BrainAgent: LLM framework analyzes brain networks with 4x better reasoning

Brain network analysis is critical for understanding cognition and disorders like Alzheimer's, but traditional deep learning methods treat it as a black-box classification problem. A team from (unknown institution) introduces BrainAgent, an agentic LLM framework that reformulates connectome classification as an iterative process. It converts raw brain networks into compact multi-level structural descriptions using brain-specific analysis tools, then retrieves relevant neuroscience knowledge and task-specific cases to ground reasoning, and finally generates structured predictions with reflective verification. On four public rs-fMRI datasets, BrainAgent consistently boosts performance of both closed-source (e.g., GPT-4) and open-source LLMs over direct prompting and standard reasoning baselines, producing more comprehensive and verifiable outputs.

The framework addresses the structure-language gap and overconfident predictions plaguing general-purpose LLMs on brain data. Ablation studies confirm each component—topology understanding, external retrieval, and reflection—contributes to accuracy and interpretability. By enabling LLMs to reason about connectomes with grounded neuroscience knowledge, BrainAgent offers a practical route toward transparent, knowledge-intensive brain network analysis, potentially aiding diagnosis and discovery of neurological biomarkers. This work bridges graph neural networks and LLMs, opening new frontiers for AI in neuroscience.