Dynamic causal modelling of electrographic seizure activity using Bayesian belief updating

Gerald K. Cooray, Biswa Sengupta, Pamela K. Douglas, Karl J. FristonNeuroImage 2016 · DOI

TL;DR

This paper introduces a faster, more efficient method for analyzing complex brain activity during epileptic seizures using Dynamic Causal Modelling (DCM). By updating beliefs sequentially across short time windows, their approach significantly reduces computation time while maintaining accuracy, paving the way for more detailed studies of seizure generation.

What Problem Does This Paper Address?

Epileptic seizures are characterized by rapid, complex changes in brain activity that can persist for hours. Analyzing this spatiotemporal evolution requires sophisticated computational models, like Dynamic Causal Modelling (DCM), which can infer the underlying neural network dynamics from electrophysiological data (EEG/ECOG). However, standard DCM inversion, which analyzes the entire dataset at once, is computationally very expensive, making it impractical for large datasets or real-time analysis. This bottleneck hinders our ability to fully understand the mechanisms driving seizure initiation, propagation, and termination. The paper addresses this by developing a more computationally efficient DCM approach.

Methods

The core innovation is a Bayesian belief updating scheme for DCM. Instead of processing the entire time series at once, the data is divided into short, overlapping epochs. For each epoch, DCM is used to estimate the model parameters (e.g., synaptic efficacy, connectivity). The key idea is that the posterior estimates from one epoch are used as the prior beliefs for the next epoch. This "belief updating" mimics how our brains might learn and adapt over time.

This approach leverages the assumption that seizure dynamics involve at least two distinct timescales: fast neuronal activity and slower changes in synaptic parameters. The Bayesian belief updating method allows these slower synaptic parameters to "drift" or change gradually between epochs, modeled as a random walk. This is contrasted with the standard DCM approach, which uses temporal basis functions to model these slow changes across the entire dataset.

The authors also detail the theoretical underpinnings, including the use of variational Bayesian inference and the Laplace approximation for model inversion. They formalize the belief updating process as a prediction step (using posteriors from the previous epoch as priors for the current one) and an update step (using new data to refine these priors).

Datasets

The authors evaluated their method using two types of datasets:

  • Simulated EEG data: Three simulations were created to represent local field potentials with controlled drifts in inhibitory connectivity, excitatory connectivity, or endogenous input. This allowed for direct comparison of inferred parameters against known ground truth.
  • Empirical EEG/ECOG data:
    • Invasive ECoG data: Collected retrospectively from a patient with focal temporal lobe epilepsy.
  • Non-invasive EEG data: From two patients with recurrent partial seizures, likely due to anti-N-methyl-D-aspartate receptor (NMDA-R) antibody encephalitis.

Key Insights and Ideas

  • Two-Timescale Dynamics are Key: The paper strongly supports the idea that understanding seizures requires considering both fast neuronal firing and slower changes in synaptic strength or network connectivity. The Bayesian belief updating method is specifically designed to capture these slower drifts.
  • Computational Efficiency is a Game-Changer: The most striking insight is the dramatic speedup achieved. This isn't just a minor improvement; it's a 10-20 fold reduction in computation time, making complex DCM analyses feasible for large datasets and potentially for real-time applications.
  • Accuracy is Preserved: Despite the significant speedup, the Bayesian belief updating method yields comparable results to standard DCM in terms of parameter estimates and variance explained. This suggests that the efficiency gain doesn't come at a substantial cost to accuracy.
  • Flexibility in Modeling Parameter Changes: The random walk model for parameter fluctuations offers a more flexible way to capture potentially abrupt changes in seizure activity compared to the more constrained temporal basis functions used in standard DCM.
  • Potential for Clinical Translation: The ability to efficiently model seizure dynamics opens doors for better understanding individual patient seizure patterns and potentially informing treatment strategies.

Key Results

  • Speedup: The Bayesian belief updating scheme was approximately 5-10 times faster than standard DCM, taking minutes instead of 1-2 hours for inversion.
  • Accuracy: The method provided similar estimates of time-varying synaptic parameters compared to standard DCM. The difference in variance explained was small, generally less than 5%.
  • Simulated Data: The method successfully recovered the true parametric fluctuations in simulated data, demonstrating its ability to infer known changes in connectivity and input.
  • Empirical Data: The analysis of both invasive ECoG and non-invasive EEG data showed that the winning model (involving changes in inhibitory, excitatory connectivity, and spectral input) explained over 98% of the data variation.
  • Volatility: The study observed an increase in parameter volatility (rate of change) prior to seizure onset, suggesting that unknown factors significantly influence network dynamics at these critical moments.

Limitations and Future Work

  • Spectral Domain vs. Time Domain: The current spectral domain approach assumes linearity and may miss important phase information crucial for certain seizure types (e.g., spike-and-wave). Future work could compare this spectral method with time-domain approaches that retain phase information.
  • Approximation of Non-linearities: The method relies on linearizing the neural mass model. While this simplifies inversion, it might not fully capture the complex non-linear dynamics of seizures.
  • Activity-Dependent Plasticity: The authors suggest that incorporating activity-dependent plasticity (where fast activity influences slow parameter changes) could further improve parameter estimates and predictive power.
  • Clinical Validation: While the method is validated on empirical data, direct validation with more invasive physiological measurements (e.g., micro-dialysis) or through clinical interventions is a promising future direction.
  • Spatial Dynamics: The current model collapses spatial dynamics into afferent input. Incorporating explicit spatial dynamics could provide a more complete picture, especially for non-invasive data where localization is more challenging.

Summary updated

BibTeX
@article{cooray2016dynamic,
  author = {Gerald K. Cooray and Biswa Sengupta and Pamela K. Douglas and Karl J. Friston},
  title = {{Dynamic causal modelling of electrographic seizure activity using Bayesian belief updating}},
  journal = {NeuroImage},
  year = {2016},
  doi = {10.1016/j.neuroimage.2015.07.063},
}
Figure