Toward a unified framework for interpreting machine-learning models in neuroimaging.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Machine learning (ML) is increasingly used in neuroscience to model brain function and behavior.
• ML models are often complex and difficult to interpret, hindering neuroscientific validation.
• Interpretability is crucial for understanding how ML models relate to brain mechanisms.

Purpose of the Study:

• To present a unified framework for assessing the interpretability of neuroimaging-based ML models.
• To provide practical tools and examples for model interpretation.
• To enhance the neuroscientific validity and understanding derived from ML models.

Main Methods:

• A unified framework with model-, feature-, and biology-level assessments.
• Practical tools and analysis examples for functional MRI data.
• Application to multivariate pattern-based predictive models.

Main Results:

• The framework offers complementary assessments for understanding ML model function.
• Provides methods to evaluate if models are based on neurobiological signals rather than confounds.
• Facilitates comprehension of what brain pathways or regions represent specific mental constructs.

Conclusions:

• The protocol aids in building more interpretable neuroimaging ML models.
• Contributes to a cumulative understanding of brain mechanisms and health.
• Interpreting ML models is an ongoing, collaborative process.