CSIRO Conversations: From promise to practice, the future of AI adoption in healthcare

Opening video   1:14
They say curiosity killed the cat.
AEHRC, Australian eHealth Research Centre.
No, we think curiosity.
Ask the right questions.
and looks at things differently to help solve the world's biggest health problems.
It's what brought us here.
Besides, we love science, and we're not kidding around.
A EHRC Digital Health.
Go on to it.

Hello and welcome. Good afternoon and welcome to CSIRO Conversations. I'm David Hansen and it's a pleasure to host today's discussion, From Promise to Practice, the Future of AI Adoption in Healthcare. We're streaming live from Meanjin, Brisbane and in particular the precinct in Fortitude Valley. Thank you for joining us online and here in person.
Before we begin, I'd like to acknowledge the traditional owners of the lands that we're meeting on here today. We're joining from Meanjin, Brisbane, and we pay our respects to the Turrbal and Yuggera peoples, their elders past and present, and all Aboriginal and Torres Strait Islander people joining us today.

Today's discussion builds on CSIRO AI trends in healthcare 2026 report. You'll find a link to this in the chat. And of course, this is what it looks like. Also, while I'm mentioning the chat, the chat is there. It would be great to have people putting in questions. We've got a number of our scientists and engineers here in the audience today and in the office who are also in the chat and they'd be more than happy to participate and answer some questions. So our AI trends for healthcare report shows that AI has moved from the sidelines to the centre of healthcare. Organisations now are using AI in many, many different ways across healthcare.
Supporting clinical decision-making, improving efficiency, and we hope enhancing patient care.
The report also highlights the responsible innovation, evidence and collaboration that will be critical to unlocking AI's future, AI's full potential. Before we get into the discussion, just a few housekeeping items. Today's session is being recorded and a recording will be made available after the event.
Online attendees can use captions and submit questions through the S&A function in the Teams Town Hall. And as I mentioned, we'll have some of our scientists and engineers on the chat as well. We'll be drawing on both pre-submitted and live questions throughout the session. For those in the room, we have microphones.
But I'm also joined on the stage by Naomi, who will be helping facilitate questions and guide us through today's case studies and discussion.
The question facing organisations today is no longer whether AI can create value, but how to deploy it safely, responsibly and at scale to take advantage of what AI and digital can offer. Healthcare offers a powerful example. It's A data intensive, highly regulated and built on trust.
Making it an ideal environment for understanding the opportunities and challenges of AI adoption.
Today, we're going to be exploring what's working in AI implementation, how AI is being integrated into real world workflows, what organisations can learn from healthcare, healthcare's experience with trust, governance and evidence, and some of the emerging AI capabilities and what they mean for the future of health.
Health.
I've got a great panel up here today, so I'm very pleased to welcome as well as Naomi, Bevan, Denis, Liesel and Kai Lin. And we're going to meet them and talk a little bit about AI through some questions.
Firstly, generative AI and large language models have changed AI over the last few years. We now have tools like ChatGPT that have made AI available to everyone. They are a bit of a mystery though. So for my first question, for our first question today, I'll be asking Bevan. Bevan is a research scientist and leads our health AI and data team. Bevan, can you explain a little bit about foundational models and how they all journey to AI.

Bevan:

Sure, thanks David. Yeah, foundational models are really the technology underpinning these huge changes that we're seeing in AI at the moment, and they are huge general purpose models. So before foundational models, we used to have to train a small model to do a specific task, say interpret this MRI image or
summarize this e-mail, and that's all they could do. Whereas foundational models have a general problem-solving ability that means they can tackle tasks that they've never seen before. And the way they do that is come through, I guess, three key advances. One is huge amounts of pre-training, so that's
where you expose the model to, say, pretty much all of the internet, and it begins to learn a really good grasp of natural language, of what images look like, and has this core ability.
And then we teach it to perform certain tasks. So that's instruction fine tuning, which is often done with human feedback. So the model is presented with tasks, it's got to try and solve that task, and it learns from human feedback where it's doing well. And then I guess underpinning all of this is a really big increase in our compute.
our hardware, our ability to use modern AI specific chips to train much, much, much larger models than we were able to before. So these are the three kind of pillars driving the rise of these foundational models.
You mentioned generative AI as well. That's really just an umbrella term that captures all these different families of models who are designed to generate novel content, to generate new language, new images. That's where we are. Fabulous. Thanks, Bevan. Now, you can read more about these technologies, generative AI, foundational models, etc. at the beginning in the first part of our AI report. And one of the other topics we mentioned in the report in the first part of the report is a genetic AI. Denis is our cloud computing and genomics expert and leads our transformational bioinformatics group. Denis, can you tell us about AI agents and how they work?
Identic AI refers to a system that can do more than answer a question or generate data. An AI agent can pursue a goal by planning, taking actions, using software tools, observing the results, and adapting of what they do next. So the key here is that they have some sort of agency themselves, which enables them to take action. So they don't rely on the human to specify every little step that they need to do. So for example, we have created an agent around our OmniFold algorithm, which predicts how the payload is folded inside the molecular delivery vehicle for vaccine development. This allows the vaccines to be produced more efficiently and be tailored to the viral evolution that is going around them. And the agent can do that by optimizing the payload.
Great, so biogentic, do you mean autonomous?

Great point, David. No, agentic does not mean autonomous. Particularly in healthcare and genomics and Biosecurity, the human in the loop is absolutely vital.

Excellent, thanks. One of the other areas that we asked, one of the areas that we're seeing AI used a lot in comes from when you visit the doctor and they're using an AI scribe. Now, they still capture data. It's still important that we capture data in electronic health records in structured and coded form.
Kylynn leads a lot of our interoperability work, including co-leading the National Sparked Accelerator, which involves working with clinicians to discuss what data should be captured in those electronic health records. Kylynn, in all our interactions with clinicians, what are they talking about when it comes to AI? We're definitely seeing an increase in the use of AI scribes and AI tools within general practice.
And they're using it really to just take notes of your conversations with your doctor. So they're using it in their day-to-day appointments, also to support their documentation of clinical notes to help fill in forms and develop care plans and other artifacts that are developed during your visit to the doctor. We are also seeing other care providers and other clinicians starting to adopt or at least explore the benefits and the risks of using AI in their care practice. We at Sparked are developing the clinical data and technical exchange standards that support the safe and effective use of AI technology, including scribes, and we are working with the clinical and technical community to help us do this.
Great. To our final panel member today, Liesel is part of a health services group working at how digital and AI changes the way we deliver health services. Liesel, when it comes to aged care and community care, where are these sort of technologies starting to be really used?

