Bidirectional long short-term memory attention neural network to estimate neural mass model parameters

Abstract

Mild Cognitive Impairment (MCI) is a precursor stage of Alzheimer's disease. The effective utilization of electroencephalography (EEG) in the analysis of neuronal populations through mathematical models of the brain is imperative for the study of the neurophysiological mechanisms underlying MCI. The research performs multi-parameter reverse identification of EEG in patients with MCI by combining the neural network model and the Jansen-Rit Neural Mass Model (J&R model) and then studying the brain function problems of MCI patients. We proposed the Bidirectional Long Short-term Memory Attention (BiLSAT) model by integrating the Bidirectional Long Short-term Memory (BiLSTM) network with the Attention mechanism. The BiLSAT model is designed using an encoder-decoder architecture that combines the BiLSAT model with the J&R model through a loss function. We compared the BiLSAT model and the Unscented Kalman Filter (UKF) algorithm through multi-parameter experiments. Degradation experiments demonstrated that the BiLSTM module and the Attention module enhance the performance of the BiLSAT model. The results of the multi-parameter experiments showed that the BiLSAT model demonstrates higher accuracy in multi-parameter identification compared to the UKF algorithm. We used the BilSAT model to estimate the parameters of the EEG data of the healthy elderly group and the MCI group. The results showed that the excitatory-inhibitory balance in the brains of patients with MCI was dysfunctional. In this study, a novel inverse identification method for the J&R model is proposed. This method is intended to address the limitations of the conventional UKF algorithm, which has been observed to suffer from inaccurate multi-parameter identification. The proposed method offers a novel perspective for future research in this field.