Gaining Brain Insights by Tapping into the Black Box: Linking Structural MRI Features to Age and Cognition using...

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Global interpretability in machine learning (ML) offers potential for neuroimaging insights.
• High dimensionality and correlations in neuroimaging data necessitate careful selection of interpretability methods.
• Existing methods like SHAP (SHapley Additive exPlanations) often assume feature independence.

Purpose of the Study:

• To evaluate various global interpretability techniques for neuroimaging data.
• To assess methods accounting for feature dependence and inherently global approaches.
• To apply these methods to predict age and fluid intelligence using UK Biobank data.

Main Methods:

• Trained XGBoost models on UK Biobank neuroimaging measures.
• Applied global interpretability techniques including SHAP and SAGE (SHapley Analysis for General Explanation).
• Compared methods with varying assumptions on feature dependence.

Main Results:

• Mean intensities in subcortical regions strongly correlate with brain aging.
• Hippocampus, cerebellum, frontal, and temporal lobes are key predictors of fluid intelligence.
• Identified distinct neuroimaging patterns associated with aging and cognitive function.

Conclusions:

• Interpretable ML methods are valuable for understanding brain function from neuroimaging data.
• Data-driven approaches using global interpretability enhance insights into brain aging and cognition.
• The study highlights the importance of considering feature dependence in neuroimaging interpretability.