Privacy-Aware Distillation of Large Language Models for Enhanced Multimorbidity Scoring

A research team developed a new method to safely use large language models (LLMs) in healthcare without compromising patient data privacy. The method involves transferring clinical knowledge from advanced LLM models into smaller, compact models using synthetic data that preserves characteristics of real patients from the UK Biobank without exposing actual personal information. This approach achieved high-quality knowledge transfer with a correlation of 0.75-0.89. Multimorbidity scores created this way improved patient survival prediction with an index up to 0.91 and showed higher genetic heritability (approximately 0.05). The research demonstrates a secure and legally compliant way to apply large language models in large-scale healthcare applications.