ESPE Yearbook of Paediatric Endocrinology

In Brief: This ‘News in focus’ article in the journal Nature describes a non peer-reviewed preprint (1) by researchers at Google Research and Google Deepmind that claims remarkable abilities for medical artificial intelligence (AI), beyond ‘simple’ pattern recognition tasks.

The preprint describes Google’s ‘Articulate Medical Intelligence Explorer’ (AMIE), an experimental conversational AI system, or ‘chatbot’. It was tested on 20 trained actors who simulated 149 clinical scenarios. They were blinded to interact with AMIE or one of 20 board-certified physicians. AMIE matched or surpassed physicians’ diagnostic accuracy in all 6 medical specialties considered (cardiology, gastroenterology, neurology, obstetrics, respiratory and internal medicine). More notably, AMIE outperformed physicians in 24 of 26 criteria for conversation, including politeness, explanation, and in expressing honesty, care and commitment.

To enable efficient self-learning, AMIE was taught to play 3 different roles: a patient with a specific condition; an empathetic clinician; and a critic who evaluates the patient-doctor interaction and provides feedback on how to improve that interaction.

Many of us are worried about the prospect of medical AI. No doubt issues such as confidentiality, reliability, access ability and responsibility need to be addressed through debates within and beyond the medical profession. However, there are many amazing advantages – and many may be beyond our current imagination as this paper illustrates. To facilitate this debate and sharing of information, the New England Journal of Medicine has established a new (sponsored) interdisciplinary journal called NEJM AI. One of its first articles, by Hoifung Poon, Microsoft Research (2), describes the potential multimodal applications of medical AI.

References: 1. Tu T, et al. Towards Conversational Diagnostic AI. 2401.05654 (arxiv.org) 2. Poon H. Multimodal Generative AI for Precision Health. NEJM AI Sponsored. December 11, 2023. https://ai.nejm.org/doi/full/10.1056/AI-S2300233