Category guidance
A proprietary model proposes structured categories and state transitions. It never sees sensitive patient values.
Healthcare · Privacy-compliant agents
CARE
A multi-stage agentic framework that reasons over conflicting clinical evidence without sending sensitive patient data to a remote model.
Evidence on device
A local model uses that guidance to acquire evidence from the patient record and keep private features on-site.
Discordance resolved
The two sides combine signs and symptoms that disagree, then decide whether organ dysfunction is worsening.
Abstract
LLM systems typically degrade when evidence is internally inconsistent. In healthcare, patient-reported symptoms can contradict medical signs. CARE studies this setting through MIMIC-DOS, an ICU dataset derived from MIMIC-IV that contains only cases with sign–symptom discordance.
Single-pass LLMs and ordinary agentic pipelines struggle to reconcile those signals. CARE splits the work: remote guidance without patient data, local evidence acquisition, and a final privacy-compliant decision. Empirically, it is stronger across key metrics than multiple baselines while keeping sensitive values off the remote model.