Privacy-Preserving LLM for Dermatology Data Retrieval

A study published on arXiv evaluated a locally deployed, privacy-preserving small language model (SLM) for retrieving structured clinical features from longitudinal dermatology records of pemphigus patients. Thirty pemphigus patients contributed 541 visit notes aggregated into full records totaling 89,336 words. Two expert dermatologists annotated 56 clinically relevant features. The SLM, Qwen3 4B Thinking 2507, was queried to retrieve these features and generate summary reports. Across 1,680 feature retrieval tasks, mean accuracy was assessed. The approach aims to reduce clinician workload and mitigate the risk of missing critical historical information during routine follow-up visits.