🤖 machine learning

Epileptic Seizure Detection in Separate Frequency Bands Using Feature Analysis and Graph Convolutional Neural Network (GCN) from Electroencephalogram (EEG) Signals

This study proposes a frequency-aware framework that decomposes EEG signals into five bands, extracts discriminative features, and utilizes a Graph Convolutional Neural Network (GCN) to model spatial electrode dependencies, achieving high seizure detection accuracy and improved interpretability on the CHB-MIT dataset.

Original authors: Ferdaus Anam Jibon, Fazlul Hasan Siddiqui, F. Deeba, Gahangir Hossain

Published 2026-04-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ferdaus Anam Jibon, Fazlul Hasan Siddiqui, F. Deeba, Gahangir Hossain

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is a bustling, 86-billion-person city. Normally, the citizens (neurons) communicate in a calm, organized rhythm, sending messages back and forth to keep the city running smoothly. But sometimes, a "power surge" happens. A group of citizens starts shouting in unison, creating a chaotic storm of electricity. This is an epileptic seizure.

Doctors use a tool called an EEG (Electroencephalogram) to listen to this city. They place 23 "microphones" (electrodes) on the scalp to record the electrical chatter. However, listening to 23 microphones at once, all day long, is like trying to hear a single conversation in a crowded stadium. It's overwhelming, and human doctors often miss the warning signs or get tired.

This paper introduces a new, super-smart "detective" system to find these seizures automatically. Here is how it works, broken down into simple concepts:

1. Tuning the Radio (Frequency Bands)

The electrical chatter in the brain isn't just one sound; it's a mix of different "pitches" or frequencies. Think of it like a radio station.

Low Pitches (Delta & Theta): Slow, heavy waves, like a deep drumbeat.
Medium Pitches (Alpha & Beta): Faster, more active waves, like a conversation or a car engine.
High Pitches (Gamma): Very fast, sharp waves.

The researchers realized that seizures don't happen at just one pitch. Sometimes the "storm" starts in the low drums, sometimes in the medium conversation. So, instead of listening to the whole radio at once, they built five separate listening stations, each tuned to a specific frequency band. This allows them to hear the specific "voice" of a seizure that might be hidden in the noise of other frequencies.

2. The "Snapshot" Strategy (Ictal Phase)

Seizures have three acts:

Pre-ictal: The calm before the storm (warning signs).
Ictal: The storm itself (the actual seizure).
Post-ictal: The calm after the storm.

Most computer systems try to learn from the whole movie. This paper decided to focus only on the "storm" scenes (the ictal phase). By training the computer only on the moments when the seizure is definitely happening, the system learns exactly what a "real" seizure looks like, making it much harder to get confused by normal brain activity.

3. The "City Map" (Graph Convolutional Networks)

This is the most creative part. Traditional AI looks at data like a spreadsheet (rows and columns). But the brain isn't a spreadsheet; it's a network. The left side of your brain talks to the right side; the front talks to the back.

The researchers used a special type of AI called a Graph Convolutional Network (GCN).

The Metaphor: Imagine the 23 microphones on the scalp are cities on a map.
The Edges: The wires connecting them are roads.
The AI: Instead of just looking at the traffic in one city, the GCN looks at the entire map. It understands that if City A starts shouting, City B (which is connected by a road) might start shouting too.

This allows the AI to understand the spatial relationship of the seizure. It knows that a seizure isn't just a loud noise in one spot; it's a wave traveling along specific roads in the brain.

4. The Detective's Toolkit (Feature Extraction)

Before the AI can make a decision, the researchers gave it a checklist of 11 clues to look for in the sound waves. These aren't just "loud" or "quiet." They are complex clues like:

How "jagged" is the wave? (Kurtosis)
Is the wave symmetrical? (Skewness)
How much is the volume fluctuating? (Variance)

They took these 11 clues from every single microphone, creating a massive, detailed profile of the brain's activity.

The Results: A Super-Smart Detective

The team tested their system on a famous database of brain recordings (CHB-MIT). Here is what they found:

The "Sweet Spot": The system was incredibly good at detecting seizures in the Alpha and Lower Beta bands (the medium frequencies). It got it right 99.5% to 99.7% of the time! It's like having a detective who never misses a clue in a specific room.
The "Confused" Band: Interestingly, the system struggled with the Higher Beta band (22–30 Hz). It was very sensitive (it screamed "Seizure!" a lot) but often wrong (false alarms). This tells us that this specific high-pitch frequency might not be the best place to look for seizures, saving doctors time by ignoring that channel.
Overall Success: When combining all the bands, the system achieved 99% accuracy.

Why Does This Matter?

Think of this system as a smart security guard for the brain.

It's Faster: It can scan hours of brain waves in seconds.
It's Fairer: It doesn't get tired or distracted like a human doctor.
It's Insightful: By breaking the signal into frequency bands, it tells doctors where and how the seizure is happening, which helps in choosing the right treatment.

In short, this paper teaches a computer to listen to the brain's "radio stations" separately, map out the connections between the speakers, and spot the exact moment a seizure storm begins, all with near-perfect accuracy.

1. Problem Statement

Epileptic seizures are neurological disorders characterized by abnormal, excessive electrical discharges in the brain. While Electroencephalogram (EEG) signals are the gold standard for diagnosis, manual interpretation by physicians is time-consuming, labor-intensive, and prone to inter-observer variability.

