Strategy	Implementation
Class weighting	class_weight="balanced" in sklearn classifiers, or scale_pos_weight in XGBoost
Oversampling	SMOTE or ADASYN via imbalanced-learn
Undersampling	RandomUnderSampler from imbalanced-learn
Loss functions	Focal loss, weighted cross-entropy
Evaluation	Use F1, AUROC, AUPRC — not accuracy

Leave-One-Patient-Out (LOPO) Evaluation

⚠️ Random window splits inflate scores. Because neighbouring 10-second windows are highly correlated, a random train_test_split across all windows leaks information and yields artificially high metrics (often > 99% ROC-AUC). For a fair, clinically meaningful estimate of generalisation to new patients, use patient-level leave-one-out cross-validation.

For clinically meaningful evaluation, use patient-level leave-one-patient-out (LOPO) cross-validation rather than random splits. The case identifier in file_source can be mapped to the corresponding subject, with chb01 and chb21 grouped as the same subject.

Reference result: XGBoost (500 trees, 590 features), leave-one-patient-out across all 24 patients:

• Mean ROC-AUC: 0.56, Median: 0.59, Range: 0.26–0.71
• Full per-patient table: apala_xgb_v1_lopo.json

---

Citation

If you use this dataset in your research, please cite the original CHB-MIT database and PhysioNet. The official citation as specified by PhysioNet is:

@article{guttag2010chbmit,
  author    = {Guttag, John},
  title     = {{CHB-MIT Scalp EEG Database}},
  journal   = {{PhysioNet}},
  year      = {2010},
  month     = jun,
  note      = {Version 1.0.0},
  doi       = {10.13026/C2K01R},
  url       = {https://doi.org/10.13026/C2K01R}
}

@phdthesis{shoeb2009application,
  title     = {Application of Machine Learning to Epileptic Seizure Onset Detection and Treatment},
  author    = {Shoeb, Ali Hossam},
  year      = {2009},
  school    = {Massachusetts Institute of Technology},
  address   = {Cambridge, MA}
}

@article{goldberger2000physiobank,
  title     = {{PhysioBank, PhysioToolkit, and PhysioNet}: Components of a New Research Resource for Complex Physiologic Signals},
  author    = {Goldberger, Ary L and Amaral, Luis A N and Glass, Leon and Hausdorff, Jeffrey M and Ivanov, Plamen Ch and Mark, Roger G and Mietus, Joseph E and Moody, George B and Peng, Chung-Kang and Stanley, H Eugene},
  journal   = {Circulation},
  volume    = {101},
  number    = {23},
  pages     = {e215--e220},
  year      = {2000},
  publisher = {American Heart Association}
}

Original dataset DOI: https://doi.org/10.13026/C2K01R

---

Related Resources

Resource	Link
Original CHB-MIT Database	physionet.org/content/chbmit/1.0.0
Neuro_Pulse Project (source code)	github.com/JayD108/Neuro_Pulse
APALA Model (trained on all 24 patients)	JayF14/Neuro_Pulse
MNE-Python (EEG toolkit)	mne.tools
imbalanced-learn	imbalanced-learn.org

---

License / Data Provenance

Source Data

This dataset is a processed and feature-engineered derivative of the CHB-MIT Scalp EEG Database v1.0.0, originally collected at Boston Children's Hospital and distributed by PhysioNet.

• Original dataset: CHB-MIT Scalp EEG Database v1.0.0
• PhysioNet page: https://physionet.org/content/chbmit/1.0.0/
• DOI: 10.13026/C2K01R
• Source license: Open Data Commons Attribution License v1.0 (ODC-By 1.0)
  • License text: https://opendatacommons.org/licenses/by/1.0/

The original CHB-MIT database is made available by PhysioNet under the Open Data Commons Attribution License v1.0 (ODC-By 1.0), which permits creation and distribution of derivative databases provided that appropriate attribution and licensing notices are retained.

Nature of This Repository

The creator of this repository does not claim ownership of the underlying EEG recordings or the original CHB-MIT database. The original CHB-MIT EEG recordings are distributed by PhysioNet under the Open Data Commons Attribution License v1.0. This repository contains only feature-engineered representations — tabular numerical features extracted and derived from the original recordings through a signal-processing pipeline. The original .edf files are not included here.

For licensing and provenance purposes, this repository is treated as a processed/feature-engineered derivative database of the CHB-MIT dataset, as it is based upon the original recordings through extraction and transformation of a substantial part of their contents.

User Obligations

Users of this derivative dataset should:

1. Attribute the original CHB-MIT Scalp EEG Database and PhysioNet as the source of the underlying data.
2. Retain the applicable ODC-By 1.0 attribution and licensing notices when redistributing or publishing work that incorporates this dataset.
3. Consult the full ODC-By 1.0 license text at https://opendatacommons.org/licenses/by/1.0/ to understand their obligations.

Attribution Notice (ODC-By 1.0)

When using or redistributing this dataset, please include the following attribution:

This dataset is derived from the CHB-MIT Scalp EEG Database v1.0.0 (Guttag, 2010), distributed by PhysioNet (https://physionet.org/content/chbmit/1.0.0/) under the Open Data Commons Attribution License v1.0 (https://opendatacommons.org/licenses/by/1.0/). DOI: 10.13026/C2K01R.

---

Built with care by Mrittunjay Dubey as part of the Neuro_Pulse project.