New LLM-powered pipeline automates patient-trial matching using multimodal medical records.
A new multimodal LLM-powered pipeline has been developed to automate patient-trial matching using unprocessed documents from electronic health records (EHRs). The system leverages reasoning-focused LLMs to interpret complex eligibility criteria, visual capabilities to analyze medical images directly, and multimodal embeddings for efficient record search. This approach eliminates the need for lossy image-to-text conversions and custom integrations, enabling scalable deployment across healthcare sites.
The pipeline was validated on both a benchmark dataset (n2c2 2018, 288 diabetic patients) and a real-world cohort (485 patients, 36 trials). It achieved a state-of-the-art 93% criterion-level accuracy on the benchmark and 87% in real-world settings, with eligibility reviews taking under 9 minutes per patient—80% faster than manual chart reviews. The main limitation is the difficulty of replicating human judgment when records lack sufficient detail. Nevertheless, this pipeline demonstrates robust, scalable automation for clinical trial matching, with potential to significantly reduce administrative burden and accelerate patient enrollment.
Source: Nature