JayF14/chb-mit-preprocessed-eeg

CHB-MIT Preprocessed EEG Dataset

Feature-Engineered EEG Data for Epileptic Seizure Detection

347,075 labeled epochs | 24 subjects | 590+ features per epoch | ~3.9 GB

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

---

---

Overview

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

The original CHB-MIT database contains continuous scalp EEG recordings from 24 pediatric patients 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 dataset was produced as part of the Neuro_Pulse project — an end-to-end seizure detection system combining classical ML, deep learning, and a real-time monitoring interface.

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
• Clinically labeled — ground truth seizure annotations from board-certified epileptologists
• Full coverage — all 24 CHB-MIT subjects, all EDF recordings processed
• Validated — zero NaN/Inf values, 100% file source traceability to original EDF files

---

Dataset at a Glance

Property	Value
Subjects	24 pediatric patients (chb01 through chb24)
Total Epochs	347,075
Seizure Epochs	22,312 (6.43%)
Non-Seizure Epochs	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 (seizure vs. non-seizure)
Label Column	target_label — 0 = non-seizure, 1 = seizure

---

Per-Subject Statistics

Subject	File Size	Epochs	Seizure	Non-Seizure	Seizure %	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, use pandas.concat with join="outer" or filter to the shared 595-column subset.

---

Column Schema

Each CSV file begins with 5 metadata columns, followed by 590+ feature columns:

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 = non-seizure, 1 = seizure

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

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"Seizure 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

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

# Load and combine all subjects
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}"
        dfs.append(df)

full_df = pd.concat(dfs, ignore_index=True, join="inner")  # inner join to keep shared columns
print(f"Combined shape: {full_df.shape}")
print(f"Seizure distribution:\n{full_df['target_label'].value_counts()}")

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

Train a seizure detection classifier in under 20 lines:

from huggingface_hub import hf_hub_download
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report

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

# 2. Prepare
META = ["epoch_id", "window_start", "window_end", "file_source", "target_label"]
X, y = df.drop(columns=META), df["target_label"]
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, stratify=y, random_state=42)

# 3. Train
clf = RandomForestClassifier(n_estimators=200, class_weight="balanced", n_jobs=-1, random_state=42)
clf.fit(X_train, y_train)

# 4. Evaluate
print(classification_report(y_test, clf.predict(X_test), target_names=["Non-Seizure", "Seizure"]))

---

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)
         -->  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

Class Imbalance

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

Strategy	Implementation
Class weighting	class_weight="balanced" in sklearn classifiers
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

Variable Column Counts

Four subjects (chb11, chb12, chb15, chb18) have more than 595 columns because their original EEG recordings include additional channels. When combining subjects:

# Option A: Keep only shared columns (safe)
full_df = pd.concat(dfs, join="inner", ignore_index=True)

# Option B: Keep all columns, fill missing with NaN
full_df = pd.concat(dfs, join="outer", ignore_index=True)

Leave-One-Subject-Out (LOSO) Evaluation

For clinically meaningful results, evaluate using LOSO cross-validation rather than random splits. The file_source and subject_id columns enable proper patient-level partitioning.

---

Citation

If you use this dataset in your research, please cite the original CHB-MIT database and PhysioNet:

@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}
}

---

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
MNE-Python (EEG toolkit)	mne.tools
imbalanced-learn	imbalanced-learn.org

---

License

This dataset is derived from the CHB-MIT Scalp EEG Database, distributed by PhysioNet under the PhysioNet Restricted Health Data License 1.5.0. By downloading or using this dataset, you agree to comply with the terms of that license. Please consult the original license before use in commercial applications.

---

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