🤖 AI Summary
Effective connectivity (EC) modeling from EEG data remains challenging due to the inherent complexity of neurodynamics and strong noise contamination. Method: This study proposes a deep learning framework integrating the biologically grounded Jansen-Rit neural mass model (JR-NMM) with a bidirectional long short-term memory (Bi-LSTM) network to enable noise-robust, local parameter inversion. Contribution/Results: We present the first systematic evaluation of deep learning for JR-NMM parameter inversion, combining Monte Carlo noise simulations and global sensitivity analysis to uncover learnable mappings between key biophysical parameters—such as synaptic gains and time constants—and observable EEG features. Experiments demonstrate high prediction accuracy and strong generalization across multiple noise levels. The results validate the feasibility and robustness of deep learning–driven whole-brain EC estimation and establish a novel, physiology-informed paradigm for functional brain connectivity analysis.
📝 Abstract
The study of effective connectivity (EC) is essential in understanding how the brain integrates and responds to various sensory inputs. Model-driven estimation of EC is a powerful approach that requires estimating global and local parameters of a generative model of neural activity. Insights gathered through this process can be used in various applications, such as studying neurodevelopmental disorders. However, accurately determining EC through generative models remains a significant challenge due to the complexity of brain dynamics and the inherent noise in neural recordings, e.g., in electroencephalography (EEG). Current model-driven methods to study EC are computationally complex and cannot scale to all brain regions as required by whole-brain analyses. To facilitate EC assessment, an inference algorithm must exhibit reliable prediction of parameters in the presence of noise. Further, the relationship between the model parameters and the neural recordings must be learnable. To progress toward these objectives, we benchmarked the performance of a Bi-LSTM model for parameter inference from the Jansen-Rit neural mass model (JR-NMM) simulated EEG under various noise conditions. Additionally, our study explores how the JR-NMM reacts to changes in key biological parameters (i.e., sensitivity analysis) like synaptic gains and time constants, a crucial step in understanding the connection between neural mechanisms and observed brain activity. Our results indicate that we can predict the local JR-NMM parameters from EEG, supporting the feasibility of our deep-learning-based inference approach. In future work, we plan to extend this framework to estimate local and global parameters from real EEG in clinically relevant applications.