CohortFlow cuts clinical cohort correction latency by up to 64.8%
Researchers report that their clinical cohort curation agent reduced correction time and model calls while maintaining similar requirement accuracy to full regeneration.
A language-model agent that updates only the affected parts of a clinical cohort specification reduced correction latency by up to 64.8% in tests, according to an arXiv preprint that has not been peer reviewed. The researchers designed CohortFlow to handle changes to eligibility criteria without repeatedly processing requirements that remain unchanged. It converts natural-language corrections into localized updates and uses recorded dependencies to identify which computations need to be rerun.
In 100 scenarios derived from clinical-trial eligibility criteria, covering 126 correction turns, CohortFlow achieved 77.0% requirement micro-F1, compared with 76.4% for full regeneration using the same 8B model. The difference was not statistically significant (p=0.493). The authors report that mean correction latency fell by 64.8% after completion and 44.3% during execution, while model calls decreased by 38.1%.
In a separate proof-of-concept evaluation using MIMIC, all 13 cohort states executed successfully. The authors say the correction trajectories preserved unaffected computations and recomputed changed or dependent ones. The results test localized steering on specified scenarios; the preprint does not establish clinical deployment performance.