LLM-based feature selection and counterfactual explanations applied to functional connectivity analysis in...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Psychiatric Disorders

Background:

• Schizophrenia (SZ) poses challenges due to unclear neural mechanisms.
• High-dimensional functional connectivity (FC) data complicates feature selection and interpretation.
• Traditional methods lack integration of neuroscience knowledge, limiting clinical relevance.

Purpose of the Study:

• To develop an innovative framework for SZ analysis using LLM-guided FC feature selection.
• To integrate LLM-encoded brain disease knowledge for biologically plausible feature selection.
• To enhance model interpretability and clinical relevance through counterfactual explanations.

Main Methods:

• A novel framework combining LLM-guided feature selection and counterfactual explanation for FC data.
• Leveraging LLM knowledge for dimensionality reduction of high-dimensional FC data.
• Generating causal intervention examples and translating them into natural language explanations.

Main Results:

• Validation on five real-world SZ datasets.
• Demonstrated improvement in model classification performance for SZ.
• Provided novel insights into SZ analysis through enhanced feature selection and interpretability.

Conclusions:

• The LLM-based FC analysis method effectively screens key FC features for brain regions in SZ.
• The approach offers improved feature selection and interpretability for SZ datasets.
• Limitations include clinical application challenges due to data heterogeneity and unoptimized hyperparameters.