Yeah, we really see the application here, David. And I think where the sort of three main areas where we really see it, which is workforce issues, we see it in the safety issues and providing safety for people based in the community, and also general wellbeing. So those might be examples like the AI scribes, like we've just heard about, people taking notes, really using them for their workflows across care providers, but also looking at things like fall sensors, where we're looking at safety issues, and then really looking into the wellbeing space where it might be things like social robotics.
Excellent, thank you. Today we're going to run through a number of case studies. I think they're all in our report, so I encourage you again to download the report. There's a link in the chat. And the first case study we're going to look at explores some of those emerging AI technologies that we heard about. This includes large language models and Aaron's going to tell us a little bit and we'll discuss it after. Multimodal AI and how this is used. Over to Aaron.
Medical imaging in and of itself naturally helps people on a personal level. It locates and provides an extent to an abnormality of a patient. We develop machine learning and artificial intelligence tools for diagnosis and prognosis, as well as treatment planning for a clinical setting. We work with clinicians, industry and patients to understand their needs. A single tool could be hosted and serve multiple healthcare systems simultaneously. And that's very cost effective and affordable. We work with medical colleges as well as other research institutes in order to collaborate and make more impactful research. Another way we're trying to look into the future is by generating medical images from radiologists' reports. And this can help with a few things like disease progression for clinicians. It will enable them to be more efficient in their job.
Particularly with the deep learning and artificial intelligence techniques that we're working on, it's very mind-boggling to see them work and to see them think like clinicians.
Great, thanks, Aaron. And we actually have, oh, did we finish there or did the video pause?
Just have a small technical issue here. We did have a question from the audience about AI scribes online, which was, are doctors obliged to let patients know when they're using AI scribes? Certainly the RECGP and others encourage it as best practice, and certainly you should be asking if you're going in.
Oh, great. Aaron, thanks. I'd like to narrow down to one application that we've developed and trialing, CXR Mate. Firstly, what does it do and how does it work?

14:00
Thanks, David. So if you've ever had an x-ray, you'll know that often several images are taken when you go into a radiology imaging room. And A radiologist then needs to analyze those images and prepare a report. And as the number of patients grows, so does radiologists' workload and their risk of burnout. So to help solve this problem, we've developed CXRMate, a multi-model large language model for chest x-ray reporting designed to support radiologists in their clinical workflow.
So multimodal means it looks at more than just the x-ray. It also looks at the patient's medical details and history, and when available, their previous x-rays and reports. And by looking at the full picture, CXRmate represents a new generation of tools for medical imaging.

Great, thanks. I know that we've already trailed a version of this with Metro South Brisbane Hospital and Health Service here in Brisbane. What did we find?

15:03
So we conducted a blinded pilot study with three consultant radiologists from the Princess Alexandra Hospital, comparing CXR Mates reports with reports written by radiologists. The results were really encouraging. CXR Mates reports were either preferred or rated equally to radiologist reports in 45% of assessments.
which was getting close to our 50% target. Its reports were consistently preferred for readability, although radiologists were still better at identifying subtle abnormalities. So the study showed us the potential of the technology and where further work is needed.

15:38
Excellent. And 45% sounds low and 50% as a target sounds low, but that is actually best practice in this case, right? That's what we're actually aiming for, to show that the reports are as good as, or can't be differentiated from the humanly written ones. Is that right?

15:55
Yes, if we reach 50%, that means the AI model performs equally to radiologists.

16:03
Great, thank you. Now, we use lots of data for this. How do we improve the model for use in Australia?

16:12
So that's the focus of our current work with the Princess Alexandra Hospital. The model used in our pilot was developed using publicly available data from the US. So we now need to understand how well it translates to an Australian health care setting. So at the moment, we're seeking ethics and governance approval to evaluate CXRM8 using de-identified
chest x-rays from the Princess Alexandra Hospital and using at least six radiologists to do the evaluation. This blinded evaluation would show us how the model performs with Australian patients and where it needs to improve before being tested in clinical practice.

16:47
Excellent. And a final question for you, Aaron, is that we did have a pre-submitted question about how we can ensure AI delivers massive efficiency and effectiveness gains in operating the healthcare system. Is that what we're aiming for here?

17:04
Yes. So the scale of the challenge is enormous. So millions of Australians undergo diagnostic imaging every year, and growing demand is placing increasing pressure on radiologists and reporting services. So tools like CXRMA can be shown to be safe, accurate, and genuinely useful.
They could help ease some of the reporting burden, giving radiologists more time to focus on complex cases and the clinical decisions that most need their expertise. For patients, more timely imaging results could mean quicker clinical decisions and that delivery of care is improved, which is our ultimate goal.

17:44
So Aaron, we've had an audience question come through and an anonymous person would like to know what else can it be used for on top of diagnosis?

17:53
So we're investigating the same technology on different tasks. One of those tasks we are focusing on now is discharge summary generation. So we want to use the same technology to look at all of the data collected during a patient's hospital admission.
to transform that into a discharge summary that a clinician can then review and make sure is correct.

