Lightweight Retrieval-Augmented Generation and Large Language Model-Based Modeling for Scalable Patient-Trial Matching

A team of researchers from Mayo Clinic and multiple universities has published a paper on arXiv proposing a lightweight framework that combines retrieval-augmented generation (RAG) with large language models (LLMs) to solve the computationally expensive problem of matching patients to clinical trials. The work, tested on four public benchmarks (n2c2, SIGIR, TREC 2021/2022) and Mayo Clinic's own multimodal dataset (MCPMD), shows that their pipeline achieves accuracy comparable to full end-to-end LLM methods while using a fraction of the compute resources.

The framework separates the matching process into two stages: first, a RAG module identifies and extracts only the clinically relevant segments from long, heterogeneous electronic health records (EHRs), dramatically reducing input complexity. Then, frozen LLMs encode these segments into informative representations, which are further compressed via dimensionality reduction before being passed to lightweight classifiers. The authors note that frozen LLMs work well for structured clinical data, but fine-tuning remains necessary for unstructured clinical narratives. This design allows the system to scale to large patient populations without the prohibitive cost of processing entire EHRs through an LLM, making it practical for real-world hospital deployment.