CARE: Privacy-Compliant Agentic Reasoning Splits Local and Remote Models for Healthcare Decision-Making

arXiv·medium signalXBlueskyLinkedInCopy link

CARE introduces a hybrid architecture where sensitive patient data is processed locally by open-source models while remote proprietary models provide high-level reasoning guidance without accessing raw data, enabling robust decision-making under evidence discordance while preserving privacy. The paper introduces MIMIC-DOS, a benchmark derived from MIMIC-IV for cases where patient-reported symptoms contradict clinical signs. Baseline single-pass LLMs and existing agentic pipelines collapse to degenerate one-class predictions on discordant evidence — CARE's multi-stage approach avoids this failure mode.