EEG-Based Seizure Detection Using Machine Learning

Seizure vs. non-seizure classification on the CHB-MIT scalp EEG dataset (subject chb01).
Pipeline: 4s windowing → bandpower features (Welch PSD) → Logistic Regression vs. Linear SVM, evaluated with time-ordered splits.

Key Takeaways

• Linear SVM achieved better overall performance on the held-out future test set (Macro-F1 = 0.9012) with fewer false positives.
• Because seizure windows are rare, we report Macro-F1 in addition to accuracy to reflect performance on the minority class.

---

Results (held-out test set)

• Logistic Regression: Accuracy 0.9833, Macro-F1 0.8442
• Linear SVM: Accuracy 0.9908, Macro-F1 0.9012

Visual Results

Confusion Matrices (Test Set)

Test Performance Comparison (Accuracy + Macro-F1)

---

Dataset

• Dataset: CHB-MIT Scalp EEG Database (subject chb01)
• Source: Kaggle mirror of CHB-MIT pediatric EEG (see References)
• Data in this repo: NOT included (EDF files are large). Please download separately.
• Sampling rate: 256 Hz
• Channels available: 23 EEG derivations per EDF file
• Channel used (baseline): F7–T7 (Channel 2) to reduce computation (engineering simplification)
• Files used (7 seizure-containing EDFs):
  chb01_03, chb01_04, chb01_15, chb01_16, chb01_18, chb01_21, chb01_26

---

Problem Setup

Sample unit: one 4-second EEG window (from the selected channel)
Label unit: one binary label per window

• 1 = seizure (window overlaps an annotated seizure interval)
• 0 = non-seizure

---

Feature Extraction (Bandpower)

For each 4-second window, we compute Welch Power Spectral Density (PSD) and convert the raw waveform into a small feature vector.

• Features: band powers in standard EEG bands: delta, theta, alpha, beta
• Feature dimension: 4 features per window
• Implementation: see the notebook for details

In short: PSD → average/integrate power within each frequency band → features.

---

Models

• Logistic Regression (L2)
• Linear SVM (LinearSVC)

Both models use:

• StandardScaler for feature standardization
• TimeSeriesSplit (walk-forward CV) for hyperparameter tuning on the training portion only

---

Evaluation (No Leakage)

• Train/Test split: time-ordered 80/20 (future is never used to predict the past)
• Cross-validation: walk-forward TimeSeriesSplit on training data
• Metrics: Accuracy, Macro-F1, Confusion Matrix

---

How to Reproduce

1. Download the CHB-MIT dataset (EDF + summary file) from Kaggle/PhysioNet.

2. Create a folder named chb01/ at the repo root, and put the EDF files inside:

   chb01/chb01_03.edf
   chb01/chb01_04.edf
   chb01/chb01_15.edf
   chb01/chb01_16.edf
   chb01/chb01_18.edf
   chb01/chb01_21.edf
   chb01/chb01_26.edf

3. Install dependencies:

   • pip install -r requirements.txt

4. Open EEG_SeizureDetection.ipynb

5. Set DATA_DIR (default is Path("chb01")), then run all cells.

---

Repository Structure

• EEG_SeizureDetection.ipynb — main notebook (full pipeline)
• assets/ — figures (confusion matrices, comparison plot)

---

Limitations

• Single subject only — trained and evaluated on chb01; results may not generalize to other patients without retraining.
• Single channel — uses only F7–T7 (Channel 2). Multi-channel analysis would likely improve performance.
• Simple features — only bandpower in four standard EEG bands. Does not include gamma band, spectral ratios, nonlinear features (e.g., sample entropy), or time-domain statistics beyond mean/std.
• Linear models only — no nonlinear classifiers (Random Forest, XGBoost, CNN/LSTM) were explored.
• No class imbalance handling — seizure windows are ~2% of the data, but no oversampling, undersampling, or class_weight adjustment was applied.
• No EEG preprocessing — no bandpass filtering or artifact rejection prior to feature extraction.
• Small positive sample size — only ~115 seizure windows total, which limits statistical reliability.

---

References

• CHB-MIT Scalp EEG Database (PhysioNet)
• Kaggle mirror used in this project
• Scikit-learn documentation (LogisticRegression, LinearSVC, TimeSeriesSplit)
• Welch PSD method (SciPy)
• Standard EEG frequency bands (background)