Truth, Trust, and Trouble: Medical AI on the Edge
This work addresses the problem of meeting industry standards for medical AI, particularly for open-source solutions, but it is incremental as it benchmarks existing models without introducing a new method.
The paper tackled the challenge of ensuring factual accuracy, usefulness, and safety in open-source LLMs for medical question answering by benchmarking models like Mistral-7B, BioMistral-7B-DARE, and AlpaCare-13B on over 1,000 health questions. AlpaCare-13B achieved the highest accuracy (91.7%) and harmlessness (0.92), while few-shot prompting improved accuracy from 78% to 85%, though all models showed reduced helpfulness on complex queries.
Large Language Models (LLMs) hold significant promise for transforming digital health by enabling automated medical question answering. However, ensuring these models meet critical industry standards for factual accuracy, usefulness, and safety remains a challenge, especially for open-source solutions. We present a rigorous benchmarking framework using a dataset of over 1,000 health questions. We assess model performance across honesty, helpfulness, and harmlessness. Our results highlight trade-offs between factual reliability and safety among evaluated models -- Mistral-7B, BioMistral-7B-DARE, and AlpaCare-13B. AlpaCare-13B achieves the highest accuracy (91.7%) and harmlessness (0.92), while domain-specific tuning in BioMistral-7B-DARE boosts safety (0.90) despite its smaller scale. Few-shot prompting improves accuracy from 78% to 85%, and all models show reduced helpfulness on complex queries, highlighting ongoing challenges in clinical QA.