Explainable machine learning approach to predict and explain the relationship between task-based fMRI and individual...

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Predicting individual cognitive differences from functional magnetic resonance imaging (fMRI) remains challenging despite extensive research.
• Existing methods often lack transparency in how they utilize brain data for predictions.

Purpose of the Study:

• To develop and evaluate an explainable machine learning (ML) framework for predicting cognitive abilities from task-based fMRI.
• To compare the predictive performance and explainability of ML algorithms against traditional methods.

Main Methods:

• Applied an explainable ML framework to task-based fMRI data from the Adolescent Brain Cognitive Development study (n=3,989).
• Compared 9 ML algorithms against mass-univariate and ordinary least squares (OLS) multiple regression for predicting 12 cognitive abilities.
• Utilized explainability techniques including SHapley Additive explanation, eNetXplorer, Accumulated Local Effects, and Friedman's H-statistic.

Main Results:

• ML algorithms significantly outperformed OLS multiple regression in out-of-sample prediction accuracy.
• The Elastic Net algorithm, a linear and additive ML model, performed comparably or better than nonlinear/interactive ML algorithms.
• Explainability methods revealed benefits of ML over OLS, showing consistency in variable importance and directionality of brain-cognition relationships.

Conclusions:

• An explainable ML framework offers enhanced prediction and interpretability for cognitive abilities using task-based fMRI data.
• This approach provides a more robust understanding of brain-cognition relationships compared to standard methodologies.