Answers are grounded in this paper and cite their sources.
Shows improved clinical decision support in diabetes management, suggesting enhanced AI capabilities.
Introduction and Objective: Large language models (LLMs) show limited subspecialty accuracy in type 2 diabetes (T2D) management. Current cloud-based medical artificial intelligence (AI) systems have privacy concerns and lack subspecialty expertise. We developed OLYMPUS (On-premise Local LLM for Medical Practice Under Secure environment), a privacy-preserving framework enabling knowledge distillation from expert teacher models to deployable student models using guideline-based synthetic data. Methods: Sophia (Llama 70B) was fine-tuned on the clinical data of 80 subjects with T2D to create an expert teacher model. Chiron generated 744 synthetic T2D cases, and fine-tuned Sophia generated expert-level solutions for all 744 cases. One endocrinologist validated each case-solution pair. Asclepius, a Qwen 14B model, was fine-tuned via low-rank adaptation on these 744 validated pairs using consumer hardware (Mac mini M4 Pro, 64GB RAM) with complete local processing. Performance was compared with GPT-4o through blinded evaluation of 40 randomly selected cases by three independent endocrinologists for treatment appropriateness, quality, guideline adherence, and clinical reasoning. Results: Asclepius significantly outperformed GPT-4o (4.08±1.00 vs 3.88±1.00, p=0.002), showing 5.3% improvement. Asclepius consistently exceeded GPT-4o across all domains: treatment appropriateness (+5.4%), treatment quality (+5.5%), guideline adherence (+4.2%), and clinical reasoning (+5.9%). All three evaluators independently rated Asclepius higher. Conclusion: In conclusion, knowledge distillation from expert teacher to deployable student model using guideline-based synthetic data achieved superior performance compared to state-of-the-art general medical AI while maintaining complete patient privacy through local processing.
No takes yet. Share an insight, caveat, or question.
PARK et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: