CARE-AD: a multi-agent large language model framework for Alzheimer's disease prediction using longitudinal clinical...

Area of Science:

• Artificial Intelligence in Medicine
• Computational Neuroscience
• Clinical Informatics

Background:

• Large language models (LLMs) demonstrate potential across various fields but have limited application in complex clinical prediction.
• Early prediction of Alzheimer's disease (AD) is crucial for timely intervention and management.

Purpose of the Study:

• To introduce CARE-AD, a novel multi-agent LLM framework for forecasting Alzheimer's disease onset.
• To evaluate the efficacy of CARE-AD in analyzing longitudinal electronic health record (EHR) notes for early AD risk assessment.

Main Methods:

• Developed CARE-AD, a framework utilizing specialized LLM agents for extracting AD-relevant signs and symptoms from EHR notes.
• Emulated a collaborative diagnostic process by assigning domain-specific evaluation tasks to individual LLM agents.
• Conducted a retrospective evaluation comparing CARE-AD against baseline single-model approaches.

Main Results:

• CARE-AD achieved higher accuracy (0.53) in predicting AD risk 10 years prior to diagnosis compared to baseline models (0.26-0.45).
• The multi-agent system demonstrated superior performance in analyzing longitudinal EHR data for AD risk forecasting.

Conclusions:

• Multi-agent LLM systems are feasible for supporting early Alzheimer's disease risk assessment.
• CARE-AD's performance highlights the potential of integrating advanced AI into clinical decision support workflows for neurodegenerative diseases.