Methods for computing the maximum performance of computational models of fMRI responses.

Area of Science:

• Computational neuroimaging
• Cognitive neuroscience
• Neuroscience methodology

Background:

• Computational neuroimaging uses models to predict brain responses (e.g., fMRI) to understand brain function.
• Prediction accuracy is limited by measurement noise, defining the 'noise ceiling'.
• Existing noise ceiling estimation methods (split-half, Monte Carlo) have limitations and unclear relations.

Purpose of the Study:

• To derive and validate a computationally efficient analytical method for estimating the noise ceiling in fMRI.
• To compare the analytical method with existing simulation-based and split-half approaches.
• To investigate the impact of noise structure, variance estimators, regularization, and model complexity on noise ceiling estimation.

Main Methods:

• Derivation of an analytical formula for the noise ceiling.
• Simulations to validate the analytical method against Monte Carlo simulations.
• Evaluation of different variance estimators and their effect on noise ceiling estimation.
• Assessment of regularization and model complexity interplay on performance relative to the noise ceiling.
• Testing the analytical method on real 7 Tesla fMRI data.

Main Results:

• The analytical noise ceiling estimation method is computationally efficient and requires less data than split-half procedures.
• Analytical estimates closely match Monte Carlo estimates in simulations.
• The analytical method provides similar estimates to the split-half method when considering across-run response variance.
• The analytical approach approaches the true noise ceiling under various simulated noise conditions.
• The method demonstrates validity on real 7 Tesla fMRI data.

Conclusions:

• A novel, analytically derived noise ceiling offers a more efficient and data-preserving alternative to existing methods in computational neuroimaging.
• This method provides reliable noise ceiling estimates comparable to established techniques, even with complex noise structures.
• The findings are applicable to real-world fMRI studies, enhancing the interpretation of computational model performance.