Multimodal LLM pipeline automates patient-trial matching with 87% real-world accuracy.
A new multimodal LLM-powered pipeline has been validated for automating patient-trial matching in healthcare using unprocessed documents from electronic health records (EHRs). The system leverages the latest reasoning-LLM capabilities to assess complex eligibility criteria and uses visual LLM features to interpret medical records without lossy image-to-text conversion. Multimodal embeddings enable efficient search across text and images, streamlining the matching process.
On the n2c2 dataset, the pipeline achieved a state-of-the-art 93% criterion-level accuracy. In real-world clinical trials, it maintained an 87% accuracy rate, with users reviewing eligibility in under 9 minutes per patient—80% faster than traditional manual chart reviews. The main limitation remains the challenge of replicating nuanced human decision-making when records are incomplete. This pipeline demonstrates the potential of LLMs to automate complex, high-stakes tasks in healthcare, improving efficiency and scalability while reducing administrative burden.
Source: Nature