Power efficient refined seizure prediction algorithm based on an enhanced benchmarking.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning significantly advances seizure prediction, but inconsistent benchmarks and overfitting hinder reliable results.
• Existing methods often suffer from inappropriate training and evaluation, leading to fluctuating or untrustworthy prediction performance.
• Lack of standardized data preparation and partitioning methods complicates model comparison and validation.

Purpose of the Study:

• To analyze data preparation and dataset partitioning methods in seizure prediction research.
• To explain the impact of these methods on prediction algorithms and model reliability.
• To propose a robust deep learning architecture for accurate and reliable seizure prediction.

Main Methods:

• Analysis of existing data preparation and dataset partitioning techniques in seizure prediction literature.
• Application of a rigorous processing procedure with appropriate sampling and leave-one-out cross-validation to prevent overfitting.
• Development of a deep learning model combining a one-dimensional convolutional neural network (1D-CNN) and a bi-directional long short-term memory (Bi-LSTM) network.

Main Results:

• The proposed 1D-CNN-BiLSTM architecture achieved 77.6% accuracy, 82.7% sensitivity, and 72.4% specificity.
• The model demonstrated superior performance compared to existing prior-art works in seizure prediction.
• The developed model is hardware-efficient, featuring 6.274k parameters and 12.825M floating-point operations.

Conclusions:

• The study presents a reliable deep learning framework for seizure prediction, overcoming common overfitting and benchmarking issues.
• The proposed 1D-CNN-BiLSTM model offers high accuracy and sensitivity, outperforming previous methods.
• The hardware-friendly nature of the model makes it suitable for implementation on memory and power-constrained devices.