Rationale and Cost of Developing AI Models for Healthcare Diagnostics, and Safeguards for Patient Data

Mr Low Wu Yang Andre asked the Coordinating Minister for Social Policies and Minister for Health (a) before the deployment of Singapore Medical Foundation AI Model (SIMFONI), whether the Ministry will subject its models to privacy testing by assessors independent of the developers to determine if they can memorise or reproduce identifiable patient information; and (b) whether test standards, findings and response protocol for any disclosure will be published.

The Singapore Medical Foundation AI Model (SIMFONI) develops artificial intelligence (AI) tools to provide clinical decision support for managing chronic conditions, prioritising conditions with high disease burden in Singapore and feasibility of development. We are starting with a couple of specialty areas and settings in the public healthcare sector. If successful, we will expand its coverage to more specialities and also possibly, to the private practitioners.

SIMFONI's models are built on existing models trained using international data. They are adapted using local patient data, they are de-identified, in a secure national platform with well-established processes, expertise and safeguards. This platform is called TRUST, which was established in 2022 precisely for research, development and innovation efforts.

To further ensure the integrity of this process, an independent safety and evaluation unit will carry out rigorous testing of SIMFONI's models before actual deployment.

Thank you, Mr Speaker. I thank the Minister for his response. While clinicians will ultimately retain responsibility for medical decisions, international experience has shown that doctors may gradually defer to AI decisions or recommendations – something known as automation bias. Will the Ministry announce specific safeguards to detect and mitigate this risk and how we ensure that AI compliments rather than subtly replaces independent clinical judgement?

We do not really announce what we are going to do to safeguard against this. This is very much in the way we approach AI, how we deploy and how we implement them.

I have given a few speeches on this matter. There is a lot of hype about AI, including in healthcare. And sometimes, you really wonder why are these proposals like that. Because those of us who work in healthcare, you know it is not simple, especially a public essential service.

We have to take a very careful use-case approach. That means, what problems are we facing, can AI solve those problems? Find the right tools. Never treat AI as a hammer going around looking for nails or a solution looking for a problem. We have enough problems. Let us see what AI can do to solve some of these problems.

And when you can find the right AI tool, make sure it is trained in the local context, taking into account security, de-identified data, patient data that is in Singapore.

When the model is ready, when you deploy it, you have to deploy in accordance with clinical protocols. There is a certain way we do things. There are existing legacy IT systems. AI is not something you plug into the socket and suddenly, it solves your problems. It does not work that way. And then, how do you interface with existing legacy IT systems? How the processes are thought through, with the clinicians still making the judgement?

It is early days, but I think we are very conscious that there is a lot of hype, a lot of use or misuse of AI. I think all this hype at some point will subside to a more realistic level, where people are more judicious, thoughtful and even, sober about using AI. And we tell ourselves, "Why wait for that process? Why do we not start off being sober and judicious?" And that is the approach that the Ministry of Health is taking.

Thank you, Speaker. Minister, I have three supplementary questions. Firstly, it is about the scope of the data that is being envisaged to be used to train this AI model. Would it encompass all health data in Singapore of all patients, including data in the National Electronic Health Record (NEHR) database?

Secondly, it is about the nature of the consent that will be sought before patient's data will be used for training. Has the Ministry considered, or will the Ministry consider, asking patients for explicit consent before their data can be used for training purposes?

And third supplementary question: I understand the Minister has given assurances that the data will be anonymised and have personal identifiers removed before they are used for training. However, as we can see, I have another Parliamentary Question today for a similar data leak incident with the Singapore Land Authority, where a lot of these processes are fallible. And sometimes, anonymisation of data does not actually pan out and personal identifiable data can leak into publicly available databases. What assurances can the Minister give to the public that their very sensitive health data will remain safe?

I think we all know AI is a breakthrough technology, but we must take a thoughtful and judicious approach. So, we have to decide what we want to do with it. If we start off thinking that we distrust this system entirely and therefore, let us scope down the data to be used to train models – even though we know the data is essential to train models that can improve healthcare, save lives, operate everything better and maybe help us address some of our very stark challenges of an ageing population with rising patient load – and if we accept that, but yet at the same time distrust the system and say, "Let's scope down the data, let's not believe that they can really be anonymised and de-dentified because there can be leaks and traceability", and "Let's seek everyone's consent before we can use", then it is a non-starter. It becomes a non-starter.

I think just take a balanced approach: recognise that this is a breakthrough technology that can do a lot of good, but to do so, we will need to use the data that we have and take the necessary precautions. Make sure that it is anonymised and de-identified. And we started this process some years ago. So, TRUST data has been used in various contexts already – so far, incident-free. I think we make sure it is anonymised and de-identified by international standards.

I think let us proceed on that basis. We must have safeguards, I agree. But we must strike a balance between leveraging that technology, but putting in place the safeguards. And we can proceed on that basis.

And at some point, Singaporeans will see that healthcare services are improving, yet quite incident-free. That is when they have confidence and assurance. I think let the action and the results speak for themselves.