MELMA-Q: A Clinical-Grade Framework for Safety Assessment of Medical Large Language Model Answers

MELMA-Q is a safety assessment framework for answers generated by medical large language models (LLMs). It includes a 30-item clinician rating questionnaire covering seven dimensions: accuracy, reasoning ability, safety, clarity, comprehensibility, practicality, and response behavior. Its purpose is to fill the gap where traditional automatic evaluation metrics fail to capture the safety dimensions of medical responses.

With the widespread application of large language models (LLMs) in scenarios like medical consultation and health Q&A, the reliability of AI-generated medical advice has become a core issue. Traditional automatic evaluation metrics (e.g., BLEU, ROUGE) cannot capture the crucial safety dimensions of medical responses. Grammatically fluent but medically inaccurate answers pose significant potential harm to patients. The MELMA framework introduces the professional perspective of clinicians to systematically assess the quality and safety of AI medical responses from seven key dimensions.

The 30 assessment items in the MELMA-Q questionnaire are distributed across seven dimensions:

1. Accuracy: Medically factual, consistent with current medical consensus, no contradictory content;
2. Reasoning ability: Demonstrates a clear chain of clinical thinking, correctly links symptoms to causes, and reasoning aligns with medical logic;
3. Safety: No harmful advice, includes safety warnings (e.g., drug interactions, contraindications), and provides appropriate handling suggestions for emergency situations;
4. Clarity: Clear organizational structure, prominent key information, no confusing expressions;
5. Comprehensibility: Language suitable for the user's health literacy level, terms are explained, and sentence structures are not complex;
6. Practicality: Provides actionable advice, includes specific guidance (e.g., medication dosage, timing of medical visits), and responds to questions in a targeted manner;
7. Response behavior: Recognizes the scope of capabilities, advises users to seek professional medical help, and responds cautiously to uncertain questions.

The core innovation of MELMA-Q lies in the introduction of clinicians' professional judgment. Its advantages include: identifying subtle medical errors, evaluating the clinical rationality of suggestions, judging the potential impact of responses on patient safety, and identifying implicit biases or inappropriate assumptions in model responses. The rating uses a standardized 30-item questionnaire, with clear scoring criteria for each item to reduce subjective bias.

For medical AI developers: Provides a systematic evaluation tool to help identify model weaknesses and make targeted improvements; For medical institutions and regulatory bodies: Provides a reproducible evaluation method for comparing the performance of different medical AI products or monitoring changes in versions of the same product; For researchers: The seven dimensions can serve as starting points for research hypotheses to explore the impact of model architecture, training data, or fine-tuning strategies on specific capability dimensions.

Limitations: Currently relies mainly on manual scoring, which may become a bottleneck for large-scale evaluations; Outlook: Develop automated auxiliary scoring tools, establish a larger network of clinician evaluators, convert evaluation criteria into computable metrics, and expand the framework to cover the evaluation of multimodal medical AI (e.g., medical image analysis models).

MELMA-Q represents an important advancement in the field of medical AI evaluation. It reminds us that when applying large language models in high-risk scenarios like healthcare, we need to strictly examine them from multiple dimensions such as accuracy, safety, and practicality. Clinicians' professional judgment is indispensable, and MELMA-Q provides a reference framework for the development of reliable medical AI.