Researchers develop AI system to predict epileptic seizures — News-Medical

Researchers have developed the CMS-NAS system to predict epileptic seizures using electroencephalography data. It combines convolutional neural networks, Mamba and spiking neural networks and automatically searches for a model architecture intended to balance prediction accuracy and model size, News-Medical reports.

Combining three types of networks

Epilepsy affects approximately 50 million people worldwide. EEG makes it possible to record the brain's electrical activity in real time and is used to diagnose and monitor the disease. At the same time, such signals are non-stationary, noisy and differ substantially among people, making their manual interpretation difficult.

As part of CMS-NAS, the developers created two hybrid blocks, ConvSpike and MambaSpike. They combine elements of convolutional, Mamba and spiking neural networks in a search space designed for seizure prediction.

More current news is available on the UA.News Telegram channel Telegram.

Model search using EEG data

First, the system converts multichannel EEG signals into time-frequency representations using the short-time Fourier transform. It then generates a set of possible architectures, trains and evaluates each of them. Models with better performance are selected to create subsequent generations through crossover and mutation.

The search space contains about 1.74 × 10¹⁷ possible architectures. To evaluate predictions, the researchers used AUC, sensitivity, the rate of false-positive alerts per hour of interictal EEG, and comparison with random prediction. Higher AUC and sensitivity values, as well as a lower false-alarm rate, are considered better indicators.