Researchers develop HypEHR, a lightweight AI model that answers medical record questions as efficiently as large language models while using a fraction of the computing power

Researchers at an unnamed institution published HypEHR, an AI model designed to answer questions about patient electronic health records (EHR — the digital files doctors keep on patients). Unlike current systems that rely on large language models (LLMs — AI systems trained to understand and generate text), HypEHR uses a compact mathematical approach called hyperbolic geometry that mirrors how medical codes and patient histories actually relate to each other.

HypEHR achieves performance comparable to LLM-based systems on standard medical question-answering tests while using significantly fewer computational resources — meaning hospitals can run it on cheaper, smaller servers. The model was pretrained to predict what diagnoses a patient will need at their next visit, which helps it better understand the medical hierarchy embedded in diagnosis codes (like ICD codes used by all healthcare systems).

For hospital IT teams and healthcare organizations, this matters because deploying AI on medical records is currently expensive and slow — every query sent to a cloud-based LLM costs money and adds latency (delay). A lightweight alternative that runs locally could reduce both costs and response times, making AI-powered search through patient records feasible for smaller healthcare providers that can't afford premium LLM services.

The code is open-source and available on GitHub, so any developer or healthcare organization can download and test it immediately without waiting for a commercial release.