Singapore medical school releases healthcare GenAI ethics checklist

Singapore medical school, Duke-NUS, has partnered with international and local researchers, as well as Singapore public health institutions, to develop a standardised ethical framework for regulators and developers of healthcare generative AI (GenAI) applications.

Participating public health institutions include the Singapore National Eye Centre (SNEC), Singapore General Hospital (SGH), and the SingHealth artificial intelligence (AI) office.

Named the Transparent Reporting of Ethics for Generative AI (TREGAI), the ethics checklist outlines nine widely accepted principles and definitions tailored to the healthcare context.

These nine principles are accountability, autonomy, equity, integrity, privacy, security, transparency, and trust.

For example, the checklist has tailored the principle of autonomy within the healthcare context to refer to the need to preserve and foster patients’ dignity, right of self-determination, and their capacity to make informed decisions.

The checklist seeks to provide a “comprehensive assessment and transparent documentation of ethical discussions in GenAI development,” according to Duke-NUS' official statement, and is maintained live on GitHub to incorporate timely developments in the GenAI space.

To subscribe to the GovInsider bulletin click here.

Broad-based regulations and guidelines are insufficient

According to the researchers, while regulations and guidelines are top of mind for practitioners seeking solutions for ethical issues for healthcare GenAI implementations, they remain insufficient.

When it comes to broad-based ethical principles, the challenge is when practitioners try to interpret them against a specific requirement of a given context.

Such guidelines also offer little help when trade-offs need to be made between ethical principles.

For example, in the healthcare setting, clinicians may face an ethical dilemma between the principles of do-no-harm and autonomy in situations where patients do not have the physical or mental capacity to make their own decisions.

The other gaps found in the discourse around healthcare GenAI ethics included the lack of discussion into other GenAI methods beyond large language models (LLMs) such as ChatGPT.

One example of a GenAI method is multimodal GenAI, which can understand and generate outputs across multiple mediums, such as image-to-text models.

Such models offer extended capabilities for healthcare but are also more complex to implement. One type of multimodal GenAI is generative adversarial networks (GANs), which can be used to generate both X-ray images and radiology reports.

Checklist used in both research and user settings

While the developers expect the checklist to be used primarily by researchers, funders and regulators, they noted the possibility of adapting the checklist for other settings.

One instance is to help healthcare practitioners disclose ethics-related considerations when using GenAI-powered products with their patients. This can “help users establish reasonable trust in [GenAI] capabilities.”

Another instance is adjusting the checklist to assess the benefits, limitations, and risks of various GenAI-generated content, including social media posts and teaching materials.

“As GenAI adoption becomes more widespread, the checklist may also increase the general public’s awareness of ethical issues that have arisen because of it.”

The ethical principles have been developed by the researchers scoping the available research around GenAI, AI ethics and healthcare.

The study is now published in The Lancet Digital Health where researchers discussed findings and explain the value of the checklist for those working in the healthcare sector.

To subscribe to the GovInsider bulletin click here.