New consensus framework aims to improve safety and transparency in medical LLM deployment

Highlights

1. Inconsistent evaluation standards have hindered the safe deployment of LLMs in clinical practice. This consensus establishes a systematic retrospective framework for pre-deployment evaluation of LLM applications in healthcare.
2. The framework is built on four core principles: scientific rigor, objectivity, comprehensiveness, and ethical compliance, and translates them into a standardized workflow for pre-deployment risk assessment.
3. The consensus defines hybrid evaluation metrics that combine quantitative indicators with qualitative expert scoring, including MOS-based assessment of accuracy, completeness, safety, practicality, and professionalism, aligned with six core evaluation areas in medical LLM use.
4. A multidisciplinary evaluation model is recommended, spanning medical experts, computer scientists, ethics experts, statisticians, and legal experts, with admission testing and quality assurance that include interrater reliability thresholds above 0.85.
5. Dataset construction emphasizes clinical authenticity, representativeness, fairness, and de-identification, with recommendations for modular expansion, version control, and compliance safeguards aligned with privacy and data protection requirements.
6. Dynamic feedback channels, dispute arbitration, and scheduled updates are recommended to support continuous improvement, safety monitoring, and adaptation to regulatory, technical, and clinical change.
7. Standardized evaluation reports are proposed to improve transparency on model details, data sources, methods, results, expert participation, and validity management.
8. The consensus also identifies major barriers to safe deployment, including bias, privacy risks, inconsistent evaluation practices, and limited data sharing, and recommends privacy-preserving multicenter evaluation approaches to support safer clinical integration of medical LLMs.

***

Reference
DOI: 10.1016/j.imed.2025.09.001