medRxivSource publication:
Loading content…
A multi-agent workflow under investigator direction carried a large electronic health record study from cohort construction and executable statistics through data-quality checks, literature retrieval and reporting, using sequential sacituzumab govitecan (SG) and trastuzumab deruxtecan (T-DXd) orders in breast cancer as the clinical case. From 4.95 million patients, 106 had orders for both drugs, 73 with T-DXd first and 33 with SG first; median second-drug record spans were 63 and 49 days, a median difference of −14 days (95% CI, −41 to 21), descriptive and unadjusted.
An end-to-end electronic health record research workflow ran from clinical question through cohort, statistics, data-quality assessment, evidence context and a draft manuscript, with sequential orders for two antibody-drug conjugates in breast cancer as the use case. Such work has typically split cohort definitions, executable analysis, data-quality review, evidence synthesis and reporting across separate steps; agentic methods were recently applied to EHR-informed trial design, and producing an observational oncology study is a complementary application. The workflow ran on the June 4, 2026 Protégé release, which contains 4.95 million patients including 156,051 with breast cancer codes, and the rule requiring orders for both drugs yielded 106 patients.
Muse Spark operated inside a Symbiotic Agent Teaming environment with tools for code execution, literature and registry retrieval and file editing, generating and running cohort, endpoint and statistical code and producing data-quality assessments, figures and manuscript text. The human side held tiered roles: investigators set the clinical question and analysis direction through Human-Guided Orchestration and served as Expert-in-the-Loop reviewers of code and evidence, adjudicating revisions and editing all outputs. Expert-in-the-Loop review replaced the agent's originally proposed filtered time-on-treatment endpoint with unfiltered first-to-last order spans and removed clinical labels whose validity was not established; the original analyst confirmed the reported medians, interquartile ranges, confidence intervals and tests matched the amended definition.
Among the 106 patients, median second-drug record spans were 63 days for T-DXd then SG and 49 days for SG then T-DXd, a median difference of −14 days with P = 0.42 for the distributional comparison. A span runs from the first to the last qualifying order for that drug, counts a single dated order as zero days, adds no dosing interval and applies no time-to-event censoring; it does not confirm administration or identify progression or discontinuation. None of six secondary comparisons had a Holm-adjusted P value below 0.05; further systemic therapy was recorded in 64% and 52% of the groups, a difference of −13 percentage points (95% CI, −32 to 7).
The demonstration covers one breast cancer use case on one database release: the June 4, 2026 Protégé release, drawing routine-care records from 39 medium-to-large US health systems linked to open claims, with 2020 to 2024 data analyzed. Intended readers are clinical researchers and engineers following EHR research methods and agentic workflows. The source proposes next evaluating whether this division of work scales, extending to larger clinically curated datasets with validated receptor status, treatment episodes, progression, mortality and reasons for discontinuation, and comparing complete workflows using matched questions, independently checked source records and prospectively recorded resource use.
Orders do not confirm administration and may be planned, revised, duplicated or canceled; spans include gaps between orders and may be truncated by the end of records. Patients were selected for having orders for both drugs, and receptor status, treatment line and performance status were not validated, so group differences cannot be read as sequencing effects. The source states that research time and resource use were not measured, no matched human-only workflow was evaluated, and AI accuracy and independent reproducibility were not benchmarked, so no quantitative efficiency advantage is claimed.