Clinical AI safety and evaluation

Research question

What evidence tells us whether a clinical AI system is useful and reliable in the context where people will use it?

Methods and findings

In a 2026 Nature Medicine correspondence, Gorenshtein and colleagues discuss why apparently conflicting evaluations can reflect differences in prompts, tasks, populations, and outcome definitions. The article argues for interpreting performance in its evaluation context; it is a methodological argument rather than a prospective clinical trial.

A separate PLOS Digital Health study evaluated ten language models using 305 neurology board examination questions. It compared base models, retrieval augmentation, and a multi-agent reasoning framework, with additional evaluation on 155 neurological MedQA cases. The findings support further investigation of structured reasoning, while examination accuracy remains different from demonstrated benefit in patient care.

Authorship

Co-author of both publications. This page describes the published collaborative work; individual contribution statements are available in the original articles.

Publications and sources

• How to meaningfully evaluate AI in clinical medicine

  Mahmud Omar, Reem Agbareia, Alon Gorenshtein, Ashwin Ramaswamy, Ankit Sakhuja, Yiftach Barash, Daniel S. W. Ting, Eyal Klang, Girish N. Nadkarni. Nature Medicine (2026).

  DOI: 10.1038/s41591-026-04350-5
• A multi-agent approach to neurological clinical reasoning

  Moran Sorka, Alon Gorenshtein, Dvir Aran, Shahar Shelly. PLOS Digital Health (2025).

  DOI: 10.1371/journal.pdig.0001106