Brain model state space reconstruction using an LSTM neural network.

Area of Science:

• Computational Neuroscience
• Machine Learning in Neuroscience
• Biomedical Signal Processing

Background:

• Kalman filtering is used for tracking neural model states and parameters in electroencephalography (EEG).
• Existing Kalman filter methods for neural mass models (NMMs) struggle with determining initial conditions and assume Gaussian state distributions.
• Accurate tracking of NMM states and parameters is crucial for brain modeling, monitoring, imaging, and control.

Purpose of the Study:

• To present a data-driven deep learning method for tracking NMM states and parameters from EEG.
• To overcome the limitations of traditional Kalman filtering, particularly the need for accurate initial conditions.
• To provide a general and efficient approach for estimating brain model variables often difficult to measure directly.

Main Methods:

• A long short-term memory (LSTM) neural network was trained as a filter on simulated EEG data generated by an NMM.
• A customized loss function enabled the LSTM filter to learn NMM behavior and output state vectors and parameters.
• The LSTM filter was tested on simulated data and applied to real EEG recordings, including those with epileptic seizures.

Main Results:

• The LSTM filter achieved high correlations (R-squared ≈ 0.99) on simulated data, demonstrating robustness to noise.
• The method proved more accurate than a nonlinear Kalman filter when initial conditions were inaccurate.
• Application to real EEG data revealed changes in connectivity strength parameters at the onset of epileptic seizures.

Conclusions:

• The proposed LSTM-based filter offers a novel, efficient, and data-driven alternative to Kalman filtering for tracking NMMs.
• This approach eliminates the need for specifying difficult-to-obtain initial state vectors and parameters.
• The method is generalizable to any NMM and has significant potential for advancing brain modeling and analysis.