Lightweight Seizure Detection Based on Multi-Scale Channel Attention

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epilepsy is a neurological disorder characterized by recurrent seizures, necessitating timely diagnosis and treatment.
• Manual analysis of electroencephalography (EEG) signals for seizure detection is labor-intensive and time-consuming.
• Existing deep learning methods for automated seizure detection often require substantial computational resources, limiting their use on portable devices.

Purpose of the Study:

• To develop a novel, lightweight neural network for efficient and accurate seizure detection using EEG signals.
• To reduce the computational complexity of seizure detection algorithms for deployment on resource-constrained devices.
• To enable the development of low-cost, portable epilepsy monitoring solutions.

Main Methods:

• Proposed a novel lightweight neural network architecture for seizure detection.
• The network utilizes pure convolutions, incorporating an inverted residual structure and a multi-scale channel attention mechanism.
• The model was evaluated on the CHB-MIT dataset.

Main Results:

• Achieved high performance metrics: 98.7% accuracy, 98.3% sensitivity, and 99.1% specificity.
• Demonstrated significantly reduced computational complexity with only 2.68M multiply-accumulate operations (MACs).
• The model is highly parameter-efficient, requiring only 88K parameters.

Conclusions:

• The proposed lightweight neural network offers an effective solution for automated seizure detection.
• Its reduced computational demands make it suitable for deployment on low-cost, portable devices.
• This advancement can improve accessibility to continuous epilepsy monitoring and management.