18:25
Thanks, Aaron. Bevan, one of the things that Aaron was talking about was training the model on our own data. And so we have had another question from the audience around sovereign AI. How important is it for us to build our own AI tools on Australian data?
Yeah, so just for those not familiar with the term, sovereign AI is this principal idea of us building, hosting, managing AI ourselves rather than having that managed externally. So there might be like at a national level, an Australian sovereign AI approach or even down to an institutional organisation.
Why do we want sovereign AI? I guess at a national level, if we begin to actually think about AI as a critical piece of infrastructure, like our electricity grid or water, we want to make sure that we have control over that, that we're not exposed to wild fluctuations in supply.
So, you want to have it in-house?
The second is that we know that although AI is quite general purpose, when it's trained in one setting and then brought to another setting, it doesn't always work as well, right? So think AI trained in a US healthcare setting, which is quite different to an Australian healthcare setting, might not translate.
well, right? So we want AI that's adaptable, that's localized, that's fit to our needs. And I think the third is it's a great economic opportunity to really build our local AI industry. Australia's got world-class research institutions.
But we sometimes struggle to translate that into a really strong local tech industry. So it'd be great to kind of support that as well. Hear, hear. Great. So that's and that's a great example from Aaron of using really new technologies and the new AI technologies that, you know, we're all using across society today.
In terms of driving efficiency and effectiveness in the health system, which is a key theme of a lot of the questions, Liesel, where are you seeing that right across the health system?
Well, there are practical examples that I think we all use across all our industries. So scheduling, triaging, funneling information, how do we use algorithms to really make workforce efficiency better, very, very needed within the health space across community care. The other area which is really interesting and I think talks about that translation of science, which we've actually done quite bit of in our team is smarter homes technology. So where can data come from the community and go into the care provider? Smart homes is one of those areas that people are increasingly interested in. People wish to stay in their homes for longer as they're aging or living with chronic disabilities or other chronic health conditions. This is a really great way of people being able to collect their own data, send it to their care provider, have efficiencies that way. And then on that topic as well, another little personal one of mine is the concept of BYO data. So all of us, oh no, not BYO, other things, BYO data, everyone, almost everyone, I can see even just looking around the room here, everyone's off.
You know, most people are wearing a smart watch, they're carrying a phone. People are actually collecting all their own health data through those mechanisms, through, you know, gait, through menstrual cycle information, through body temperature, BP, all those things. Increasingly, we're seeing people take that data, head to their care provider and go, I would like this considered as part of my care, and that actually helps create some efficiencies as well in the in the clinic rooms.
Excellent, thanks. And yeah, when we can think about that BYO data and the fact that we can access a lot of our clinical data now via My Health Record, this brings lots of opportunities for AI across healthcare and community care. Thanks. One of the things which Bevan and Aaron were talking about was using existing data to train the models.
That makes it really important for our data to be accurate, complete and much easier if it's structured and coded. Here at the Australian Health Research Centre, we're leading the Sparked National Accelerator, as we mentioned earlier, to agree on the structures and language used to collect that clinical data. I think we've got a little video about it.
This is the story of your healthcare data. You are unique. So is your healthcare data. Ever wondered why you have to tell your health story to doctors and health practitioners over and over again? Today, giving you access to your data has been tricky, as health systems often speak different languages. Why can't everyone just be on the same page, right?
Enter Sparked, Australia's FHIR accelerator. Putting your health journey front and centre, the Spark community is standardising the way your data is recorded and shared through the Australian adoption of the International Data and Exchange Standard, FHIR. Because your health data is exactly that, yours.
Captured at the point of care, it follows you seamlessly, giving you and your healthcare providers a more complete picture of your health situation, with access to the right data, in the right format, at the right time. Now, what Bright Spark thought of that? Sparked better standards, better health, better outcomes.
Excellent, thanks, and I acknowledge the Department of Health and the Australian Digital Health Agency and HL7 as our partners in the Spark program. Kylynn, tell us about the Spark community that we've developed over the last three years. Thanks, David. So while the health care system in Australia collects heaps and heaps of data, sometimes health information gets lost as we move between hospitals, into allied health, into specialists. So sometimes we have trouble finding that data. Also, the computer systems within that system speak different languages, and SPARC is working to find a shared common clinical and technical language between them in the form of data standards. By having that shared common language, those data standards, health information will be able to be shared safely and accurately to empower consumers and care providers. Sparked is a collaboration between government, clinicians, care providers, software vendors, consumers, and anyone else who has an interest to understand those clinical and technical requirements for how health information can be recorded and shared to improve patient outcomes. We at CSIRO, with our partners, are coordinating our community of over 1500 members to ensure that our clinical data standards and our technical data standards are working together and are fit for purpose. These data standards form the foundations of national infrastructure and national policy, but they also provide a solid foundation to ground the use of AI in healthcare. So we know that AI is not only using good quality information to drive it, but also outputting a good quality information that is safe to use. So if AI is being used to create a summary of your visit to the GP, it's using good quality documentation of what happened to create that summary, and that the summary then is an accurate representation of what happened in that visit.
Excellent. Thanks, Kylynn. So, we've now got the Spark community, which is building the structured and coded data that we want in our electronic health records, and the AI scribes who are capturing narrative text. Let's have a look at a video from Dion talking about what we're doing in that space.
Hi, I'm Dion. I make tools that help AI better understand doctors. So the problem, I suppose, at least in the general practice world, is that the clinicians want to and still need to produce these text clinical notes. So the problem with it being locked up in text data is it's not available to things like clinical decision support, longitudinal analysis of a patient's record, all the way through to population health analysis. We're going to solve the problem by helping existing systems and the new generation systems like AI scribes be able to create structured and coded data more easily and quickly and help them provide an interface to clinicians where they can work through their clinical note and get structured and coded data without providing a big impost on the clinicians. The impact is that the information that's currently locked up in text data is made available to those sorts of AI and non-AI based analytics that help us improve healthcare for people and the whole population. Essentially, these systems are about providing help to the doctor, they're not replacing the doctor. But the more they know about you and the more that they can get quick access to that information about you, just like the doctor, the better help that they can provide.
Great. Thanks, Dion. And we did have a question in the text asking about using AI scribes in outpatients to get structured data. In all sorts of places, AI scribes are being used across healthcare. We want to make sure that they collect really good structured coded data for our healthcare records. Okay, we're going to move on to talking about another case study now, and that's one around trust, evidence, and governance.
Adopting AI in healthcare and again, there's a section in our report. I'm going to keep reminding you about our report. There's a section in our report talking about some of the issues in evidence and governance. And adopting AI in healthcare requires the health system to understand consumer and clinician perspectives and the regulatory frameworks that AI must must sit within. Our next case study focuses on ensuring AI is integrated safely and responsible. And our first vignette about that comes from Andrew Goodman, who's one of our Indigenous scientists. So over to Andrew.
My name is Dr. Andrew Goodman. I'm an Aboriginal man from Central West Queensland and I am a research scientist leading a program of work to explore the self-determination of AI's place within healthcare as appropriate to Aboriginal and Torres Strait Islander peoples. AI systems are only as good as the data they're trained on. And if that data is unrepresentative of a community or biased in any way, then we risk coming out with biased outcomes. We've advanced this program of work in collaboration with Indigenous leaders across our country to explore the opportunities and challenges artificial intelligence has within healthcare. Our report has provided us a great evidence base to see the opportunities that exist in AI and healthcare for Aboriginal and Torres Strait Islander peoples. And what excites me is that on an international stage, our peoples in Australia are leading the charge of what Culturally equity-based AI system could look like.
Thanks, Andrew. Bevan, you were involved in this work with Andrew. What were some of the important take home points in terms of Indigenous data sovereignty? Yeah, this was a great project to be involved in. And as Andrew pointed out, we do have to be conscious or careful about how these AI systems are trained.
particularly if they underrepresent certain populations. But at the beginning, we also mentioned how these foundational models are really good general purpose models. So in this case, we were actually able to take a foundational model and Andrew led an effort of essentially developing a key set of instructions or a culturally appropriate criteria or rubric that we were then able to operationalize to teach the model to act in this way such that it was ensuring Indigenous data sovereignty and
and able to critically assess information to ensure it was culturally appropriate. So I think it was a great project, and it was done with a small team that shows, I guess, somewhat the democratization of these foundational models that we were able to
make use of. Excellent. Thanks, Bevan. Our report also emphasizes that as healthcare becomes increasingly reliant on AI-enabled tools, the need for quality assurance, robust evidence and strong standards becomes even more important. We're now going to hear from the head of our QMS, a quality management system, sorry, Yan.
So, you have an idea for software as a medical device, and you want to take it to market. We get it. You're excited. But wait, certification is tricky, like this cake. It's layered, packed full of ingredients, and has gone through many processes before being served. Imagine then, we need to get this cake certified, just like your product. The auditors, aka baking judges, aren't only looking at the end product; they are looking at your recipe, the ingredients, the method, and asking if you can justify the baking time. No use baking a cake, no one can eat, right? By partnering with CSIRO's AEHRC, we can collaborate early on the research design and development process and ensure deliverables comply with the ISO 13485 standard. LQMS certification.
ensures everything we bake has been through the right test kitchen, making us the master chefs of digital health tech. We can know early whether or not something qualifies as a medical device and take the necessary steps to ensure it is fit for commercialization. This supports sustainable research and safe market ready devices. AEHRC's quality management system, making sure you get to bake your cake with CSIRO and eat the cake too.
Okay, and now, okay, and now I think we've got the real Yan joining us. Thanks, Yan. Great video. Frameworks for quality insurance and meeting regulation are top of mind in a lot of industries. We began the implementation journey back in 2022.
Tell us about some of the thinking behind doing this in terms of resourcing and the rationale behind it.

