Seizure detection by convolutional neural network-based analysis of scalp electroencephalography plot images

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epileptologists visually analyze EEG plot images for seizure detection.
• Automated methods analyze spectro-temporal or complex EEG features.
• Convolutional neural networks (CNNs) possess human-like visual recognition capabilities.

Purpose of the Study:

• To explore image-based seizure detection using CNNs on EEG data.
• To compare CNN performance against automated seizure detection software.
• To develop a clinically practical seizure detection index.

Main Methods:

• EEG data were filtered, segmented into time windows, and converted to plot images.
• CNNs classified EEG images as 'seizure' or 'non-seizure'.
• Performance was evaluated using true positive rates and false alarm rates.

Main Results:

• CNNs achieved a median true positive rate of 74% (1-s window), outperforming BESA (20%) and Persyst (31%).
• Median seizure detection rate per minute was 100% for CNNs, compared to 73.3% (BESA) and 81.7% (Persyst).
• CNNs produced a low false alarm rate of 0.2 per hour.

Conclusions:

• CNNs enable seizure detection through artificial visual recognition of EEG images.
• This method offers a practical alternative to manual inspection by epileptologists.
• Improving CNN performance requires diverse seizure pattern training data.