Unsupervised mapping of causal relations between brain lesions and behavior

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• Human lesion studies are crucial for understanding brain-behavior causality.
• A key challenge is determining the appropriate granularity for analyzing brain regions and behaviors.
• Existing methods struggle to identify optimal analysis levels and uncover complex lesion-behavior relationships.

Purpose of the Study:

• To introduce Causal Feature Learning (CFL), a data-driven approach for lesion-behavior mapping.
• To enable the simultaneous learning of analysis granularity and lesion-behavior associations.
• To discover novel, cross-cutting lesion-behavior maps without pre-specifying brain regions or outcomes.

Main Methods:

• Developed and applied the Causal Feature Learning (CFL) framework.
• Utilized simulated datasets to test CFL's robustness compared to Canonical Correlation Analysis (CCA).
• Applied CFL to a large dataset of human lesion subjects, including language, visuospatial, and depression symptom data.

Main Results:

• CFL successfully recovered lesion-behavior maps in simulated data where CCA failed.
• Validated CFL by identifying known lesion-behavior maps for language and visuospatial deficits.
• Demonstrated CFL's capability to uncover new associations between lesions and depression symptoms.

Conclusions:

• CFL offers a powerful, data-driven method for advancing lesion-behavior mapping.
• The approach overcomes limitations of traditional methods by learning optimal analysis levels.
• CFL facilitates the discovery of new insights into how brain lesions impact cognitive functions and emotional states.