33:26
Thanks, David, and hi, everyone. I'm the real Yan. Okay, so the main driver was research translation. But I can't speak for every industry, but I definitely can share our experience in software as a medical device. Now, if we want to move AI research into clinical practice, we need to meet the TGA requirements. In another industry or another country, that may be a different regulator or assurance framework. So back in 2022, we decided to invest in a quality management system.
And a QMS is expensive. It takes people to build and run it, platforms to manage the evidence and documentation, processes that teams need to follow, and ongoing effort around the certification. So it was a significant investment for us but we still made that investment because we saw QMS as a key part of translating AI in research into the real world. So in our area, the QMS is really the bridge between an AI prototype in research and real world use. It helps us to be the evidence as we go. So when the AI technology is ready, the pathway to the real world is already there.

34:57
Excellent. So what were some of the biggest technical, organizational and governance barriers to enterprise AI adoption?

35:05
Well, from my experience, one of the biggest barriers was simply resources. Back then, when we first started, it was basically me and an external consultant trying to get the QMS in place. We had a target timeline for implementation and certification.
While at the same time, the AI projects were moving so quickly. So we were building the quality system while the technology itself was still evolving. And also cultural change. AI researchers are used to experimenting, iterating quickly and changing direction as they learn. And then you introduce a QMS with more documentation and review. So it naturally can feel like an extra layer of work. And governance is moving to AI regulation and assurance expectations are still evolving. So, you're not building for a completely fixed environment. The processes need to be strong enough for today, but flexible enough for tomorrow to adapt as the technology and regulation change.

36:23
How did we solve some of those problems?

36:26
Oh wow, it definitely wasn't overnight. It was really a case of trialing, learning what worked, adjusting as we went, and building the capability over time. On resources, we started small and gradually brought more people in as the need became clearer. Now, on the cultural side, we learned pretty quickly that you can't just hand researchers a new process and say, hey, this is how we do things from now. We had to make it practical, reduce unnecessary burden, and help people understand.
why the evidence and documentation actually matter. One of the biggest changes was bringing quality in early. So instead of finishing an AI prototype and then asking what do we need to prove, we started thinking about that from the very beginning.
And governance has been similar. AI and regulation are both moving so quickly. So LQMS can't be static. We need to keep reviewing it and adapting it as the technology and expectations change. And so the biggest lesson is that quality works best when it becomes part of how you build AI, not something you think about at the end.

37:52
Fabulous. And now that we have got certification of our QMS, we must have a few tools ready to go through approval. How are we going?

38:01
Oh, yes, we are very excited about that. We are focusing on one of our technology. It's called SCT Gen. and that will be the first technology that goes through the certification, and hopefully that people can use it very soon.

38:20
Right, thanks. Thanks, Yan. So, so far today, we've talked a little bit about some of the new AI technologies, large language models, foundational models, generative AI, multimodal AI. We've seen one of those tools in terms of CXRMate, which is a chest X-ray tool for which uses large language models and multimodal AI. We've talked about some of the regulatory things as well as communities and I think it's time to actually now have a look at a few more AI tools that we've been developing over many, many years. We've been using AI for well over 20 years, it's been great to be talking about it more. Since ChatGPT and the democratization of AI, we feel like we can talk about AI more without people's eyes glazing over. And it's great to be here at CSIRO Conversations doing that. So we're going to have a little bit of a look around and there's been a number of audience questions around different ways AI can be used, from better medical science to optimized diagnostics and improved efficiencies, of course. So the first, we're going to have a look at a few examples through a few videos, and then come back to the panel and discuss it. And the first one is from Anne Kylynn, who's one of our research scientists, talking about how she's using AI to aid with medical research. Over to Anne. Anne Kylynn, I'm a research scientist. So I like to imagine RNA vaccine being like small letters that we give to the body or to the cells, and then the cells learn from those letters.
to how to reproduce like a germ or something that will attack the body and then form this simulated germ, then it will know how to protect itself. So when those letters go away naturally, then the body actually remembers how to protect itself. And so mRNA therapeutic, it's an mRNA that you need to package in a delivery vector which is whether a lipid nanoparticle or a virus. But this is a big challenge because RNA is highly flexible. It constantly shape and reshape. It is negatively charged on the surface. And so it's constantly interacting with environments, even repulse itself. So it's very, very, very hard to predict, also is very hard to predict the carriers, the delivery vehicle, all of that, those packaging at the moment, the way it is predicted is trial and error. So they would formulate, test, and reformulate until they find like a good match. So it's consumed a lot of effort, energy, material, and time. To tackle that, we've developed 2 of our workflows using AI and using new technology approaches. So we developed computational or AI approaches to tackle those bottlenecks. And we already developed workflow. The first one is Omnifold. So Omnifold has been developed to predict how single-stranded DNA actually fold into compactness and it used machine learning to train on this, and it trains on your own simulated data to predict the folding and predict how it was going to package pretty much. The other computational approach that we have developed is can build. So can build also is a computational approach.
It predicts how a whole virus is in 3D based on just one protein. And this way we can modify it, engineer it, visualize it in 3D, and so on. Our way to tackle this packaging problem is to combine those AI approaches to find a good combination between a mRNA and the right carrier.
If we want to package more efficiently mRNA to the vector, we will be able to reduce the effort, manufacturing effort, the cost, the time to manufacture, and especially the time to patient.
Excellent, thanks. And Denis, your group's work on mRNA that we've just heard of, but of course right across the group, in using AI to enable medical research is really fabulous. What other examples do you want to tell us about? So from Anne's example, I think it's very clear that AI can help greatly and understand RNA and
genomic data in general. However, our genomes is one of the most sensitive and private information there is, so we don't want to expose it to AI directly. So we were working closely with the Genomics Alliance for the Global Alliance for Genomics and Health in creating new standards of how to uphold good data stewardship without hampering the innovation of AI.
So for example, we created Ask Beacon, which is a framework where conditions can query genomic data using natural language by keeping the beacon protocol between the AI and the sensitive genomic data.
Another example is Genomator, which creates a synthetic version of the original genome, which can then be exposed to the AI safely as the confidential information is removed by preserving the disease information so research can happen. Denis, just while you're in the hot seat, we've had a great question come through from Manuel.
They ask, it is clear that not all models used on one continent could be used on another. Genomic models differ.
How much data might be required to obtain a reliable analysis using AI tools?
Yes. So genomic data is vast and the complexity of it is not even understood yet. Therefore, it is unrealistic to assume that the models that we currently build can understand and can synthesize the information that is really encoded in the genome. However, the genome has in itself a certain language, and language models are exactly geared towards that. So therefore, the local interactions of how genes are regulated, how proteins are translated, and what they do, that can actually be understood, and this is what the models focus on at the moment. Thanks, Denis.
Thanks. And then because our DNA is global, that would then actually be tools that could be used around the world. Exactly. And not only between humans, but also other species as well. And I think that's a huge, huge benefit of really understanding how one health can actually fit together and how organisms in the real world interact with each other.
Fabulous and fascinating. Thanks. As I mentioned, we're looking at a few examples of different AI tools that we've developed over many years. One area where we think AI scribes will need to interact with are clinical guidelines. Here's Bevan again, but via video talking about how GPs might be able to access clinical guidelines using AI.

