Improved data quality and statistical power of trial-level event-related potentials with Bayesian random-shift Gaussian processes

Dustin Pluta, Beniamino Hadj-Amar, Meng Li, Yongxiang Zhao, Francesco Versace, Marina VannucciScientific 2024 · DOI

TL;DR

This paper introduces the Bayesian Random Phase-Amplitude Gaussian Process (RPAGP) model, a novel Bayesian framework for analyzing individual trial event-related potentials (ERPs) in EEG data. RPAGP significantly improves data quality and statistical power by de-noising trial-level signals, leading to more sensitive and accurate detection of experimental condition differences compared to traditional averaging methods.

What Problem Does This Paper Address?

Analyzing electroencephalogram (EEG) data often involves studying event-related potentials (ERPs), which are brain responses time-locked to specific events. Traditionally, researchers average ERPs across multiple subjects and trials to reduce noise and identify group-level effects. However, this averaging process can obscure important trial-to-trial and subject-to-subject variability, potentially masking subtle but scientifically relevant findings. The challenge lies in accurately modeling and inferring these individual trial ERPs, which are inherently noisy and can vary in amplitude and timing. Existing methods for trial-level analysis often require subjective initialization or make restrictive assumptions, limiting their flexibility and interpretability.

Methods

The core contribution is the Bayesian Random Phase-Amplitude Gaussian Process (RPAGP) model. This framework treats individual trial ERPs as a common structural signal that is then modified by trial-specific amplitude and phase (latency) shifts, plus ongoing brain activity. Key methodological aspects include:

  • Gaussian Process (GP) Prior: A GP prior is used to model the underlying structural signal, allowing for flexible and data-driven estimation of its shape.
  • Random Phase-Amplitude Shifts: Trial-specific amplitude scaling and latency shifts are incorporated to account for individual trial variations.
  • Autoregressive Process: Ongoing brain activity is modeled using an autoregressive process to capture temporal dependencies.
  • Bayesian Inference: The model uses a Bayesian approach, providing not only point estimates but also uncertainty quantification for all parameters, including trial-specific signals. This allows for robust statistical inference.
  • MCMC Sampling: Markov Chain Monte Carlo (MCMC) methods are employed for posterior inference of model parameters.
  • De-noising: A key outcome of the RPAGP model is the generation of "de-noised" trial predictions, which effectively separate the signal from the noise.

Datasets

The study utilizes data from a previously published experiment involving 161 subjects observing emotionally arousing images (high, low, and neutral arousal levels). This dataset was originally reported in Versace et al. (2016). No new datasets were introduced.

Key Insights and Ideas

  • De-noising is Crucial for Power: The most significant insight is that by effectively de-noising individual trial ERPs, RPAGP dramatically increases statistical power. This is achieved not by boosting the signal itself, but by reducing the uncertainty (narrower confidence intervals) around the estimates.
  • Trial-Level Variability Matters: The model explicitly accounts for trial-level variations in amplitude and latency, which are often ignored in traditional averaging methods. This allows for a more nuanced understanding of brain responses.
  • Bayesian Framework for Interpretability: The Bayesian approach provides posterior distributions for all estimated parameters, offering a rich understanding of uncertainty and enabling more robust statistical comparisons.
  • Flexibility Over Existing Methods: RPAGP offers greater flexibility than previous trial-level ERP analysis methods (like dVCA or ASEO) by avoiding subjective initialization of component waveforms and allowing for continuous latency estimation.

Key Results

  • Improved Statistical Power: RPAGP detected expected differences in the Late Positive Potential (LPP) between emotional conditions in a significantly higher proportion of subjects (86.3% for High-Neutral) compared to existing methods like ASEO (45.3%), empirical bootstrap (EMP) (44.1%), and ANOVA (42.9%).
  • Enhanced Data Quality (SNR): The de-noising process of RPAGP resulted in a substantial improvement in signal-to-noise ratio (SNR), with an average 15-fold increase compared to empirical means.
  • Narrower Confidence Intervals: RPAGP produced substantially narrower credible intervals for LPP mean differences compared to the empirical bootstrap method, leading to more precise estimates and increased statistical power.
  • No Increase in False Positives: The method was shown to substantially increase the detection of true positives without increasing the probability of false positives.
  • Accurate Structural Signal Estimation: Simulations demonstrated that RPAGP provides more accurate estimates of the true structural signal compared to empirical means, especially when trial-specific latencies are present.

Limitations and Future Work

  • Computational Cost: Fitting the RPAGP model for a single subject took approximately 10 hours on a MacBook Pro, indicating a significant computational cost, though parallel processing was used for multiple subjects.
  • Assumed Common Structural Signal: For the presented analyses, a common structural signal was assumed across all trials within a condition. The authors suggest that extending the model to allow for different structural signals across experimental categories could be a valuable future direction.
  • Multiple Component Waveforms: While the current model assumes a single structural signal, future work could explore incorporating multiple component waveforms, similar to the framework in previous VSPOA models, to capture more complex ERP structures.
  • Trend Analysis: The paper briefly touches on trend analysis of LPP means over the course of the experiment, but suggests that a formal analysis of trends using the proposed model is an interesting avenue for future research.

Summary updated

Abstract

Studies of cognitive processes via electroencephalogram (EEG) recordings often analyze group-level event-related potentials (ERPs) averaged over multiple subjects and trials. This averaging procedure can obscure scientifically relevant variability across subjects and trials, but has been necessary due to the difficulties posed by inference of trial-level ERPs. We introduce the Bayesian Random Phase-Amplitude Gaussian Process (RPAGP) model, for inference of trial-level amplitude, latency, and ERP waveforms. We apply RPAGP to data from a study of ERP responses to emotionally arousing images. The model estimates of trial-specific signals are shown to greatly improve statistical power in detecting significant differences in experimental conditions compared to existing methods. Our results suggest that replacing the observed data with the de-noised RPAGP predictions can potentially improve the sensitivity and accuracy of many of the existing ERP analysis pipelines.
BibTeX
@inproceedings{pluta2024improved,
  author = {Dustin Pluta and Beniamino Hadj-Amar and Meng Li and Yongxiang Zhao and Francesco Versace and Marina Vannucci},
  title = {{Improved data quality and statistical power of trial-level event-related potentials with Bayesian random-shift Gaussian processes}},
  booktitle = {Scientific},
  year = {2024},
  doi = {10.1038/s41598-024-59579-2},
}
Figure