Traj-CoA: Multi-Agent LLM System for Lung Cancer Risk Prediction
A team of researchers has introduced Traj-CoA, a multi-agent framework that utilizes large language models (LLMs) for modeling patient trajectories, particularly in predicting lung cancer risk. This innovative system tackles the complexities of lengthy and noisy electronic health records (EHR) by utilizing a series of worker agents that sequentially process EHR data, condensing essential events into a collective long-term memory module known as EHRMem. This module minimizes noise while maintaining a detailed timeline. A concluding manager agent integrates the summaries from the worker agents and the timeline to generate predictions. In a zero-shot one-year lung cancer risk prediction challenge using five-year EHR data, Traj-CoA surpassed baseline performances across four categories, demonstrating clinically relevant temporal reasoning and establishing itself as a reliable method for clinical prediction tasks.
Key facts
- Traj-CoA is a multi-agent system for patient trajectory modeling using LLMs.
- It uses a chain of worker agents to process EHR data sequentially.
- Critical events are distilled into a shared memory module called EHRMem.
- A manager agent synthesizes summaries and timeline for predictions.
- Tested on zero-shot one-year lung cancer risk prediction from five-year EHR data.
- Outperformed baselines from four categories.
- Exhibits clinically aligned temporal reasoning.
- Described as robust and generalizable.
Entities
—