Dual deep neural network-based classifiers to detect experimental seizures.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biomedical Engineering

Background:

• Manual review of electroencephalograms (EEGs) for seizure detection is time-consuming and requires specialized expertise.
• Developing automated systems is crucial for efficient and reliable seizure detection in continuous EEG monitoring.

Purpose of the Study:

• To develop an efficient and robust automated system for detecting experimental seizures from continuous EEG data.
• To combine spectral analysis with deep neural networks for enhanced seizure detection accuracy.

Main Methods:

• A dual deep neural network approach was employed, utilizing spectral analysis (periodograms) of EEG segments.
• The first network discriminated seizure from non-seizure segments; the second network refined classification using misclassified non-seizure data.
• Simple pre- and post-processing steps were integrated to optimize performance and reduce computational load.

Main Results:

• The automated system achieved 100% sensitivity and 98% positive predictive value in detecting seizures across 4,272 hours of test EEG data, with only 6 false positives.
• The entire classification process for 8,977 hours of EEG data, including training and testing, was completed in just 2.28 hours on a personal computer.
• The system demonstrated high accuracy and a low computational burden, proving its practical feasibility.

Conclusions:

• Combining spectral analysis, dual deep neural networks, and rule-based processing provides an accurate and efficient method for automated convulsive seizure detection.
• The developed algorithm offers a feasible solution for large-scale EEG analysis, significantly reducing the manual workload.
• This approach highlights the potential of AI in improving neurological monitoring and diagnosis.