MedGuard-Referral: Minimum-Necessary Disclosure in Clinical Referral Summarization

Clinical referral summaries should communicate the information a receiving specialist needs without exposing unrelated patient details. Standard de-identification addresses names, contact details, and other direct identifiers, but it does not decide whether a correct clinical or social fact is necessary for a particular referral. We study this problem as minimum-necessary disclosure in clinical referral summarization. We introduce MedGuard-Referral, a structured pipeline that extracts candidate facts from a source note, validates the intermediate representation, filters facts according to specialty-specific relevance and privacy criteria, and generates the final summary from the retained facts. We evaluate five systems on 100 synthetic English referral cases spanning cardiology, dermatology, endocrinology, neurology, and psychiatry, using a frozen 70/15/15 development, validation, and sealed-test split. All systems use one local model configuration, qwen3:8b-q4\_K\_M through Ollama, and are scored with a deterministic evaluator rather than an LLM judge. On the 15 sealed-test cases, MedGuard achieved a trustworthy score of 0.871, compared with 0.793 for post-filtering, with a mean paired difference of 0.078, a 95% bootstrap interval of $[0.001459, 0.162932]$, and a preregistered one-sided paired sign-flip value of $p=0.042$. No direct-identifier leaks were detected for any system. Sensitive-information leakage occurred in five vanilla cases, seven privacy-prompted cases, four pre-redaction cases, two post-filtering cases, and no MedGuard cases. These results provide preliminary benchmark evidence that deciding which facts may enter a referral summary before generation can reduce unnecessary disclosure. The study remains limited by synthetic English data, a 15-case sealed test, one local model, deterministic lexical evaluation, no clinician review, and no real patient records.