BrainPrompt+: Multi-Level Brain Prompt Learning for Knowledge-Guided Neurological Disorder Identification.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Neurological disorders like Alzheimer's, Parkinson's, and Autism Spectrum Disorder present diagnostic challenges due to subtle symptoms and complex brain activity.
• Resting-state functional MRI (rs-fMRI) and Graph Neural Networks (GNNs) offer potential for disease classification but face limitations in graph construction, domain knowledge integration, and metadata fusion.

Purpose of the Study:

• To develop a novel knowledge-guided framework, BrainPrompt+, that overcomes limitations of existing GNN-based methods for neurological disorder identification.
• To integrate large language models (LLMs) with multi-level natural language prompts to unify imaging, clinical, and external knowledge.

Main Methods:

• Proposed BrainPrompt+, a framework integrating LLMs with five types of natural language prompts: spectral, spatial, ROI, disease, and subject.
• Encoded prompts using a frozen LLM and incorporated them into a GNN pipeline for semantically enriched and interpretable brain network analysis.
• Validated the framework on three rs-fMRI datasets.

Main Results:

• BrainPrompt+ consistently outperformed state-of-the-art baselines, achieving accuracy improvements of up to 8.93%.
• Biomarker analysis confirmed the model's interpretability, with highlighted Regions of Interest (ROIs) aligning with established neuroscience findings.

Conclusions:

• BrainPrompt+ establishes a flexible and generalizable paradigm for knowledge-guided brain network analysis in neurological disorders.
• The framework offers a promising approach for more accurate and interpretable diagnosis of conditions like Alzheimer's and Parkinson's disease.