DCSENets: Interpretable deep learning for patient-independent seizure classification using enhanced EEG-based...

Area of Science:

• Medical imaging and signal processing
• Artificial intelligence in healthcare
• Neurology and epilepsy research

Background:

• Epileptic activity detection in EEG signals is time-consuming for neurologists.
• Current computer-aided diagnosis systems face challenges due to EEG signal complexity and patient variability.
• Short-time Fourier transform (STFT) offers time-varying frequency analysis but has limitations in time-frequency resolution.

Purpose of the Study:

• To develop a novel STFT spectrogram construction algorithm with taper functions for high-resolution EEG channel extraction.
• To enhance the accuracy and interpretability of computer-aided seizure diagnosis.
• To achieve patient-independent seizure classification using deep learning models.

Main Methods:

• Extraction of seizure and non-seizure segments from the CHB-MIT EEG dataset.
• Application of taper functions (Hann, Gaussian) to STFT spectrograms to minimize edge effects.
• Implementation of Dilated Convolutional Squeeze and Excitation Networks (DCSENets) with leave-one-patient-out cross-validation for classification.
• Integration of Grad CAM for deep learning model interpretability.

Main Results:

• The proposed DCSENets achieved an average accuracy of approximately 87% with and without taper functions.
• Performance metrics indicated similarity in train-test sample distribution for most patients.
• Grad CAM visualization improved the transparency of the deep learning model's decision-making process.

Conclusions:

• The novel STFT spectrogram construction with taper functions improves seizure diagnosis accuracy.
• The DCSENets model provides a transparent and interpretable tool for neurologists.
• This approach offers enhanced visualized spectrograms and a reliable computer-aided diagnosis system for epilepsy.