Stacked regressions and structured variance partitioning for interpretable brain maps.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Functional brain mapping relies on linking brain activity to stimulus properties.
• Correlated stimulus features in naturalistic stimuli complicate traditional brain mapping and statistical analysis.

Purpose of the Study:

• To develop robust methods for functional brain mapping with correlated stimulus features.
• To improve the interpretability and accuracy of brain maps derived from complex, naturalistic stimuli.

Main Methods:

• Developed a stacking algorithm combining multiple encoding models, each using different stimulus attributes.
• Introduced structured variance partitioning to account for known relationships between features.
• Validated methods through simulation and application to fMRI data.

Main Results:

• The combined encoding model predicts brain activity as well as or better than individual models.
• Model weights provide interpretable insights into the importance of each feature space for voxel activity.
• Structured variance partitioning effectively addresses correlations between feature spaces.

Conclusions:

• The proposed methods enhance the ability to map brain activity to complex, correlated stimulus properties.
• These techniques offer a robust framework for analyzing brain responses to naturalistic stimuli and neural network features.
• A Python package is released to facilitate the application of these brain mapping approaches.