Memristive Neural Networks for Predicting Seizure Activity

Area of Science:

• Neuroscience
• Artificial Intelligence
• Materials Science

Background:

• Chronic epileptiform activity in mice was induced using pilocarpine.
• Local field potentials (LFP) were recorded from the hippocampus and medial entorhinal cortex of control and epileptic mice.
• This study investigates predicting pathological neuronal activity.

Purpose of the Study:

• To assess the predictive capabilities of neuronal activity data for epileptiform activity.
• To develop and implement a deep artificial neural network (ANN) for predicting epileptiform activity.
• To demonstrate the ANN's implementation using memristive devices.

Main Methods:

• A deep long short-term memory (LSTM) artificial neural network (ANN) was developed.
• The ANN was trained using supervised learning on LFP recordings from mice with induced chronic epileptiform activity.
• The ANN implementation utilized memristive devices based on redox processes in metal-oxide-metal films.

Main Results:

• The deep ANN successfully predicted epileptiform activity 40 milliseconds before its occurrence.
• A high evaluation metric (low root-mean-square error of 0.019) was achieved, demonstrating the approach's efficacy.
• The developed neural network architecture showed significant potential for accurate prediction.

Conclusions:

• The deep ANN can predict pathological neuronal activity, including focal epileptic seizures in mice, before they occur.
• The approach offers potential for building long-term prognoses of disease progression based on LFP data.
• This memristive device-based ANN presents a novel method for analyzing and predicting pathological neuronal activity, potentially improving epilepsy diagnosis and prognostication.