Epileptic seizures, characterized by abnormal and excessive electrical activity in the brain, pose a significant global health burden. Electroencephalogram (EEG) signals remain the cornerstone of seizure diagnosis because they capture both temporal and spatial neural dynamics. While recent deep learning methods have reported high detection accuracy, they often function as black boxes, lacking the interpretability needed for clinical trust and neurophysiological insight.
A new study, published on arXiv, introduces a frequency-aware framework that addresses this gap. The researchers—Jibon, Ferdaus Anam, Siddiqui, Fazlul Hasan, Deeba, Hossain, and Gahangir—decompose raw EEG signals into five distinct frequency bands: delta, theta, alpha, lower beta, and higher beta. From each band, they extract eleven discriminative features that characterize seizure-related patterns. These features are then fed into a Graph Convolutional Neural Network (GCN) , which models the spatial dependencies among EEG electrodes by treating each electrode as a node in a graph.
Frequency Band Decomposition and Feature Engineering
The rationale behind frequency-band separation is that seizure activity often manifests differently across spectral ranges. The five bands span from low-frequency delta (0.5-4 Hz) to higher beta (12.5-30 Hz). For each band, the eleven features include statistical and entropy-based measures that capture amplitude, variability, and complexity. This approach contrasts with conventional broadband EEG analysis, which mixes all frequencies and can obscure band-specific seizure signatures.
Graph Convolutional Neural Network for Spatial Modeling
EEG electrodes are arranged spatially on the scalp, and seizure activity often propagates across regions. The GCN architecture takes advantage of this by constructing a graph where nodes represent electrodes and edges represent spatial proximity or functional connectivity. Through convolutional operations on this graph, the network learns to recognize patterns of neural synchrony and propagation that are characteristic of ictal (seizure) phases. This spatial reasoning is a key differentiator from traditional convolutional neural networks that treat EEG channels as independent input dimensions.
Experimental Results and Performance
Experiments were conducted on the CHB-MIT scalp EEG dataset, a widely used benchmark for seizure detection. The model achieved the following per-band accuracies:
| Frequency Band | Accuracy (%) |
|---|---|
| Delta | 97.1 |
| Theta | 97.13 |
| Alpha | 99.5 |
| Lower Beta | 99.7 |
| Higher Beta | 51.4 |
| Broadband | 99.01 |
The high accuracy on alpha (99.5%) and lower beta (99.7%) bands underscores the strong discriminative capability of mid-frequency activity during seizures. The notably lower performance on the higher beta band (51.4%) suggests that this frequency range carries less seizure-related information or may be dominated by noise. The overall broadband accuracy of 99.01% demonstrates that the frequency-aware approach competes with state-of-the-art methods while offering superior interpretability.
The results highlight the strong discriminative capability of mid-frequency bands and reveal frequency-specific seizure patterns.
Implications for Enterprise AI and Explainable Models
While this research focuses on medical diagnostics, its methodological contributions are relevant to enterprise technology leaders evaluating AI for critical decision-making. The use of graph neural networks to model spatial dependencies in sensor data can be applied to other domains, such as predictive maintenance on IoT sensor networks or anomaly detection in supply chain logistics. Moreover, the explicit integration of domain knowledge—here, neuroscience frequency bands—into the model architecture provides a template for building interpretable AI systems that engineers and business stakeholders can trust. As organizations push for explainable AI under regulatory scrutiny, approaches like this one, which decompose inputs into meaningfully interpretable components, offer a pathway to both accuracy and transparency.
The study is available on arXiv and represents a step toward clinically viable, explainable seizure detection systems. Its code and data are associated with the CHB-MIT dataset.
