Tiny CNN for Seizure Prediction in Wearable Biomedical Devices.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Medical Technology

Background:

• Epilepsy affects millions globally, necessitating improved prediction and management strategies.
• Artificial intelligence (AI) offers promising avenues for enhancing epilepsy therapy and seizure prediction.
• Deep learning models, particularly Convolutional Neural Networks (CNNs), excel at detecting ictogenesis through complex data representations.

Purpose of the Study:

• To develop and evaluate a novel, compact deep learning model for accurate epileptic seizure prediction.
• To assess the performance of the proposed model against existing state-of-the-art methods.
• To investigate the feasibility of implementing the model on hardware-friendly wearable devices.

Main Methods:

• A tiny one-dimensional stacked convolutional neural network (1DSCNN) was designed.
• Short-time Fourier transform (STFT) was utilized as the input feature extraction method.
• The model was trained and tested on the American Epilepsy Society Seizure Prediction Challenge dataset.
• 4-bit quantization was applied to assess model size reduction and performance impact.

Main Results:

• The proposed 1DSCNN achieved high performance: 94.44% average sensitivity, 0.011/h average false prediction rate (FPR), and 0.979 average area under the curve (AUC).
• The model boasts a small size of 21.32kB, outperforming recent state-of-the-art methods.
• 4-bit quantization reduced model size by 7.08x with minimal AUC precision loss (0.51%).

Conclusions:

• The developed 1DSCNN demonstrates superior performance for epileptic seizure prediction.
• The model's small size and efficiency make it highly suitable for real-time, hardware-based wearable applications.
• This AI-driven approach holds significant potential for improving epilepsy management and patient care.