Collaborative Multi-Agent Screening Framework for Efficient Eligibility Evaluation of Clinical Trial Patients

Abstract Background Clinical trial patient eligibility screening is a core and labor‑intensive step in clinical research, which traditionally relies on manual review of electronic health records (EHRs) and suffers from high time costs, human errors, and inconsistent screening criteria. With the rapid development of large language models (LLMs), intelligent EHR analysis has become a feasible solution to automate patient recruitment. However, existing single‑LLM inference and single‑agent screening methods often fail to parse complex inclusion and exclusion rules, yielding unstable prediction and substantial false‑positive or false‑negative outputs. Methods This paper proposes a modular multi‑agent collaborative framework built upon LangGraph for clinical trial patient eligibility evaluation. The workflow is decomposed into four specialized agents. LLMs are only used for unstructured text understanding, while eligibility matching adopts deterministic rule‑driven computation to avoid LLM output randomness. Retrospective simulation experiments are conducted on an in‑house programmatically generated synthetic EHR cohort. Results Our multi‑agent framework achieves perfect accuracy, precision, recall and F1‑score on simulation cohort, outperforming the end‑to‑end single‑agent baseline. Conclusions This work provides a proof‑of‑concept intelligent automation paradigm for clinical‑trial patient screening validated on an in‑house synthetic EHR cohort; further real‑world clinical validation is required before practical clinical deployment, and the framework can be further migrated toward multicenter and cross‑border clinical trial recruitment scenarios.