CohortFlow records 77.0% micro-F1 on 126 correction turns with 38.1% fewer calls than full regeneration
CohortFlow achieves statistically equivalent accuracy to full regeneration while cutting correction latency 64.8% and model calls 38.1% through localized dependency tracking. The 100-scenario, 126-turn evaluation and MIMIC proof-of-concept confirm selective recomputation preserves valid prior results. The method targets the dominant cost of repeated full reprocessing in evolving clinical cohort definitions.
The arXiv preprint 2610.36100 introduces CohortFlow, an agent that stores explicit semantic dependencies between eligibility criteria and cohort operations. Natural-language corrections trigger selective recomputation rather than full re-execution of the graph. On 100 derived scenarios the system completed 126 correction turns while preserving unaffected subgraphs.
Localized steering produced 77.0% requirement micro-F1 against 76.4% for full regeneration with the same 8B model (p=0.493). Mean correction latency fell 64.8% post-completion and 44.3% during execution; model calls dropped 38.1%. A MIMIC proof-of-concept executed all 13 cohort states without error and demonstrated node-level preservation of prior results.
The approach directly addresses repeated full scans of static criteria that dominate current clinical cohort pipelines. By maintaining an auditable dependency graph it reduces both compute cost and the risk of unintended drift when criteria evolve. Operational impact appears largest in longitudinal studies where eligibility rules are revised quarterly or more often.
Next steps include integration with live EHR query engines and measurement of end-to-end cohort refresh times on production workloads exceeding 10 000 patients.
CohortFlow: Production MIMIC-IV refresh latency drops below 30% of baseline on at least 50 live queries within 12 months of open-source release.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2610.36100)
- [2]Supporting Source(https://physionet.org/content/mimiciv/3.1/)