Modeling plasticity during epileptogenesis by long short term memory neural networks

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Epilepsy pathogenesis involves synaptic pathway changes.
• Data-driven models, like artificial neural networks (ANNs), can capture synaptic plasticity effects.
• Understanding these changes is crucial for developing novel epilepsy treatments.

Purpose of the Study:

• To propose and validate Long Short-Term Memory (LSTM) as an ideal architecture for modeling synaptic plasticity in epilepsy.
• To compare LSTM's performance against other ANNs (RNN, FFNN) in predicting epilepsy indicators.

Main Methods:

• Induced synaptic plasticity (potentiation/depotentiation) in animal models using kindling and transcranial magnetic stimulation.
• Recorded after-discharge (AD) and seizure behavior as plasticity indicators over six days.
• Trained LSTM, RNN, and Feedforward Neural Networks (FFNN) to predict AD and seizure behavior.

Main Results:

• LSTM demonstrated superior prediction accuracy for both seizure behavior (0.91 ± 0.01) and AD (0.82 ± 0.01).
• LSTM significantly outperformed RNN (seizure: 0.77 ± 0.02, AD: 0.74 ± 0.08) and FFNN (seizure: 0.59 ± 0.02, AD: 0.42 ± 0.1).

Conclusions:

• LSTM networks are highly effective for modeling synaptic plasticity changes relevant to epilepsy.
• The findings support LSTM's utility in advancing epilepsy research and treatment development.