Epileptic Spike Detection Using Neural Networks With Linear-Phase Convolutions.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Automated diagnostic-aid technologies for epileptic electroencephalogram (EEG) testing are crucial due to a shortage of skilled professionals.
• Effective frequency filtering is essential for identifying biomarkers like epileptic spike waves in EEG signals.

Purpose of the Study:

• To introduce a novel class of neural networks (NNs) incorporating data-driven, linear-phase finite impulse response (FIR) filters as a preprocessor.
• To enhance the extraction of biomarkers from EEG data without waveform distortion.

Main Methods:

• Developed NNs with a first layer comprising a bank of linear-phase FIR filters acting as data-driven bandpass filters.
• Trained the NNs using a large dataset of 15,833 clinical EEG epileptic spike waveforms from 50 patients.
• Compared the proposed filters against no preprocessing and discrete wavelet transform (DWT) in experimental validation.

Main Results:

• The trained data-driven filter bank effectively functioned as multiple bandpass filters, particularly passing frequencies around 10-30 Hz.
• The proposed method demonstrated high performance in detecting epileptic spikes.
• Achieved an area under the receiver operating characteristic curve (AUC) of 0.967 in mean intersubject validation.

Conclusions:

• The proposed data-driven filter bank within NNs offers a significant advancement in EEG signal processing for epilepsy diagnostics.
• This approach enhances biomarker extraction and improves the accuracy of automated seizure detection.
• The method shows promise for developing more reliable and efficient diagnostic-aid technologies in clinical neurology.