Traj-Evolve: A Self-Evolving Multi-Agent System for Patient Trajectory Modeling in Lung Cancer Early Detection
Traj-Evolve introduces a self-evolving multi-agent system that models patient trajectories from longitudinal electronic health records for lung cancer early detection. The system combines an Experience Pool for retrieval-augmented few-shot learning with multi-agent reinforcement learning to optimize collaboration, outperforming nine baselines on both general and never-smoker populations.
Traj-Evolve addresses a critical challenge in clinical AI: reasoning over sparse, noisy, and extended temporal sequences from patient records while learning from accumulated clinical experience. Traditional LLM-based approaches handle long contexts but process patients in isolation, missing the diagnostic insights clinicians gain from comparing cases with similar histories. This work bridges that gap through a dual-mechanism architecture that mirrors human clinical reasoning.
The system's innovation lies in combining non-parametric memory (Experience Pool) with parametric optimization (MARL). The Experience Pool indexes high-quality reasoning traces for rapid retrieval of similar patients, while reward-ranked fine-tuning improves agent coordination. The leave-one-out cross-retrieval strategy elegantly aligns training and inference behaviors, addressing a common gap in retrieval-augmented systems.
For the healthcare AI sector, Traj-Evolve demonstrates that complementary mechanisms can address distinct diagnostic objectives: retrieval improves specificity (reducing false positives) while reinforcement learning improves sensitivity (reducing false negatives). This nuanced performance trade-off is clinically significant, particularly for early cancer detection where both misses and overdiagnosis carry substantial costs.
The analysis revealing that optimal retrieval shifts from diverse to specific samples as the Experience Pool expands offers practical insights for deployment. Healthcare institutions implementing similar systems should anticipate that population-level performance dynamics evolve as case libraries grow. Future work likely involves scaling to multi-condition diagnosis, integrating imaging data more deeply, and validating on prospective patient cohorts rather than historical records.
- →Traj-Evolve combines an Experience Pool and multi-agent reinforcement learning to model patient trajectories from five years of multimodal EHR data
- →The system outperforms nine baselines on lung cancer prediction in both general and never-smoker populations through retrieval-augmented few-shot learning
- →Experience Pool retrieval primarily improves diagnostic specificity while MARL optimization improves sensitivity, enabling tunable performance trade-offs
- →Optimal retrieval strategy shifts from diverse to specific patient samples as the Experience Pool expands, suggesting dynamic deployment considerations
- →Leave-one-out cross-retrieval alignment unifies training and inference-time behavior, addressing a common limitation in retrieval-augmented systems