A weighted cluster kernel PCA prediction model for multi-subject brain imaging data.

Area of Science:

• Neuroimaging
• Machine Learning
• Biostatistics

Background:

• Brain imaging data show potential for predicting psychiatric conditions and cognitive functions.
• Challenges in prediction include high dimensionality, noisy data, and inter-subject variability.
• Existing methods often overlook voxel clustering and subject-specific differences.

Purpose of the Study:

• To develop an advanced predictive model for brain imaging data.
• To address limitations of current dimension reduction techniques like PCA.
• To incorporate neuroanatomic structures and subject heterogeneity into prediction models.

Main Methods:

• Proposed a weighted cluster kernel PCA predictive model.
• Voxel data were clustered using neuroanatomic parcellation or data-driven approaches.
• Extracted cluster-specific principal features and employed a weighted estimation method.

Main Results:

• The proposed method effectively handles high dimensionality and complex data structures.
• It allows for assessing the importance of different brain regions in prediction.
• The model captures nonlinear relationships and mitigates overfitting for outlier subjects.

Conclusions:

• The weighted cluster kernel PCA model offers a robust approach for brain imaging prediction.
• It enhances predictive accuracy by leveraging clustered voxel information and subject weighting.
• Demonstrated effectiveness through simulations and a real fMRI data analysis.