Ethical Considerations for the Translational Application and Review of Biomedical Research Involving AI - A Briefing Document

Aug 23, 2026

A briefing document outlining ethical considerations for translating biomedical AI into healthcare, with guidance focused on three intertwined issues across the R&D pipeline: bias, appropriate human involvement, and risk prediction. The material links each theme to Singapore research practice through local case-study examples and emphasizes how to evaluate implementation choices using concepts of utility, equity, and reasonable risk. Recommendations are intended to support researchers, institutions, and IRBs when navigating translational clinical research and real-world evaluation.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages40

Reading Time1 h 41 min

AccessFree to download

Introduction

What’s in this document?

Who is it for?

What should you know?

Methodology

Bias—Is it ever justifiable to implement biased AI?

What is bias in AI?

Why not just remove the bias?

How should utility and equity be considered in evaluating the implementation of biased AI?

What principles apply in deciding when implementing a biased AI is justified?

Strategic Measures to Mitigate Bias in Practice – What can you do?

Human Involvement – How and to what extent?

What does human involvement in AI mean?

Leaving Humans Out of the Loop – What Are the Main Concerns?

What principles apply in deciding the role of human involvement?

Strategic Measures in Practice to Ensure Appropriate Human Involvement – What can you do?

The Risks of Risk Prediction – How much risk is reasonable in an already risky world?

What is risk prediction in AI?

Assessing Risk – Are AI risks exceptional?

Moving Forward with Implementation – How should risk mitigation strategies be evaluated?

What principles apply in deciding a risk threshold for implementation and risk mitigation strategies?

Strategic Measures in Practice to Evaluate Reasonable Risk Involvement – What can you do?

Conclusion

Annex

Ethical Considerations for the Translational Application and Review of Biomedical Research Involving AI—A Briefing Document Acknowledgements This document would not have been possible without strong collaborative efforts among the involved parties. Alexa Nord-Bronzyk is the main author of this document, working closely with Michael Dunn asthe co-author. There was strong support from the NUS Centre for Biomedical Ethics team, especially Sumytra Menon, Jerry Menikoff, and Julian Savulescu. The foundation for the arguments presented in this document draws upon ongoing work from Science, Health and Policy-relevant Ethics in Singapore (SHAPES), an NUS Centre for Biomedical Ethics initiative supported by the Singapore Ministry of Health’s National Medical Research Council. The authors involved include Kathryn Muyskens, Harisan Nasir, Angela Ballantyne, Murali and Julian Savulescu, Yonghui Ma, Jerry Menikoff, and James Hallinan.1 The SHAPES AI working group was also instrumental in the completion of this document. Reviewers include Tsung-Ling, Ma Yonghui马永慧, Angela Ballantyne, and Pavitra Krishnaswamy. An extended thankyou to the wider working group for their ongoing support. The SHAPES team is deeply grateful to all involved in the development of this document. Thankyou all for your time, effort, and invaluable insights. 1 See methodology for further details on the involvement of these authors. Contents: 1. Introduction 1.1. What’s in this document? 1.2. Who is it for? 1.3. What should you know? 1.4. Methodology 2. Bias—Is it ever justifiable to implement biased AI? 2.1. What is bias in AI? 2.2. Case Study: The Permissibility of Biased AI in a Biased World: An Ethical Analysis of AI for Screening and Referrals for Diabetic Retinopathy in Singapore 2.3. Why not just remove the bias? 2.4. How should utility and equity be considered in evaluating the implementation of biased AI? 2.5. What principles apply in deciding when implementing a biased AI is justified? 2.6. Strategic Measures to Mitigate Bias in Practice – What can you do? 3. Human Involvement – How and to what extent? 3.1. What does human involvement in AI mean? 3.2. Case study: Spine AI: Medical Imaging For Lumbar Spinal Stenosis 3.3. Leaving Humans Out of the Loop – What Are the Main Concerns? 3.4. What are the criteria for kicking humans out of the loop?– Benefits vs. Costs 3.5. What principles apply in deciding the role of human involvement? 3.6. Strategic Measures in Practice to Ensure Appropriate Human Involvement – What can you do? 4. The Risks of Risk Prediction – How much risk is reasonable in an already risky world? 4.1. What is risk prediction in AI? 4.2. Case Study Score for Emergency Risk Prediction (SERP) --Machine Learning Triage Tool For Estimating Mortality After Emergency Admissions 4.3. SERP Risk Prediction – What is the Main Concern? 4.4. Assessing Risk – Are AI risks exceptional? 4.5. To Implement or Not-What Are the Tradeoffs? 4.6. Moving Forward with Implementation – How should risk mitigation strategies be evaluated? 4.7. What principles apply in deciding a risk threshold for implementation as well as risk mitigation strategies? 4.8. Strategic Measures in Practice to Evaluate Reasonable Risk Involvement – What can you do? 5. Conclusion 6. Annex 1. Introduction 1.1 What’s in this document? This document discusses key ethical considerations surrounding the pipeline of research and development activities focused on translating AI into healthcare. The focus is narrowed to three pertinent ethical issues: bias, human involvement, and risk prediction. Each theme is discussed in relation to a Singapore research case study and offers recommendations grounded in an understanding of local research practice. The guidance presented intends to offer focused insights into resolving local real-world challenges.2 1.2 Who is it for? This document is for researchers, research institution

FAQs

What three ethical issues does the document focus on for translational AI in healthcare?

The document focuses on bias, the extent and criteria for human involvement, and risk prediction. Each issue is discussed in relation to translational research and implementation decisions.

Who is the guidance intended for?

It is intended for researchers, research institutions, and IRBs in Singapore, including those translating new technology into practice and those evaluating it once implemented, such as clinicians.

How does the document define and frame “artificial intelligence” for its purposes?

It endorses the Nuffield Foundation definition of AI as technology that performs tasks considered intelligent, while noting that beliefs about what counts as intelligent may change over time. It also references AI tools rather than AI in general to keep the scope appropriate.

Aug 23, 2026

Upload by นรินทร์