C²SP-Net: Joint Compression and Classification Network for Epilepsy Seizure Prediction

Area of Science:

• Biomedical Engineering
• Neuroscience
• Signal Processing

Background:

• Brain-machine interface (BMI) technology enables seizure prediction.
• High data volumes and computational demands of electrophysiological signals are bottlenecks for BMI systems, especially in wearable devices.
• Existing data compression methods require complex steps, hindering real-time seizure prediction.

Purpose of the Study:

• To develop an efficient framework for joint signal compression, seizure prediction, and signal reconstruction.
• To address the bandwidth and computational limitations in BMI-based seizure prediction.
• To minimize energy consumption in power-critical medical devices.

Main Methods:

• Proposed C²SP-Net, a framework integrating in-sensor compression, prediction, and reconstruction.
• Utilized a plug-and-play in-sensor compression matrix to reduce transmission bandwidth.
• Evaluated energy efficiency, prediction accuracy, sensitivity, false prediction rate, and reconstruction quality across various compression ratios.

Main Results:

• C²SP-Net achieves joint compression, prediction, and reconstruction with minimal computational overhead.
• The framework significantly reduces transmission bandwidth requirements.
• Demonstrated energy efficiency and superior prediction accuracy compared to state-of-the-art methods.
• Achieved an average prediction accuracy loss of only 0.6% with compression ratios from 1/2 to 1/16.

Conclusions:

• C²SP-Net offers an energy-efficient solution for BMI-based seizure prediction.
• The proposed framework overcomes bandwidth and computational limitations.
• High-fidelity signal reconstruction is possible alongside efficient compression and accurate prediction.