Time-resolved dynamic computational modeling of human EEG recordings reveals gradients of generative mechanisms for the MMN response

Arnaud Poublan-Couzardot, Françoise Lecaignard, Enrico Fucci, Richard J. Davidson, Jérémie Mattout, Antoine Lutz, Oussama AbdounPLOS Computational Biology 2023 · DOI

TL;DR

This paper uses computational modeling to investigate the neural mechanisms behind the mismatch negativity (MMN) brain response. By analyzing EEG data in a time-resolved manner, they found evidence for distinct processes occurring at different stages of the MMN, suggesting a temporal gradient of adaptation and Bayesian learning mechanisms. This work offers a more nuanced understanding of auditory deviance detection and highlights the importance of dynamic, time-resolved modeling.

What Problem Does This Paper Address?

The mismatch negativity (MMN) is a well-studied electrophysiological marker of auditory deviance detection. However, the precise neurophysiological mechanisms generating the MMN remain debated. Two main theoretical accounts exist: adaptation, where neurons reduce their response to repeated stimuli, making them more sensitive to deviations, and Bayesian learning (or predictive coding), where the brain continuously updates its predictions about incoming stimuli and generates a "prediction error" signal when these predictions are violated. While both theories have explanatory power, direct comparisons and understanding their interplay, especially over time, have been limited by methodological challenges. This paper aims to resolve this debate by using a time-resolved computational approach to compare these competing hypotheses on human EEG data.

Methods

The authors employed a Bayesian modeling framework using variational Bayes inference (VBA toolbox) to model trial-by-trial EEG data. They developed several generative models based on the competing theories:

  • Static Models:
    • Null Model: Predicts a constant response.
  • Deviant Detection (DD) Model: Predicts a response only to deviant stimuli, without considering history.
  • Adaptation Models (SSA): Implemented as exponential decay and recovery processes, reflecting stimulus-specific adaptation. Seven models with different time constants ($\tau_a$) were tested.
  • Bayesian Learning (BL) Models: Based on predictive coding, these models assume the brain learns probabilities of stimuli. Three variants were implemented, differing in their "output" or observation function:
  • BL$_{surp}$ (Shannon Surprise): Encodes surprise about the stimulus category.
  • BL$_{pwpe}$ (Precision-Weighted Prediction Error): Computes the difference between predicted and actual sensory data, weighted by precision.
  • BL$_{madj}$ (Model Adjustment): Quantifies the change in belief about model parameters. Six models with different temporal integration parameters ($\tau_1$) were tested for each BL variant.

Crucially, the analysis was performed time-resolved, examining 10ms time windows across the MMN latency range (50-250ms). Model comparison was done using fixed-effect Bayesian Model Comparison (BMC) at the family level, followed by Bayesian Model Averaging (BMA) to estimate the optimal temporal parameters. Model recovery analyses using simulated data were conducted to assess the reliability and distinguishability of the models under varying signal-to-noise ratios (SNR).

Datasets

The study utilized two independent EEG datasets from previously published experiments:

  • Study 1: Data from 31 healthy participants (mean age 42.9 ± 10.4 years) collected during a passive auditory oddball task.
  • Study 2: Data from 66 healthy participants (mean age 52 ± 7.7 years) from a replication study, also using a passive auditory oddball task.

Both datasets involved standard and deviant tones presented in an oddball paradigm. Control conditions were used for modeling. No new benchmarks were introduced; existing datasets were re-analyzed with a novel modeling approach.

Key Insights and Ideas

  • Temporal Dissociation: The most striking finding is the evidence for distinct mechanisms dominating different temporal phases of the MMN. This challenges the idea of a single, unified MMN generator.
  • Early Adaptation Dominance: Adaptation models (SSA) were found to best explain the early part of the MMN response (approx. 80-170ms), suggesting that neuronal fatigue plays a significant role in detecting early deviance.
  • Later Bayesian Learning: Bayesian learning models, particularly the "model adjustment" variant (BL$_{madj}$), became more dominant in later MMN windows (approx. 170-230ms), indicating that higher-level predictive processes and belief updating are crucial for processing deviance over time.
  • Gradient of Adaptation: Within the adaptation family, a gradient of adaptation rates was observed, with slower adaptation dominating earlier and faster adaptation later. This echoes findings in animal studies about hierarchical processing in the auditory system.
  • SNR Matters: The quality of the EEG data (signal-to-noise ratio) significantly impacts model identifiability. Low SNR can lead to biased model selection, highlighting the importance of data quality and careful interpretation.
  • Model Recovery is Crucial: The study emphasizes the necessity of model recovery analyses to validate BMC results and understand model distinguishability under different conditions.

Key Results

  • Early MMN (80-170ms): Adaptation models (SSA) showed significantly higher evidence against the null model across both studies.
  • Mid-MMN (160-180ms): Model adjustment (BL$_{madj}$) models began to compete with SSA models, suggesting a transition in generative mechanisms.
  • Late MMN (up to 230ms): Bayesian learning models, particularly BL$_{surp}$ and BL$_{madj}$, showed increasing evidence, with BL$_{surp}$ exhibiting the highest, though moderate, evidence.
  • Model Recovery: Simulations showed that models could be recovered with high accuracy at high SNR, but confusion increased significantly with lower SNR. BL$_{surp}$ and BL$_{pwpe}$ were highly correlated and difficult to disambiguate.
  • Goodness-of-Fit: The best-fitting models explained a small but consistent proportion of the EEG variance (0.5-1.8%), comparable to previous studies. The ratio of observed to maximum possible variance ($R_{obs}^2/R_{max}^2$) indicated that the selected models provided a good fit to the data.

Limitations and Future Work

  • High-Pass Filtering: The authors acknowledge that the aggressive high-pass filtering used in the preprocessing of the EEG data might have removed some low-frequency drifts but advise future studies to use lower thresholds when possible.
  • Model Space Limitations: While the model space was comprehensive, certain models (e.g., "confidence corrected surprise" BL model) were not included and could be explored in future work.
  • Disambiguation of BL Outputs: BL$_{surp}$ and BL$_{pwpe}$ were difficult to distinguish, even under optimal conditions. Future research could explore alternative model formulations or experimental designs to better separate these mechanisms.
  • Cortical Source Localization: The current study used scalp EEG, limiting the ability to perform cortically resolved modeling. Future work should leverage MEG or EEG-MEG fusion data for source reconstruction.
  • Biophysical Plausibility: While the models capture computational principles, future work could focus on developing more biophysically plausible implementations of these mechanisms.
  • Animal-Human Homology: While parallels were drawn to animal studies, differences in experimental paradigms and measures warrant caution. Further research is needed to confirm these homologies.
  • Fine-Grained Temporal Dynamics: The study suggests a temporal gradient, but future research could investigate even finer temporal resolutions and explore the interplay of these mechanisms with other cognitive processes.

Summary updated

BibTeX
@article{poublancouzardot2023time,
  author = {Arnaud Poublan-Couzardot and Françoise Lecaignard and Enrico Fucci and Richard J. Davidson and Jérémie Mattout and Antoine Lutz and Oussama Abdoun},
  title = {{Time-resolved dynamic computational modeling of human EEG recordings reveals gradients of generative mechanisms for the MMN response}},
  journal = {PLOS Computational Biology},
  year = {2023},
  doi = {10.1371/journal.pcbi.1010557},
}
Figure