Probabilistic Forecasting and Information-Theoretic Analysis of Multivariate fMRI Dynamics

Area of Science:

• Neuroscience
• Information Theory
• Machine Learning

Background:

• Functional magnetic resonance imaging (fMRI) signals possess complex temporal structures influenced by neural activity, physiology, and noise.
• Analyzing these signals requires advanced methods to capture multivariate dynamics and uncertainty.

Purpose of the Study:

• To frame region-of-interest (ROI)-level fMRI analysis as a probabilistic forecasting challenge.
• To investigate the predictability of blood-oxygen-level-dependent (BOLD) activity using an information-theoretic lens.
• To assess various forecasting models for their ability to predict future BOLD signals.

Main Methods:

• Modeled multiregional BOLD activity as a finite-memory stochastic process using the Natural Scenes Dataset.
• Trained and compared forecasting models: linear regression, exponential smoothing, recurrent neural networks, and transformers.
• Estimated information-theoretic quantities (entropy, predictive information) from model-derived distributions without Gaussian assumptions.

Main Results:

• Transformer models significantly outperformed a persistence baseline (p=0.001) in forecasting BOLD activity.
• Achieved a high predictive information fraction (η=75.49%) with transformer models.
• Directed information analysis indicated short-term prediction relies on within-ROI autoregressive structures.

Conclusions:

• Probabilistic forecasting and information theory can characterize fMRI dynamics' predictability and uncertainty.
• The framework offers insights into the directional organization of large-scale brain activity.
• Potential applications include neuroengineering and neural-state inference.