Embedding Anatomical or Functional Knowledge in Whole-Brain Multiple Kernel Learning Models.

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Pattern recognition models are widely used in neuroimaging for cognitive and clinical research.
• Current models often lack the integration of prior knowledge about brain structure and function.
• Incorporating anatomical or functional brain region information can enhance model performance.

Purpose of the Study:

• To present a novel approach using group sparsity to model whole-brain patterns as combinations of regional patterns.
• To leverage Multiple Kernel Learning (MKL) to learn the contribution of predefined brain regions to predictive models.
• To investigate the impact of brain region grouping on model generalisation and identify relevant regions.

Main Methods:

• Developed a sparse Multiple Kernel Learning (MKL) framework.
• Modeled the whole brain multivariate pattern as a combination of regional patterns.
• Utilized brain atlases to define anatomical or functional regions for grouping.

Main Results:

• The proposed MKL approach demonstrated improved generalisation performance when atlas-based grouping aligned with the data.
• The method successfully identified subsets of relevant brain regions for predictive modeling.
• Comparative analysis using three atlases and three datasets validated the effectiveness of the grouping strategy.

Conclusions:

• The novel group sparse MKL approach effectively integrates prior knowledge of brain regions into neuroimaging pattern recognition.
• This method offers improved predictive accuracy and interpretability by highlighting key brain regions.
• The approach is implemented in the open-source Pattern Recognition for Neuroimaging Toolbox (PRoNTo).