My name is Bevan. I lead a lot of the AI and health work at the Australian Health Research Centre, and I've been working on GuideStream, which is an AI agent to give GPs personalised individual access to clinical guidelines. Imagine as a GP, you really never know who's going to walk through the door or what you're trying to do.
treat or how you're trying to help people. And we know these GPs use clinical guidelines, which are these really big, dense, long 400-page PDF documents. So what we built is GuideStream, which is an AI agent that is able to help GPs access all this information that are locked in these documents. What it's going to mean is that GPs are able to really draw on the latest evidence be able to understand exactly what the current state-of-the-art is in medicine. It will allow them to understand individual settings for that patient. But it's really the adaptability that I think is very exciting.
Excellent. Thanks, Bevan. How will the GPs use AI scribes with AI clinical guidelines?
Yes, I guess this goes back to Denis's earlier point around agentic AI, which is really when we, instead of having these siloed individual systems going on, we're able to have a scene where things can coordinate with each other. So, for example, if I'm a GP and I'm...seeing a patient, say, Kylynn here, sorry, you're sick today, and the scribe is recording this, and maybe I prescribed some antibiotics that are not working, and I say, oh, well, we might have to think about what new antibiotics you might need. They are a scribe is recording all this, and it can then automatically go off to a system like GuideStream and say, okay, probably she's going to need some different set of antibiotics, and it can automatically go and do a full search and work out for this particular organism what's the best antibiotic, and have that ready so that when I finish the consultation or when I want to go and describe the antibiotic, I can very quickly look up and see what that guideline recommendation has said and, you know, really choose the best treatment option. Bevan, you're on a roll. So I've got another question for you. On your favourite topic, what is your opinion on people using consumer AI tools like ChatGPT to self-diagnose?
It's your favourite topic, isn't it? It is an interesting topic. We have run some studies like this, and there's obviously some real risks for people, but in other cases, we've shown that it's enabled people to understand certain conditions or understand what's going on with their healthcare at the moment.
So I guess the crux of it is to ensure that this is done in a way that is informative and is also done using tools that have actually been evaluated for these purposes. So we need to rigorously evaluate these tools to ensure they are correct. And once again, these foundational models are able to generalize. So, you know, if we carefully look at this particular problem, we can definitely help people with their healthcare journey. But of course, people should always seek advice from their doctors, correct? Yeah. Excellent, thanks. Thanks, Bevan. And, Kylynn, when we think about the SPARK program and the adoption of the standards that we've seen vendors now implement in their health records that are now being used across Australia, what are some of the
impediments to prevent some of the AI technology that we're discussing being implemented and how do we resolve those? Yep, definitely. As Dion mentioned in one of the videos, a lot of healthcare information is currently hidden away in lots and lots of free text. So it's hard, it's not structured, and it's not coded. So we need to kind of work on that. But also information is also locked away and require different specific technologies to access. So by using standards to provide instructions for how AI can access and use with appropriate consent and privacy guidelines, we can help to unlock the potential for those AI technologies.
Fabulous, thanks. We've got one last case study before we start moving into questions, and I know there's a lot coming in, so thank you for those questions, and you can keep putting them in the Q&A in the Teams Town Hall. The last case study is from Sankalp, who leads our Health Data Analytics team. He'll be talking about, we're going to watch it. as a video, talking about our digital twin for making the health system more efficient. Health intelligence team uses artificial intelligence, statistics, simulation and operations research to inform and improve health system productivity and safety and support operational and clinical decision making. We can, for  instance, predict with a high degree of certainty when the health system is going to be at capacity, when the emergency departments are going to be full.
We can use solutions like digital twins to simulate responses and work out how to best address capacity crisis. Machine learning is a field in AI that looks for patterns in data. We take the data that's in the healthcare system and use that to build personalized risk profiles of patients. Now, thanks to electronic medical records at hospitals, we can develop and deploy our predictive models in near real time to provide precision decision support at the point of care. Now, for doing this, we employ state-of-the-art machine learning models like random forests and gradient boosting trees. We also develop some novel methods that are better suited for use in healthcare.
Excellent, thanks. Liesel, this is an example where we're using AI to make decisions about how resources are used across the healthcare system. Again, looking at how AI might help in, for example, aged or community or disability services, are they the sort of uses we're seeing?

If I could pick anything in this space, I would say one of the most exciting uses is actually in the prediction space. So AI really is beginning to play a role where we can look at prediction of agitation, prediction of falls, and prediction of maybe wounds and things like infections beginning to come about.
These are massive, hard-hitting issues for aged care, disability, and have a huge impact right across the health system if they're not picked up early enough.
Right, and you mentioned disability there. We've done some work with the National Disability Insurance Agency looking at perceptions of AI use in disability. Can you tell us a bit about that work? Yeah, there's a lot of excitement across that space. We worked with the NDIS on the AI-enabled assistive technologies report, so really looking at what we need to understand about using assistive technologies in the people living with disability community. One of the things which I think really stood out to me when we did that report is when we spoke to people, they really want their quality of life improved, you know, their day-to-day tasks improved as well, and they're looking to technology to help with some of those things. But also really there were, you know, there's a bit of a gap there around the safety for them. People who are saying, you know, if I have to give up some of my safety and security across this space, I will if it means my quality of life is better. I think that's a real problem that we're facing and we have to keep going with regulation evidence, you know, research to really make sure it's safe.

Great, thank you. Thanks, Liesel. Okay, I hope everyone's enjoyed hearing about those case studies as well as everything else. And I'll remind you again, the report's available. There's a link in the chat. But now we've got to, I'm going to hand over to Naomi, who's been monitoring by her phone here, the Teams chat and the Teams Q&A. And we've had some amazing questions in. So over to you, Naomi.

