Truth, Trust, and Trouble: Medical AI on the Edge

TL;DR

This paper introduces a benchmarking framework for evaluating medical AI models on factual accuracy, safety, and helpfulness using over 1,000 health questions, revealing trade-offs and challenges in clinical question answering.

Contribution

It provides a comprehensive evaluation of open-source medical LLMs, highlighting their strengths and limitations in safety and accuracy, and demonstrates the impact of prompting techniques.

Findings

AlpaCare-13B achieves 91.7% accuracy and 0.92 harmlessness.

BioMistral-7B-DARE improves safety with domain-specific tuning.

Few-shot prompting increases accuracy from 78% to 85%.

Abstract

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.