An LLM-assisted framework for accelerated and verifiable clinical hypothesis testing from electronic health records.

Acquiring insights from electronic health records (EHRs) is slowed by manual analytical workflows that limit scalability and reproducibility. We present LATCH (LLM-Assisted Testing of Clinical Hypotheses), an agentic framework that converts natural language clinical hypotheses into fully auditable analyses on structured EHR data. LATCH integrates LLM-assisted semantic layers with deterministic execution pipelines to automate cohort construction, statistical analysis, and result reporting, while isolating patient-level data from LLM-involved steps. Using diabetes as a model disease, LATCH reproduced findings from 20 published studies within 3-15 minutes per study. Beyond replication, LATCH enabled study extensions and new insight generation through simple natural language hypothesis modifications. We demonstrated LATCH across 102 hypothesis tests spanning reproduction, extension, and insight generation. We systematically stress-tested LATCH to characterize its limitations and operational boundaries. LATCH provides a scalable framework for reproducible real-world evidence generation, reducing analytical bottlenecks and improving reliability of AI-assisted biomedical discovery while preserving human oversight.