Existing automated detection methods face several limitations:

Traditional Machine Learning: Relies on handcrafted features and often fails to capture complex spatial relationships between EEG channels (brain regions).
Standard Deep Learning (CNN/RNN): While effective at feature extraction, these models often assume a Euclidean grid structure. This fails to represent the non-Euclidean geometry of brain connectivity, where EEG electrodes are spatially distributed across lobes rather than arranged in a grid.
Broadband Analysis: Many approaches analyze the entire EEG spectrum (broadband) simultaneously, potentially obscuring frequency-specific seizure characteristics that occur in distinct bands (e.g., Alpha vs. Beta).
Interpretability: Deep learning models often act as "black boxes," lacking the neurophysiological insight required for clinical trust.

2. Methodology

The proposed framework addresses these issues by combining frequency-band-specific analysis, handcrafted feature engineering, and Graph Convolutional Networks (GCN) to model the brain's spatial topology.

A. Data Preprocessing

Dataset: The study utilizes the CHB-MIT Scalp EEG Dataset, containing recordings from 23 patients (23 channels) sampled at 256 Hz.
Frequency Decomposition: Raw EEG signals are decomposed into five distinct frequency bands using band-pass filtering:
- Delta (0.5–4 Hz)
- Theta (4–8 Hz)
- Alpha (8–13 Hz)
- Lower Beta (13–22 Hz)
- Higher Beta (22–30 Hz)
Segmentation: Long EEG files are segmented into 6-second windows.
Labeling: Only the ictal phase (seizure onset to offset) is used for training to ensure high precision. Non-seizure segments are extracted from inter-ictal periods.
Class Imbalance: The SMOTE (Synthetic Minority Over-sampling Technique) is applied to balance the dataset, as seizure events are significantly fewer than non-seizure events.

B. Feature Extraction

Instead of relying solely on raw data, the authors extract 11 discriminative features from the ictal phase for each of the 23 channels:

Power Spectral Density (PSD)
Hjorth Parameters (Activity, Mobility, Complexity)
Kurtosis
Skewness
Standard Deviation
Maximum Amplitude
Variance
Median
Mean

Input Representation: For a 6-second window, the input matrix size is $23 \text{ (channels)} \times 11 \text{ (features)} \times 6 \text{ (seconds)} = 1,518$ features.

C. Graph Construction

To capture spatial dependencies, the EEG channels are modeled as a graph $G = (V, E)$ :

Nodes ( $V$ ): The 23 EEG electrodes.
Edges ( $E$ ): Connections are defined based on the physical adjacency of electrodes in the standard 10-20 montage system.
Adjacency Matrix: A binary matrix ($0$ or $1$) represents direct connections between neighboring electrodes.

D. Graph Convolutional Network (GCN) Architecture

The GCN processes the graph-structured data to learn both local node features and global spatial dependencies:

Feature Propagation: Aggregates information from neighboring nodes using a normalized adjacency matrix (incorporating self-loops).
Linear Transformation & Activation: Applies learnable weight matrices and ReLU activation functions across multiple layers.
Classification: A Softmax classifier in the final layer predicts the binary label (Seizure vs. Non-Seizure) for each segment.
Training: The model is trained using the Adam optimizer with a learning rate of 0.01 over 500 epochs, minimizing cross-entropy loss.

3. Key Contributions

Frequency-Specific Analysis: The study demonstrates that analyzing EEG signals in separate frequency bands yields different levels of diagnostic accuracy, challenging the efficacy of single broadband models.
Ictal-Phase Focus: By isolating and training specifically on the ictal phase (active seizure), the model achieves higher precision and reliability compared to models trained on mixed phases.
GCN for EEG Topology: The integration of GCNs allows the model to respect the non-Euclidean geometry of the brain, effectively modeling the propagation of seizure activity across anatomical lobes.
Hybrid Approach: Combines interpretable, handcrafted signal features (time/frequency domain) with deep learning, offering a balance between performance and neurophysiological insight.

4. Experimental Results

The model was evaluated on the CHB-MIT dataset using 5-fold cross-validation.

Performance Metrics (Broadband vs. Specific Bands):

Frequency Band	Accuracy (%)	Specificity (%)	Sensitivity (%)
Delta	97.10	95.00	98.60
Theta	97.13	95.80	96.70
Alpha	99.50	99.38	99.20
Lower Beta	99.70	99.70	99.40
Higher Beta	51.40	3.00	100.00
Broadband (All)	99.01	99.60	99.03

Key Findings:
- Mid-Frequency Dominance: The Alpha and Lower Beta bands demonstrated superior performance, approaching near-perfect classification (99.7% accuracy for Lower Beta).
- Higher Beta Failure: The Higher Beta band (22–30 Hz) performed poorly (51.4% accuracy) with extremely low specificity (3%), indicating this band contains minimal discriminative information for seizures and leads to high false positives.
- Overall Performance: The broadband model achieved 99.01% accuracy, 98.06% sensitivity, and 99.03% specificity, with an F1-score of 96.67% and AUC of 97.67%.

5. Significance and Conclusion

This research highlights that frequency-specific analysis is crucial for epileptic seizure detection. The study proves that:

Spatial Modeling Matters: GCNs outperform traditional methods by capturing the spatial propagation of seizures across brain regions.
Band Selection is Critical: Not all frequency bands are equally informative. The Alpha and Lower Beta bands are the most reliable indicators for seizure activity, while Higher Beta bands may introduce noise.
Clinical Applicability: The model is lightweight, interpretable (via feature selection), and highly accurate, making it suitable for real-time, resource-constrained clinical environments.
Future Directions: The findings suggest that future seizure detection systems should prioritize mid-frequency bands and utilize graph-based architectures to improve diagnostic precision and reduce false positives.