A large language model-based self-learning and critical agent framework for multimodal Alzheimer's disease diagnosis

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Medical Diagnostics

Background:

• Alzheimer's disease diagnosis relies on integrating complex, multimodal data.
• Current diagnostic methods can be time-consuming and lack transparency in reasoning.
• Large Language Models (LLMs) offer potential for advanced data analysis and decision support.

Purpose of the Study:

• To develop and evaluate a training-free LLM multiagent framework for simulating team-based Alzheimer's disease diagnosis.
• To enhance the interpretability and generalizability of diagnostic processes using multimodal data.
• To leverage LLMs for transparent and accurate neuropsychological diagnosis.

Main Methods:

• A self-learning and critical multiagent workflow was designed, comprising single-dimension and senior agents with intelligence and experience-retrieval modules.
• The framework was evaluated on 1,362 Alzheimer's Disease Neuroimaging Initiative (ADNI) records across different cognitive statuses.
• No end-to-end fine-tuning or supervision on ADNI labels was performed, ensuring a training-free approach.

Main Results:

• The integrated multiagent framework achieved an 88% F1 score, 88% sensitivity, and 90% precision for classifying cognitively normal, mild cognitive impairment, and Alzheimer's disease.
• The system provided case-level rationales, demonstrating transparent reasoning capabilities.
• Sequential self-learning and critique significantly improved performance compared to early agent iterations.

Conclusions:

• A label-free LLM multiagent approach effectively integrates heterogeneous data modalities for Alzheimer's disease diagnosis.
• The framework demonstrates competitive accuracy and retains diagnostic experience, offering transparent reasoning.
• This approach shows promise for real-world clinical support in complex neuropsychological diagnosis.