We've had some really good questions come through. And I guess like overall, in terms of the pre-questions and in terms of the questions coming through the chat, two of the main themes that are coming through are the use of AI as well as the safe and responsible use of AI. So broadly, the questions kind of fall into one of those two categories. And I might start with a couple that have come through on the chat. Sammy asks, how do you ensure doctors don't end up over-relying on AI output? And I know this is something that does concern patients. David, would you like to tap in here? We know that from plenty of research that just having a human in the loop or asking them to review the output doesn't work to prevent eventual cognitive degradation. Yeah, look, I think there was a case on ABC News last week reporting where an iScribe had mentioned, someone was, I think was addicted to something and that had made their way into the health record. And, you know, it was a mistake picked up by the consumer. And so, and they've had to go through and really work hard at having that changed in their health record.
So it's a great example. Firstly, make sure if you, the GP, I think in my experience, GPs will tell you when they're using a scribe. I think you can ask then to look at what the scribe said and make sure that you're happy with the output and through my health record now we have, you know, we can access some of those reports and other things. So again, be using those sort of, that sort of infrastructure to make sure. Carlin, has there been any chat about that through the Spark community? Yeah, we're definitely thinking about the clinicians. We've had a lot of discussions around safe use of AI.
How can we make sure that we have the human in the loop, but also how do we trust the outputs of AI? And we also started to think about how can a consumer interact with some of those outputs of AI, whether they're patient summaries or, you know, records of the encounter or the visit that they've had. So it's definitely a topic that we're talking about at the moment, haven’t got answers yet, but we're definitely thinking about it.
And Syed asks, and I'm going to ask this question in light of the fact that we have probably some of the smartest people in the nation up here on this panel, but they are not legal experts. So in cases where a misdiagnosis occurs due to deficiencies in an AI model, who's held liable?

I think we're not lawyers..[unclear]… But again, I'd emphasize the need, for something you said earlier, Naomi, in terms of talking to your doctor, not just the AI. And I'd probably also say, maybe it hasn't changed. Maybe we've all been advised at various times to get a second opinion. So doctors aren't infallible either and they've made misdiagnoses in the past. So, so, and you know… For not a lawyer, I've got a lot to say here. And the third thing I'll say very quickly is health literacy. You know, I think it's really important that we continue to build our health literacy across our community so that people know an AI literacy and digital literacy, that people know that if it comes from an AI, you know, how do I check and how do I know to trust it? So I think there's obviously a legal question there about being held responsible, and that's why we have something like a quality management system, so that we've got a product that we know we can trust to meet the regulations. But then there's also the kind of common sense approach in terms of dealing with it.

Thanks, David. Now, this one might be for Bevan, might be for Aaron. Just see what happens. Do patients get asked whether their X-rays may be used to train the AI?
Lots of people very interested in the answer to this question. Bevan, can you take that one?

Yeah, sure. I mean, the data that we are working with here is, we're very fortunate to have some large collections of public data that obviously those patients have consented to make those x-rays available for us to train. And I think we're in a position where we're lucky in that there are those big data sets, but we then need to be quite careful that we're not kind of biased to those particular data sets. As I said before, that's US specific data. So we're working at the moment with Queensland Health to try and train models specifically on local data, so the system is then in tune with our particular needs and wants. But yeah, I think in those cases, there's a pretty rigorous ethics process to ensure that people who are providing that data have given consent to do it.

Yeah, and that is a big effort.

Naomi, am I allowed to say something?

Of course.

I do a lot of the…I help with a lot of the consent processes on projects in the team. And something I'd really like to say as a person who came from a health background is, you know, we're all patients. Everyone now has the right to ask where their data is going as a patient, as a carer, as a clinician, whoever you are, you are allowed to say, can you disclose where my data is going? That is a really important part of the Australian privacy principles. So I just wanted everyone to know that. Where is my data going?

And Megan, and this is also for you, Liesel, so you know. Enjoy your time in the spotlight. Older people can get very stressed using basic technology and often just give up navigating complex government aged care systems, policy and administration. Not my mum and dad, but... Is there anything in development that can help them understand get the help them get more support services for doing this kind of thing?

That is a fabulous question and I actually don't think it's just older people. I think we've got some of the most complicated processes in our country. And I think that there is a real role, what I have started seeing, which I was really excited to see,
some of, particularly in aged care, actually, some people are beginning to hire technology officers and those kinds of roles. So I think there is absolutely a whole new workforce. We talk about AI taking away jobs. It's not taking away jobs from what I can see. There are lots of new kind of positions that actually could
be made available to assist people with absolutely those kinds of things.

That's wonderful to hear. And David, I think this is one for you from Bianca. What level of attention is being given to training clinical staff and perhaps, you know, the entire workforce on what AI tools are capable of, how to integrate them into practice workflows? where they perform well and where they are prone to producing problematic or unreliable results. Are we doing any work in that space?

So I think the healthcare system has obviously undergone a lot of digital transformation over the past 20, 30 years. But in particular, the introduction of electronic medical records in all of our hospitals, which is, you know, probably started 15 or so years ago and continues. And, you know, that training them, training people on that, on those systems is really important, but just as important is the kind of digital literacy. And that is happening in various ways and at various stages across the healthcare system. But we've seen that, you know, obviously the interest in AI and Generif AI since the introduction of ChatGPT over the last few years is driving that. And we're seeing all of the colleges now, colleges of health and medicine, you know, be providing their members with information on how to safe use of AI and things like that. So, it's definitely underway. You know, could it be better? Probably, but I know everyone's taking it seriously.

Thanks, David. So, if we've still got Yan on the line, this is definitely one for her from Anton who asks, what's the process to partner with you on SAMD research assessment and approval software as a medical device. For those of you unfamiliar with that acronym.
Do we have Yan.

1:02:43
Hi, can you see me or you can hear me? OK.

1:02:44
Oh, yeah, yeah, we can hear you…We can hear you, go ahead.

1:02:50
All right. Okay, yeah, talk to us. There is a, I imagine that we would have our contact details in the AI report. So shoot us an e-mail. We can definitely discuss about it.

1:03:01
We absolutely do, Yan.

Fabulous.
Look forward to hearing from you, Anton. Okay, so we've had some submitted questions and I might move on to some of those now. What's the difference, and anyone jump in here, between treating AI as a technology project and treating it as organisational transformation?

Well, I think that's generally how healthcare has been treating digital projects. It's not just a digital project. These are health translation projects, and we need to think about AI in the same way that we've thought about those projects. I guess, Kylynn, for any thoughts on that? from the work with Sparked?

Yeah, I think, just like David said, I think it is with all digital projects. And I think for AI to deliver its maximum potential, we really need to think about when we implement it, not just consider it the technology, but also the context in which it's being used. And this includes the system that it's working within the care team, like the whole care team, as well as the consumer and their carers themselves. If we don't bring them all together to do this, we may end up with a really shiny tool that no one wants to use or it's not safe to use or actually is more work. So I think we need to really think about how it fits in all the parts. And I think that's really important with, you know, we tend to think about these projects and what's happening with AI in terms of the clinicians, but patients as well in terms of digital transformation. We spend a lot of time, Liesel, working with hospitals around Queensland on implementing or supporting them to use our mother tool, which is for women with gestational diabetes. What are we learning about? Are our patients are embracing that sort of thing? The appetite is definitely there. I would say one of the biggest learnings we've had, which I think we see across many of our projects, is that the consumer must be involved. Right from the start, if you do not involve your consumers or your end user, whoever that is, right from the get-go,
There's just no point. They must be, you know, they must be your champions. They must be the people informing the work that you do, the changes you make, the upgrades, the research has to be there.

