Generative AI a cornerstone of NUHS’ healthcare IT strategy

Dr Ngiam Kee Yuan is a pioneering surgeon who has enhanced surgical outcomes with robotic thyroid surgery, one-stop thyroid clinics, and other initiatives. But within Singapore’s healthcare sector, he played the role of Group Chief Technology Officer at the National University Health System (NUHS), one of three regional healthcare clusters, until recently and is now Head of Academic Informatics Office.

The healthcare leader has been hard at work building RUSSELL-GPT, a platform that combines different large language models (LLMs) via Amazon Web Services (AWS) Bedrock to help clinical staff.

“For many things that we do, language models are quite applicable because the nature of healthcare is information in unstructured formats, such as text, images, and videos,” he says.

Today, more than a thousand NUHS clinicians use RUSSELL-GPT to draft referrals and memos, summarise patient data, and perform other tasks. This has reduced the time required for administrative tasks by 40 per cent, he shares.

The next step for GenAI: doing away with prompting

Yet his long-term plans for RUSSELL-GPT are even more ambitious. In conversation with GovInsider, he shares that his team is building event-driven Generative AI (GenAI) models that can automatically execute tasks.

This can be done by streaming data straight into the LLMs. When a certain event occurs, this would trigger the LLM to run automatically. For example, when a patient can go home, the model could automatically draft a discharge letter for a doctor to review, without any prompting.

Such event-driven language models can help save time and reduce administrative burden, he explains.

“This is unprecedented and it's certainly not something that you can do without some form of streaming capability,” Dr Ngiam says. In 2022, NUHS developed ENDEAVOUR AI, which integrates live data to produce insights, such as predicting wait times and optimising bed capacity.

The challenge now is training a single model to do many such tasks, he says. 

“How do we input all these tasks into a single machine, so that when the note comes in, it can output five things?” he explains.

Reducing administrative load on healthcare staff

While the team is in the process of building event-driven tools, RUSSELL-GPT’s current capabilities are already reducing the administrative load for doctors, he shares.

It currently helps doctors with language tasks, such as summarisation, writing referral letters, and classifying diagnoses. Depending on the task, RUSSELL-GPT can direct questions to the correct model.

Initially, doctors were unsure of how to use the tool and needed some training to become comfortable. Now, they are actively experimenting with the model to help themselves, he shares.

“Therein lies the importance of making sure that this is secure and available to our users. If you make something hard for them to use, nobody will try it,” he shares.

The team is currently collecting data to examine the impact of the platform in more detail, but the goal is to reduce the administrative work to “nothing”, he says.

NUHS is also currently developing a model that can draw on guidelines and evidence-based research to break down complex information, provide suggestions and enhance decision-making for clinical staff who are reviewing tumours as part of cancer care.

Unlike previous efforts to build similar models led by tech companies, their project aims to solve this problem from a “clinically driven perspective”, he shares.

Keeping data secure

One of the key features of RUSSELL-GPT is that patient data is not anonymised, as that would severely limit its usefulness.

 “Today we have thousands of transactions on Bedrock. If for every transaction, doctors had to scan through, remove personal information, and then run it, it wouldn’t be used.”

This is why the team worked closely with AWS to ensure that Amazon Bedrock models would be available within Singapore and that all data would securely reside within Singapore, he shares.

RUSSELL-GPT also adheres to Singapore’s prevailing data privacy laws and relies on the platform-level security offered by the government’s Healthcare Commercial Cloud, which comes with standard security tools such as identity access management and surveillance systems.

Finally, to reduce the risks associated with AI hallucination, the cluster ensures that users need to watch instructional videos and answer a quiz.

“The video and the quiz emphasise again and again that there are hallucinations… it will never be intended for clinical decision-making, it is simply supposed to support you,” he says.

They also include other features to reduce the risk of inaccuracies, such as clearly labeling where information is retrieved from.

Not a silver bullet

“Me personally, I want to press a button and say, you know, reply to all these emails for me. If I never have to reply, that will save two hours a day for me,” he says.

Although Dr Ngiam is optimistic about the role of LLMs within healthcare, he cautions that they do not solve every problem.

Language models are less capable at clustering information and providing predictive insights than classical AI models, he explains.

“They are incredibly useful and versatile tools, but when it comes to specific tasks, classical AI will still be the mainstay,” he shares.

AWS is constantly innovating in partnership with customers. During the last 18 months, AWS has launched more than twice as many machine learning (ML) and generative AI features into general availability than the other major cloud providers combined. Today, AWS is working with governments to kickstart their generative AI journey with a new $50 million impact initiative for worldwide public sector organisations.