TL;DR
This paper introduces a new dynamic causal modeling (DCM) framework that accounts for both the phase and amplitude of neural oscillations, moving beyond traditional phase-only models. This allows for a more nuanced understanding of brain connectivity, particularly in states where amplitude dynamics play a significant role, such as during anesthesia.
What Problem Does This Paper Address?
Neuroscience often models brain activity as coupled oscillators, where interactions are primarily described by phase differences. While mathematically convenient, this "phase-only" approach can be restrictive, failing to capture important dynamics when amplitude variations are significant. This limitation is particularly relevant in understanding brain states like anesthesia, where oscillatory dynamics change considerably. The authors aim to bridge this gap by developing a more comprehensive model that explicitly incorporates amplitude dynamics alongside phase.
Methods
The core of this work is the extension of the Dynamic Causal Modelling (DCM) framework. The authors modify the standard neuronal state equation to accommodate phase-amplitude interactions. This involves:
- Generalizing the state variable: Representing the neural activity as a complex variable in polar coordinates (amplitude and phase).
- Introducing separate coupling matrices: Creating distinct matrices for phase-phase, amplitude-amplitude, and phase-amplitude coupling, allowing for their independent quantification.
- Leveraging Hopf Bifurcation: The model is designed for oscillators operating near a Hopf bifurcation, a point where a stable equilibrium can lose stability and give rise to periodic oscillations. This theoretical underpinning allows for dynamics to evolve beyond simple limit cycles.
- Bayesian Model Inversion and Reduction: The authors employ Bayesian methods to estimate model parameters and compare different model structures (e.g., phase-only vs. phase-amplitude) to determine the best fit for the data.
Datasets
The proposed model is evaluated on several datasets:
- Model-generated data: Data simulated from a system of three coupled oscillators to establish construct validity.
- Coupled pendula simulation: A physical simulation of three coupled pendula used to generate data for testing phase and amplitude dynamics under varying coupling strengths.
- Rodent LFP data: Local field potential recordings from rodent auditory cortices under different levels of isoflurane anesthesia.
- Macaque fMRI data: Functional magnetic resonance imaging data from macaque monkeys in both awake and anesthetized states.
Key Insights and Ideas
- Amplitude matters, especially in altered states: The study strongly suggests that amplitude dynamics are not merely noise but carry crucial information about neural states, particularly during anesthesia.
- Phase-only models are insufficient for strong coupling: While effective under weak coupling, phase-only models fail to capture the full picture when interactions are stronger or when amplitude variations are pronounced.
- A unified framework for phase and amplitude: The proposed DCM framework provides a principled way to disentangle and quantify the contributions of both phase and amplitude to neural connectivity.
- Model comparison reveals the importance of amplitude: Bayesian model reduction consistently favors models that include amplitude dynamics in certain scenarios, highlighting its significance.
Key Results
- Model-generated data: The phase-amplitude model accurately recovers parameters and identifies the correct generative model structure, demonstrating its validity.
- Coupled pendula simulation: The best model explaining the data shifted from "no phase or amplitude" and "phase-only" under weak coupling to the "phase-amplitude" model under strong coupling.
- Neuroimaging data:
- LFP data: Lower doses of anesthesia were best explained by models without phase or amplitude, while higher doses favored amplitude-only models.
- fMRI data: Awake states were best described by the phase-amplitude model, whereas anesthetized states were better explained by the amplitude-only model.
- Neuroimaging metrics: The phase-amplitude model generally outperformed the phase-only model in capturing metrics like Kuramoto order parameter, cross-correlation, and phase-lag index, with the exception of spectral entropy.
Limitations and Future Work
- Waveform shape: The authors acknowledge that sinusoidal oscillators may not be ideal for all neural signals and suggest exploring alternative waveforms in future work.
- Computational complexity: While the framework is designed to be tractable, incorporating amplitude increases model complexity, which can be a factor in model selection.
- Future directions: The authors propose that this model could offer insights into pathophysiological states characterized by disrupted synchrony, such as epilepsy and movement disorders, by examining how alterations in phase and amplitude contribute to these conditions. They also suggest exploring variations in correlation across levels of consciousness.
Summary updated