AI Safety Evaluation Guide for Healthcare

The Healthcare Sub-Working Group, part of the Business Demonstration Working Group of the AI Safety Institute (AISI), developed the Healthcare AI Safety Evaluation Perspectives Guide to enable the safe social implementation of generative AI technologies in healthcare.

While generative AI is being used to improve physicians' work efficiency and support communication with patients, challenges include health damage caused by hallucinations, advanced privacy protection, and ensuring security.

Based on AISI's AI Safety Evaluation Perspectives Guide, this guide is a practitioner-oriented resource that presents specific evaluation methods reflecting the sensitivities and risks unique to medicine and healthcare.

The guide aims to ensure safety from the development and design stages of generative AI and help healthcare providers create an environment where business value and safety and security can be achieved together.

Because failing to evaluate misinformation and dangerous information related to medicine and health could directly harm patients' lives or health, systematizing evaluation perspectives is relevant to risk management when using AI.

As the Basic Plan for Artificial Intelligence calls for promoting the development, demonstration, introduction, and social implementation of AI in medicine and healthcare, and Trustworthy AI is being discussed internationally as an important issue, this guide organizes AI safety evaluation in healthcare from a practical perspective.

The Healthcare SWG aims to update the guide as appropriate in response to technological developments, social conditions, and international regulatory trends, thereby contributing to safe social implementation and sustainable business value creation.

The guide defines five phases aligned with the AI lifecycle: product design, model selection, product implementation, product validation, and product introduction and operation. It organizes the points that should be evaluated at each stage.

There are 10 evaluation perspectives, providing a multidimensional organization of the matters required when applying AISI's evaluation perspectives to healthcare. They cover risks ranging from universal risks to risks specific to healthcare, and also describe the risks anticipated if each perspective is not evaluated.

The guide is structured for practical use, including by companies with few specialists. A practical guide with concrete examples such as prompts and agent skills is also scheduled for publication to support evaluations based on the evaluation perspectives.

The participating organizations are Ubie Corporation, Awarefy Inc., Ajinomoto Co., Inc., SB Intuitions Corp., CMIC Holdings Co., Ltd., SherLOCK Inc., Japan Digital Health Alliance (JaDHA), MICIN, Inc., The Tokyo Foundation, and Mitsubishi Research Institute, Inc. Ubie Corporation serves as the SWG leader, and Mitsubishi Research Institute serves as the SWG secretariat.