Comparison of different input modalities and network structures for deep learning-based seizure detection.

Area of Science:

• Computational Neuroscience
• Machine Learning in Medicine
• Epilepsy Research

Background:

• Manual electroencephalogram (EEG) review for seizure detection is time-consuming and prone to errors.
• Machine learning and deep learning approaches have been explored to automate seizure detection.
• Deep learning eliminates the need for manual feature extraction, simplifying the process.

Purpose of the Study:

• To systematically compare different input modalities and neural network architectures for automated seizure detection.
• To identify the optimal combination of input data and network structure for improved seizure detection accuracy.

Main Methods:

• Utilized intracranial EEG data from an epilepsy mouse model.
• Extracted features from 5-second segments: raw time-series EEG, periodogram, short-time Fourier transform (STFT) images, and raw EEG waveform images.
• Implemented and compared Fully Connected Neural Networks (FCNN), Recurrent Neural Networks (RNN), and Convolutional Neural Networks (CNN).

Main Results:

• CNN demonstrated superior performance compared to FCNN and RNN across all input modalities.
• The highest Area Under the Curve (AUC) of 0.993 was achieved using 2D images of raw EEG waveforms with CNN.
• CNN effectively learns spatially-invariant representations from 2D EEG waveform images for seizure pattern recognition.

Conclusions:

• Convolutional Neural Networks (CNNs) are highly suitable for automated seizure detection.
• Representing raw EEG waveforms as 2D images significantly enhances detection performance.
• This approach offers a robust and accurate method for identifying seizures in EEG data.