Bayesian inference for stationary points in Gaussian process regression models for event-related potentials analysis.

Area of Science:

• Statistics
• Machine Learning
• Neuroscience

Background:

• Stationary points in functions are vital for model interpretability and feature identification.
• Existing methods for inferring stationary points can be computationally complex and prone to mis-specification.
• Interpreting electroencephalography (EEG) signals often involves averaging, which can obscure individual-level details.

Purpose of the Study:

• To develop a semiparametric Bayesian model for efficient inference of stationary point locations and function estimation.
• To address computational challenges and mis-specification issues in stationary point detection.
• To apply the model to electroencephalography (EEG) data for individual-level analysis of event-related potentials (ERPs).

Main Methods:

• Utilized Gaussian processes as flexible priors for nonparametric function estimation.
• Imposed derivative constraints on the Gaussian processes to control function shape.
• Developed an inferential strategy focusing on at least one stationary point to enhance efficiency and avoid dimensional complexity.

Main Results:

• The proposed model successfully infers stationary point locations and estimates the underlying function.
• Applied to EEG data, the method automatically identifies characteristic components and their latencies at the individual level.
• Demonstrated that the model avoids excessive averaging, preserving individual variability in ERPs.

Conclusions:

• The semiparametric Bayesian model provides an efficient and robust approach for stationary point inference.
• The method offers a powerful tool for analyzing individual-level EEG data, particularly ERPs.
• The application to speech perception in different age groups highlights the model's utility in revealing age-related changes in cognitive processes.