Now, this is perhaps one for David. Where do you see the greatest near-term value from AI? Breakthrough clinical innovation or making health systems work more effectively?

Yeah, look, really good question. And we're seeing AI as we think today are used right across the spectrum from clinical innovation, health service efficiency and medical research, of course, where I think it's going to play an increasing role in driving efficiencies in medical research. But often digital innovations, and I think it'll be the same for AI innovations, both sides of that coin. So yes, they'll drive efficiencies, but they'll also drive clinical improvement. And so I think that's a big part of when we think about an AI project, how we make sure that it does both. And so that sort of gentle interplay between the two.

Would you say that's similar for this question here? So are there examples of AI replacing systemic burden on healthcare providers or is it primarily being used to accelerate existing workflows? And I think what you're saying is it's never primarily one thing. It's always that nice balance, isn't it?

Well, it's often things will have a primary purpose, but we need to make sure that it doesn't, you know, increase burden in other areas. So, and that can happen with digital projects, as we all know.

I love these questions so much: Do we have a library of AI so we can search for a reliable AI tool according to what we need? Sounds like a Harry Potter. I think we're all used to going to the App Store and searching for the apps that we want, and so...

Look, I don't think yet, but I know this is something which organisations, particularly big organisations, need to be thinking about is the AI governance of how AI is used in their organisations. And so having a suite of AI tools that have been tested and
and implemented in the workflow gives confidence, I think will be part of AI tools in organisations in the future. And I know all of our health departments around Australia are implementing AI governance policies and procedures, and so that will start to...that, I guess, put in place that sort of infrastructure that can support the adoption of AI tools across the health system. Bevan, you might want to add something there about AI evaluation and managing upgrades.

Yeah, that, but even before that, I think what David's been talking about in terms of different tools working together, that's really where our interoperability work and the work that Spark does is really important because provides the foundation that all those tools can interact and access the same data. On the evaluation piece, that's really been a strong focus. Lately, there's been a strong growth, obviously, in AI adoption, but we've lagged behind in terms of evaluation. And some of that's actually because it's quite tricky to evaluate these tools to understand, to assess where they're working well or not, because they're doing complex tasks, and how do we map them against what a best practice would be? I think this is actually a strong future line of research around AI evaluation and actually having specific agents that are there just to stress test different scenarios and to run simulations and to, you know, critically assess these tools. So I think that work is in early stages, but it's really growing and promising.

Denis, I reckon this one has your name on it in very shiny lights. Looking to the future, how soon will AI be capable of using the very latest in genetic research to identify genetic and polygenetic medical conditions and be able to contribute to that research itself?

We already doing it. Is the short answer. The longer answer though, the devil is in the detail. So obviously these frontier models jumped on genomics and proteomics are sort of the first area to tackle because there's lots of repetitive work that can be automated and AI is really good at that. But going deeper into actually doing something that impacts on health, that requires understanding. And as we said earlier, understanding is currently lacking.
So to me, there's a huge area of new AI research that needs to happen where the 3 billion letters in our genome and each one of those connections between those mind-blowing complexity, that is what needs to be tackled in order to really do something about complex diseases.

Okay, now I've got a quick fire question for the whole of the panel. Panel, what gives you the most tingles - I love that idea of like tingling - about AI's potential role in scientific discovery over the coming decade? Bevan, we might start with you. What gives you the tingles?

Yeah. specifically for scientific discovery? [audience laughs]

Correct, yes.

I guess it's the ability to really run long scenarios of simulations that we wouldn't otherwise be able to do, to propose you have some idea, some hypothesis.
and you want to test this and understand whether it's fruitful or not, to be able to fire up an agent or a group of agents that are able to trawl through all the information that's out there and look for the little pieces that might connect together and give them back to you, and then the ability to maybe even run experiments to say, test this out for me.
So it's really like one scientist or one person has suddenly inherited a super lab that they can, you know, that they can test all different things very quickly. And I think that will really accelerate our ability to kind of discover new things. So Bevan and his super lab.

What about you, Kylynn?

Well, my work is kind of less focused around the, you know, scientific discovery. But for me, I think really it's the problem we're trying to solve, which is really empowering our clinicians and our consumers, or our patients, to have the information that they need to make decisions about their health care. So I think maybe I'm a bit more outcome focused, but definitely I can see a huge potential there for that going forward.
Excellent. So I think when it comes to scientific discovery, Denis talked about it a little bit at the beginning, or maybe it was halfway through with Anne's talk, and that's large language models for our genome. And I think, you know, there's some really interesting work, and maybe it's not just our genome. You know, genetics and the use of genome technologies in healthcare
It's still really, it's not in its infancy, but there's lots of lots of new things that can possibly do, you know, when we're thinking about antimicrobial resistance, for instance, genetic technologies for have the ability to create cures for people, for individuals with a particular infections and that's going to really need AI. So I think there's lots of things that are going to happen in that space. So much for the quick fire. Liesel. I rather like Kylynn pinching off that.
I really want to see what science can do to further make people more of an expert in their health than they already are.
Empowerment.
Yeah, tapping into that exactly. This idea of democratization, that you don't have to be a super expert in one area to contribute, as long as you are enabled through AI to actually speak the same language as the rest of the experts, then ideas can flow more easily. Thanks, Denis. Now, we had a few questions. on the chat about the economic consequences of using AI. What skills should leaders invest in today, David, that will still matter five years from now? Well, I've got a quick fire answer for this one, which is just the literacy side of things. So digital AI literacy. So, and I think that's going to, I think, well, maybe I'll make it a longer answer. So digital literacy - this has been something we've been talking about for a long time, particularly as we introduce new digital technologies into hospitals. But people have to, and there's still errors that can happen. There's some great examples in the literature and on YouTube, of course, of errors happening in the way people use digital technologies. It's going to be the same with AI, but maybe even on steroids because we really don't know exactly how they came up with that answer. So, you know, having the literacy to really understand, well, how did they come up with that answer and is it reasonable is probably going to be something which is going to stand as in good stead for a long time to come.
And Liesel, I think this one's for you. As AI capabilities continue to evolve rapidly, how can universities, industry and organisations all work together to ensure graduates develop not only the technical AI skills, but also the practical experience and ethical judgment, and I think I'll underline that one, needed to deploy AI successfully?
I feel like there needs to be a really shared understanding of this direction to make them all work together, but I would very much go towards slightly different to David, but I feel like it's incredibly important that people moving forward are encouraged to have skills of creativity, critical thinking,
literature, communication skills, all the stuff that we don't necessarily associate with AI. But the reason we got here is because people have those skills. That's how we got here by being creative. So I think that's really, really important. Yay. Bevan, anything to add?
Yeah, I just think we'll just need to continue to...Think about how we...how we teach people to think more strategically, to think more long term, to think more along, you know, what the particular goals that they're after and really have a more strategic approach to things there.
Thanks, Bevan. And David, as AI becomes embedded in everyday work, what should organisations be doing today to protect privacy and maintain trust? And you know, I think that this is at the front of everyone on the call's minds. Yeah, look, you know, lots, I guess, is 1 answer. There's your answer, people.
But in some ways it also comes back to our responsibilities and using digital in terms of our passwords, making sure that we do take the hint to change our passwords and make it secure, and understanding where our data goes. As Liesel said, feel free, you know, we should be empowered as consumers ask where our data is going. And this is our final pre-submitted question. There's a lot of AI solution work going on. How do we best coordinate efforts to better use investment and share the learnings?
I mean, I think I think this is a good one for you, as a leader in AI. Look, I can I, I'm gonna use my team as an example, which is that, you know, we've got over 100 scientists and engineers across the Australian eHealth Research Centre, and they do a great job of talking to each other about how they can leverage each other's technologies. So, yes, there's a lot you can do using digital and everything, but it does come back to teamwork and working together, and then we're very lucky.
Well, maybe not lucky, but we're fortunate that we are leading the Sparked program, and that gives us the ability to interact with, well, over 1500 people who are part of that community. And that also drives a lot of our research and development, as do all of the fabulous clinicians, particularly across Queensland Health. And I mentioned Queensland Health. This is a joint venture partner with CSIRO in the Australian Health Research Centre, and across Queensland health, working with clinicians that we've built relationships with over a long time.
So I can see in the audience, we've got a lot of our people, but I would like to give the audience an opportunity here today to ask any questions that are burning in their minds at the moment after this wonderful discussion. So does anyone have a question that they would like to put to the panel? I can see one over there.
Yeah, hi, my name's Dan Ken. I just wanted to ask the question, sort of touching on the economic side of things. Do we think that into the future, because of the need for large data centres and all the other sort of stuff that comes with building an AI infrastructure, whether that has a flow-on effect to the cost of healthcare rather than a downward pressure and upward pressure?

