Skip to content

CHBMIT_Seizure

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

CHBMIT_Seizure

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
original_scripts
original_scripts
 
 
scripts
scripts
 
 
CHBMIT_DATASET_INFO.md
CHBMIT_DATASET_INFO.md
 
 
README.md
README.md
 
 

README.md

CHB-MIT: Seizure Detection

Binary seizure detection on CHB-MIT Scalp EEG Database using DIVER preprocessing pipeline.

Quick Info

Item Value
Dataset CHB-MIT Scalp EEG Database
Subjects 21 pediatric epilepsy patients
Task Binary seizure detection
Channels 16 bipolar channels (Double Banana montage)
Sampling Rate 256 Hz → 500 Hz (resampled)
Segment Length 10 seconds
Output Shape (16, 10, 500)
Data Format v2-keymodified (ISRUC-compatible)

Documentation

Quick Start

1. Install Dependencies

pip install numpy scipy mne lmdb pyedflib

2. Run Preprocessing (v2 Recommended)

cd scripts
python preprocessing_chbmit_v2-keymodified.py \
  --data_path /path/to/CHB-MIT/dataset \
  --lmdb_path /path/to/output/CHBMIT_Seizure_v2 \
  --elc_file standard_1005.elc

3. Migrate Existing v1 Data (Optional)

If you have existing v1 LMDB, convert to v2 format:

python migrate_chbmit_v1_to_v2.py \
  --input_lmdb /path/to/CHBMIT_Seizure \
  --output_lmdb /path/to/CHBMIT_Seizure_v2

4. Validate Output

python check_lmdb_chbmit_v2-compatible.py --lmdb_path /path/to/CHBMIT_Seizure_v2

📦 Data Format

v2-keymodified (Current, ISRUC-Compatible) ✅

File: preprocessing_chbmit_v2-keymodified.py

{
    "sample": np.array,           # (16, 10, 500) - Changed from 'signal'
    "label": int,                 # 0=non-seizure, 1=seizure
    "data_info": {                # Changed from 'metadata', unified format
        # ISRUC-compatible fields
        "Dataset": "CHBMIT-Seizure",
        "modality": "EEG",
        "release": "1.0.0",
        "subject_id": str,        # e.g., "chb01"
        "task": "seizure-detection",
        "resampling_rate": 500,
        "original_sampling_rate": 256,
        "segment_index": int,
        "start_time": float,
        "channel_names": list,    # 16 bipolar channel names
        "electrode_pairs": dict,  # Bipolar electrode pairs
        "electrode_positions": dict,  # 3D coordinates
        "xyz_id": np.ndarray,     # (16, 3) positions ordered by channel_names

        # CHBMIT-specific fields
        "segment_id": str,        # e.g., "chb01_01_0"
        "split": str,             # "train", "val", "test"
        "is_oversampled": bool    # Seizure oversampling flag
    }
}

Why v2?

  • Unified data dictionary format across ISRUC, CHBMIT, ADFTD
  • Electrode information merged into data_info (no separate elc_info)
  • Consistent key names: sample (not signal), data_info (not metadata)
  • Better dataset interoperability and model training compatibility

v1 (Legacy, for Reference Only)

File: preprocessing_chbmit.py (kept for backward compatibility)

Click to see v1 format (not recommended for new projects) 
{
    "signal": np.array,           # (16, 10, 500)
    "label": int,
    "elc_info": dict,             # Separate electrode info
    "metadata": {
        "segment_id": str,
        "split": str,
        "is_oversampled": bool,
        "original_index": int,
        "original_sr": 256,
        "target_sr": 500
    }
}

Issues with v1:

  • Inconsistent key names vs ISRUC format
  • Electrode info separated from metadata
  • Missing dataset-level metadata (Dataset, modality, task)
  • Harder to maintain unified data loaders

📁 Directory Structure

CHBMIT_Seizure/
├── README.md                                    # This file - unified documentation
├── CHBMIT_DATASET_INFO.md                      # Detailed dataset info
│
├── scripts/                                     # Preprocessing and validation
│   ├── preprocessing_chbmit.py                      # v1 (legacy)
│   ├── preprocessing_chbmit_v2-keymodified.py       # v2 (recommended) ✅
│   ├── check_lmdb_chbmit.py                         # v1 validator
│   ├── check_lmdb_chbmit_v2-compatible.py           # v1/v2 validator ✅
│   ├── migrate_chbmit_v1_to_v2.py                   # v1→v2 migration ✅
│   ├── run_preprocessing_chbmit.sh                  # Runner script
│   └── standard_1005.elc                            # Electrode locations
│
├── logs/                                        # Processing logs
│   └── run_20251121_190529.log
│
└── lmdb_output/                                # LMDB databases (not in repo)
    ├── CHBMIT_Seizure/                              # v1 data
    └── CHBMIT_Seizure_v2/                           # v2 data ✅

