BloodGPT: Prefetch-and-execute agent for MedAgentBench clinical EHR tasks

Read full text Summarise with VIRAT
Authors: Vasilii Lazuka, Mykyta Yastreb
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Source code for the BloodGPT prefetch-and-execute agent described in: Lazuka, V. & Yastreb, N. "Prefetch-and-execute architecture achieves 100% success rate on MedAgentBench clinical EHR tasks." Nature Communications (2026, submitted).The agent separates FHIR data retrieval from clinical reasoning by pre-loading all patient resources into memory before the LLM engages, then constraining computation to sandboxed Python code with a structured question decomposer guiding the execution loop. A transport-layer FHIR R4 sanitizer canonicalizes LLM-generated payloads before POST. The architecture contains no task-type-specific prompts or hardcoded clinical knowledge — all codes, thresholds, and drug names are read from the task context at runtime. Evaluated on MedAgentBench v1 (300 physician-authored tasks, FHIR R4 server, 100 patients, 785,207 records). Achieves 100.0% success rate across 5 independent runs (1,500 consecutive tasks) using Gemini 3 Flash (agent) and Gemini 3 Pro (decomposer). Requires a running MedAgentBench HAPI FHIR server: https://github.com/stanfordmlgroup/MedAgentBench https://github.com/ericoericochen/medagentbenchv2 BloodGPT

Fields of Study & Subjects

Artificial Intelligence in Healthcare and Education Machine Learning in Healthcare Electronic Health Records Systems

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
28 March 2026
Language
EN