Joint estimation of source dynamics and interactions from MEG data

Narayan Puthanmadam Subramaniyam, Filip Tronarp, Simo Särkkä, Lauri ParkkonenNetwork 2025 · DOI

TL;DR

This paper introduces JEDI-MEG, a novel algorithm that jointly estimates the location, activity, and interactions of neural sources from MEG data. By overcoming the limitations of traditional two-step methods, JEDI-MEG provides more accurate and physiologically plausible functional connectivity estimates, paving the way for a deeper understanding of brain dynamics.

What Problem Does This Paper Address?

Understanding how different brain regions communicate (functional connectivity) is crucial for deciphering brain function. Magnetoencephalography (MEG) is a powerful tool for measuring brain activity with high temporal resolution. However, estimating functional connectivity from MEG data is challenging due to the "inverse problem" – inferring the location and activity of internal brain sources from external sensor measurements.

Current methods typically tackle this in two separate steps: first, estimating the location and time series of neural sources, and second, calculating connectivity between these estimated sources. This sequential approach is suboptimal because it doesn't leverage the information that connectivity patterns might provide about source activity, and vice-versa. This can lead to "spatial leakage" and inaccurate connectivity estimates, especially when dealing with complex interactions. The paper aims to address this by developing a unified approach that estimates sources and their interactions simultaneously.

Methods

The core of the paper is the JEDI-MEG (Joint Estimation of source Dynamics and Interactions from MEG data) algorithm. It frames the problem within a state-space model (SSM).

  • State-Space Model: This model describes how the system (brain sources) evolves over time. It includes:
  • Source Locations: Modeled as a random walk.
  • Source Amplitudes: Modeled using a Pth order Vector Autoregressive (Pth order MVAR) model. The coefficients of this MVAR model directly represent the interactions (connectivity) between sources.
  • Observation Model: Relates the source activity to the MEG sensor data via a lead-field matrix (which describes how each source's activity projects to the sensors).
  • Bayesian Filtering and Estimation: JEDI-MEG employs a Stochastic Approximation Expectation-Maximization (SAEM) algorithm, a variant of the Expectation-Maximization (EM) algorithm. This is combined with Rao-Blackwellization (RB) and a particle Markov chain Monte Carlo (PMCMC) method.
  • Stochastic E-step: Uses a particle filter (specifically, a conditional particle filter with ancestor sampling) to approximate the posterior distribution of source locations and amplitudes.
  • Maximization M-step: Uses a Maximum Likelihood (ML) approach to estimate the MVAR coefficients (connectivity) and noise covariance parameters.

This joint estimation approach treats the connectivity estimation as a system identification problem within the state-space framework, allowing for a more integrated and accurate inference.

Datasets

The paper uses both simulated and real datasets for evaluation:

  • MVAR-based Simulations: Synthetic MEG data generated using MVAR models. Two scenarios were simulated:
  • Type-I: Unidirectional interaction between two sources, with a third non-interacting source.
  • Type-II: Unidirectional interaction among all three sources. These simulations were used to evaluate source localization accuracy, connectivity estimation error (using generalized PDC - gPDC), and direction of interaction accuracy (TPR/FPR).

  • ECoG-based Simulations: Real electrocorticography (ECoG) data from face perception experiments was used as a basis for realistic MEG signal simulations. This approach aims to capture more complex, non-linear neural characteristics.

  • Real MEG Data: Publicly available MEG data from subjects viewing images of human faces (Wakeman & Henson, 2015). This dataset was used to demonstrate the algorithm's performance on real-world data and to assess the physiological plausibility of the results.

Key Insights and Ideas

  • Joint Estimation is Superior: The fundamental insight is that simultaneously estimating source locations, amplitudes, and their interactions leads to more accurate and reliable functional connectivity estimates compared to the traditional two-step approach.
  • Overcoming Spatial Leakage: By integrating source estimation and connectivity inference, JEDI-MEG mitigates the problem of spatial leakage, which can introduce spurious connections in sequential methods.
  • System Identification for Connectivity: Framing connectivity estimation as a system identification problem within a state-space model provides a powerful and principled way to infer directed interactions.
  • Robustness to Noise and Modeling Errors: The algorithm demonstrates good performance even with low signal-to-noise ratios (SNR) and simulated modeling errors, suggesting its practical utility.
  • Physiologically Plausible Results: When applied to real MEG data, JEDI-MEG produces source locations and connectivity patterns that align with known neuroanatomy and previous findings in face processing research.

