Researchers build AI that learns from past cases to predict lung cancer
New system Traj-Evolve outperforms 9 baselines by learning from prior patient cases...
Researchers from the University of Washington and Fred Hutchinson Cancer Center today unveiled Traj-Evolve, a groundbreaking self-evolving multi-agent AI system designed to revolutionize lung cancer early detection. Published on arXiv (cs.AI) on June 1, 2026, this system tackles the critical challenge of modeling patient trajectories from sparse, noisy electronic health records (EHRs) over extended periods.
Traj-Evolve introduces two innovative mechanisms: an Experience Pool (ExPool) that functions as non-parametric memory by indexing reasoning traces from prior patients, and a multi-agent reinforcement learning (MARL) framework that optimizes inter-agent collaboration. In testing against five years of multimodal EHR data, Traj-Evolve outperformed nine strong baselines across overall populations and a particularly challenging never-smoker subgroup. The system demonstrates complementary benefits—ExPool improves specificity while MARL enhances sensitivity—achieved through a unique leave-one-out cross-retrieval strategy that aligns training and inference behavior.
- Traj-Evolve outperforms 9 baselines in lung cancer prediction using up to 5 years of multimodal EHR data
- Combines non-parametric memory (ExPool) with MARL for improved specificity and sensitivity
- System dynamically learns from prior patient cases, mirroring clinical experience accumulation
Why It Matters
Could significantly improve early lung cancer detection accuracy while reducing false positives/negatives in high-risk patients.