Script Descriptions

Script Version Description
preprocessing_chbmit_v2-keymodified.py v2 ✅ Main preprocessing (ISRUC-compatible format)
check_lmdb_chbmit_v2-compatible.py v2 ✅ Validator (supports v1 & v2, shows format version)
migrate_chbmit_v1_to_v2.py Migration ✅ Convert existing v1 LMDB to v2 format
add_xyz_id_to_lmdb.py Utility ✅ Add xyz_id field to existing v2 LMDB
validate_xyz_id.py Utility ✅ Validate xyz_id alignment with channel_names
preprocessing_chbmit.py v1 Legacy preprocessing (kept for reference)
check_lmdb_chbmit.py v1 Legacy validator

🔄 Preprocessing Pipeline

Raw CHB-MIT Data (.edf + -summary.txt)
    ↓
[Parse Summary Files]  ← Extract seizure timestamps
    ↓
[Load EDF & Extract Channels]  ← 16 bipolar channels
    ↓
[Segment - Regular]  ← 10-second segments, 10-second step
    ↓
[Segment - Oversampled]  ← Seizure ±1s, 5-second step
    ↓
[Merge & Label]  ← 0=non-seizure, 1=seizure
    ↓
[Resample]  ← 256 Hz → 500 Hz
    ↓
[Reshape]  ← (16, 2560) → (16, 10, 500)
    ↓
[Add Metadata]  ← Dataset info, electrode positions
    ↓
[Store in LMDB]  ← v2 format with data_info
    ↓
CHBMIT_Seizure_v2/

📊 Dataset Statistics

Data Split

  • Train: chb01-20 (17 patients, excluding chb12, chb13, chb17)
  • Validation: chb21-22 (2 patients)
  • Test: chb23-24 (2 patients)

Segments per Split

Total Samples: 327,834
├── Train:  287,341 segments (87.6%)
├── Val:     23,065 segments  (7.0%)
└── Test:    17,428 segments  (5.3%)

Class Distribution

Before Oversampling:

  • Non-seizure : Seizure ≈ 99:1 (severe imbalance)

After Oversampling:

  • Non-seizure : Seizure ≈ 30:1 (improved, but still use class weights!)
  • Seizure segments: 5-second sliding window within seizure ±1s

Storage Size

  • LMDB database: ~99 GB
  • Per segment: ~300 KB (signal + metadata)

🧠 Bipolar Channel Configuration

16 channels following Double Banana montage (10-20 system):

Left Lateral Chain:
├── FP1-F7  (Frontopolar to Frontal left)
├── F7-T7   (Frontal to Temporal left)
├── T7-P7   (Temporal to Parietal left)
└── P7-O1   (Parietal to Occipital left)

Right Lateral Chain:
├── FP2-F8  (Frontopolar to Frontal right)
├── F8-T8   (Frontal to Temporal right)
├── T8-P8   (Temporal to Parietal right)
└── P8-O2   (Parietal to Occipital right)

Left Parasagittal Chain:
├── FP1-F3  (Frontopolar to Frontal left)
├── F3-C3   (Frontal to Central left)
├── C3-P3   (Central to Parietal left)
└── P3-O1   (Parietal to Occipital left)

Right Parasagittal Chain:
├── FP2-F4  (Frontopolar to Frontal right)
├── F4-C4   (Frontal to Central right)
├── C4-P4   (Central to Parietal right)
└── P4-O2   (Parietal to Occipital right)

Individual Electrodes Used:

  • FP1, FP2, F3, F4, F7, F8
  • C3, C4, T7, T8, P3, P4
  • P7, P8, O1, O2

🔁 Oversampling Strategy

Regular Segments (All data)

  • Step: 10 seconds (non-overlapping)
  • Purpose: Standard sampling

Oversampled Segments (Seizure only)

  • Base segments: Same as regular
  • Additional segments: 5-second step within seizure ±1 second
  • Effective rate: ~3× oversampling for seizure periods
  • Flag: is_oversampled: True in metadata

Example

Seizure at 100-110 seconds:

Regular:
├── [90-100]:  label=0  (non-seizure)
├── [100-110]: label=1  (seizure) ← Base
└── [110-120]: label=0  (non-seizure)

Oversampled (99-111s range, 5s step):
├── [99-109]:  label=1, is_oversampled=True
├── [104-114]: label=1, is_oversampled=True
└── [109-119]: label=1, is_oversampled=True

→ Total: 4 seizure segments (1 base + 3 oversampled)

💡 Usage Notes

  1. Use v2 format for new projects (ISRUC-compatible)
  2. Migrate v1 data using migrate_chbmit_v1_to_v2.py
  3. Apply class weights even after oversampling (still 30:1 imbalance)
  4. Consider patient-level splitting to avoid data leakage
  5. Check logs/ for detailed preprocessing statistics

📖 Citation

@inproceedings{shoeb2009chbmit,
  title={Application of Machine Learning to Epileptic Seizure Detection},
  author={Shoeb, A.},
  booktitle={Proceedings of the 26th International Conference on Machine Learning},
  year={2009}
}

🔗 Links

📌 Version Information

  • Data Format: v2-keymodified (ISRUC-compatible)
  • Last Updated: 2025-11-27
  • Migration Available: Yes (v1 → v2)
  • Recommended: Use v2 for all new projects