VirtualTrialHarness arXiv:2611.05551 Achieves 95.1% Reference-Matched Cohort Outputs
VirtualTrialHarness uses agent coordination and explicit contracts to operationalize virtual trial protocols, delivering 95.1% reference-matched outputs. Results on synthetic environments, Mayo cases, and MIMIC-IV tasks indicate reduced drift in source selection and cohort construction. The approach supports consistent virtual trial execution across heterogeneous clinical data representations.
VirtualTrialHarness coordinates procedural skills and explicit contracts across workflow stages for virtual trial operationalization. It selects reusable skills for source selection, patient linkage, and temporal interpretation while enforcing consistent study decisions. The framework was evaluated on VTBenchmark and deployed within the Mayo Clinic Virtual Trials Program for investigator review of cohort construction.
Across three complex synthetic environments VTHarness reached 95.1% success versus 60.4% for retrieval from the same procedural corpus. Procedural skills improved source recovery for three language models and raised source overlap with analyst code in 13 Mayo Clinic cases. Site-specific adaptation produced reference-matched execution on eight MIMIC-IV tasks.
The 34.7-point gap demonstrates that contract-mediated agent coordination reduces representation drift in complex clinical schemas. Operational impact appears in faster investigator validation cycles and reduced manual reconciliation of local data conventions. This pattern aligns with prior agent frameworks that separate decision contracts from retrieval in regulated domains.
Next steps include scaling the contract layer to additional EHR systems and measuring endpoint concordance against traditional trial analyses.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2611.05551)
- [2]Supporting Source(https://physionet.org/content/mimiciv/2.2/)