PrivMeSA is a privacy‑aware self‑evolving multi‑agent system for clinical use. A locally deployed LLM agent can consult more capable remote models when needed, but sending raw patient data risks exposure. PrivMeSA places the disclosure decision with the local agent and uses reinforcement learning to balance task accuracy, direct disclosure risk, and registry‑based re‑identification risk. Privacy is evaluated over the full outbound transcript of each encounter. Remote specialists may request additional information, yet the local agent first queries a "lesson memory" that distills completed consultations into generalized clinical guidance and retrieves relevant lessons for similar future cases, avoiding another remote exchange of sensitive data. The memory grows without extra outcome labels or parameter updates. On an emergency‑department benchmark built from MIMIC‑IV‑ED records, PrivMeSA improves mean task accuracy by up to 15.8 percentage points over plain delegation, reduces personal detail disclosure from 98.0% to 0.2%, and drops the share of cases where a patient can be narrowed to ten or fewer registry entries from 74% to 0%.
Review