Key Results

  • Source Localization: JEDI-MEG consistently achieved higher accuracy (measured by the number of correctly identified sources within a 10mm error margin) than LCMV and MCMV beamformer-based methods across various SNR levels in MVAR simulations. It also showed lower mean localization error (SLE).
  • Connectivity Estimation (MVAR Simulations):
    • False Positive Rate (FPR): JEDI-MEG significantly outperformed the two-step approaches (LCMV/MCMV + MVAR fitting) by yielding substantially lower FPR, indicating fewer incorrect connections identified.
  • True Positive Rate (TPR): While TPR was comparable across methods at low SNRs, JEDI-MEG showed a trend towards higher TPR at higher SNRs, suggesting it identifies more true connections without introducing many false ones.
  • Reconstruction Error (RE): JEDI-MEG achieved the lowest reconstruction error for the gPDC matrix.
  • ECoG-based Simulations: JEDI-MEG demonstrated comparable or superior source localization accuracy to MCMV, particularly in challenging low-SNR scenarios with modeling errors. It also showed lower FPR in connectivity estimation.
  • Real MEG Data: The algorithm successfully identified sources in brain regions known to be involved in face processing (occipital, temporal, frontal cortices) and produced connectivity estimates that were largely consistent across subjects and aligned with prior research.

Limitations and Future Work

  • Number of Sources: The algorithm requires the number of sources to be specified a priori. While it can handle extra sources by assigning them near-zero amplitudes, setting this number too high can increase computational cost. The authors suggest that if the number is set too low, the algorithm will only find the strongest sources.
  • Static Connectivity: The current implementation estimates static connectivity, which may be suboptimal for non-stationary data like event-related potentials (ERPs).
  • Event-Related Transients: The analysis did not explicitly exclude event-related transients, meaning connectivity might be driven by evoked activity rather than intrinsic brain activity.
  • Future Work:
    • Develop methods to estimate dynamic functional connectivity from non-stationary MEG/EEG data, potentially by combining JEDI-MEG with dynamic Kalman filtering approaches.
  • Explore variational Bayes or ensemble Kalman filters for whole-brain dynamic functional connectivity estimation.
  • Extend the algorithm to EEG data, which should be straightforward.
  • Investigate methods to automatically determine the number of sources.
  • Address the issue of event-related transients.

Summary updated

Abstract

Current techniques to estimate directed functional connectivity from magnetoencephalography (MEG) signals involve two sequential steps: (a) estimation of the sources and their amplitude time series from the MEG data and (b) estimation of directed interactions between the source time series. However, such a sequential approach is not optimal as it leads to spurious connectivity due to spatial leakage. Here, we present an algorithm to jointly estimate the source and connectivity parameters using Bayesian filtering. We refer to this new algorithm as JEDI-MEG (Joint Estimation of source Dynamics and Interactions from MEG data). By formulating a state-space model for the locations and amplitudes of a given number of sources, we show that estimation of their connections can be reduced to a system identification problem. Using simulated MEG data, we show that the joint approach provides a more accurate reconstruction of connectivity parameters than the conventional two-step approach. Using real MEG responses to visually presented faces in 16 subjects, we also demonstrate that our method gives source and connectivity estimates that are both physiologically plausible and largely consistent across subjects. In conclusion, the proposed joint estimation approach outperforms the traditional two-step approach in determining functional connectivity in MEG data.
BibTeX
@inproceedings{subramaniyam2025joint,
  author = {Narayan Puthanmadam Subramaniyam and Filip Tronarp and Simo Särkkä and Lauri Parkkonen},
  title = {{Joint estimation of source dynamics and interactions from MEG data}},
  booktitle = {Network},
  year = {2025},
  doi = {10.1162/netn\_a\_00453},
}
Figure