Yeah, I guess it is. It's a cost rising in one area. I don't know how much it is relative to the savings in another area. And that's not specific to healthcare either. And perhaps healthcare is a bit better because any small change you do in terms of efficient.
efficiency in healthcare often is a massive change. So other sectors, I guess, that are head of healthcare are seeing that, you know, it is worthwhile. So I would expect it to see in healthcare as well.
I think in the safety and quality literature, they talk about safety, quality and efficiency as being, you know, three corners. If you increase safety, you increase quality and often increase efficiency. And a lot of that, I think, can be done with AI. So there's opportunities to make our health system more efficient.
and you know, we're not, there's plenty of demand for health services, and so, so you know, making it more efficient means hopefully we can we can really meet that that demand with perhaps, you know, less money.

Thanks. Sorry, I've got another question if that's all right. Just talking about preventative healthcare. Obviously, our healthcare system is built around people generally visiting services and actually being treated for unwellness or illness in that way. Do you see a large proportion of shifting towards a preventative healthcare sort of approach with an artificial intelligence sort of mechanism, agentic AI and the agents that exist?
Do you want to have a go at Liesel? Yeah, maybe Denis too, actually. I do hope that that will be the case. I think this is part of that, you know, BYO data in my mind is actually that people will have the ability to measure their own data to some degree, consult on that with their healthcare providers and then reduce some of the inefficiencies that currently exist. One example I can think of immediately is people who have to go to their care providers for blood pressure measurements. A very simple thing. Some people have to do it very, very regularly. Could they actually just be swapping that for doing it at home themselves, taking the data to more sparse appointments, reduces, you know, reduces the impact in other areas, that kind of thing. Denis, did you want to add? Yeah, some of the risk obviously is encoded in your genome, and therefore, if that is routinely used in order to assess what your susceptibilities are, and then your treatment, long-term treatment plan assesses that or takes that into consideration then maybe some of the chronic diseases can even be avoided altogether in the 1st place. But coming back to our understanding is so little at this stage that even AI in its current form can help that. So it's ongoing, traditional healthcare research that needs to keep going. And then, hopefully, AI can accelerate that.
There's a technical advance there too, which is in the prediction space, we used to have to train individual model to try and predict this one thing, then individual model to try and predict that thing. Nowadays, these general purpose models are able to work across a whole suite of different ways, which makes prediction much easier.
you no longer have to try and adjust on that, just get that data. So the ability to build predictive models is much, much easier with foundational models as well. So yeah. But I think one quick thing on that also, Liesel, with the gestational diabetes work, where we're supporting women with gestational diabetes, once they've had their baby, do they then take on more prevention to getting further diabetes? Yeah, so this is an area that we're actually doing a lot of preliminary work in at the moment. It's such an important space because a lot of women aren't actually aware that their long-term risks after having gestational diabetes are really significant. So we're looking at a couple of bodies of work here of discharge planning from our mother platform looking at women being able to take the data onwards into their care planning in the future and subsequent pregnancies. It's a really, really important space that we're looking at and needs to be done.
Any other questions from the audience?
Right.
I have multiple of mine, so...
Let's go to this one. This is a really important question. Given Australia is a country of high number of immigrants, sorry, migrants. I always can mix up these with migrants. And we, well, one comes in and one goes out. Yeah, they're both, aren't they? Yeah, okay. PhD in literature. [audience laughs]
And we know race, genetics, gender, age, et cetera, affect one's health differently. Does CSIRO AI work always used to be identified or synthetic data? Liesel, I think this is your one.

We use all of the above. Actually, it's very project dependent.
go to great lengths to understand when we design research projects, whether we will need live data, whether we will need synthetic data, whether we will need demographic information if we think it's going to impact on a particular issue or solution that we're looking for or question. So it's actually both. But wherever possible, we do actually use de-identified data, whether it's synthetic or not. I'm just going to squeeze two more questions in because we're high achievers here.

Anna has asked, in contrast to the question of what gives you the tingles - what makes you anxious? In fact, what doesn't give you tingles? What gives you shingles. [audience laughs]
Bevan, do you want to start? Quick fire. Quick fire.

Yeah, I think it's around deploying things that haven't been scientifically validated. You know, we need to test these things the way we test new drugs. You know, it needs to go through a careful process like the QMS process, or, or yeah, evaluated rigorously, and that's why programs like ours are so important, isn't it,

Denis?

Currently, AI is probably the cheapest that it will ever be. And therefore, what I'm a little bit worried about is that we're building all these innovation infrastructure and systems around it. And then at the end, we find out that it was never sustainable. That is something we need to consider.

Liesel, what gives you the shingles?

I really worry that people think that digital solutions like AI are the be all and end all, and we actually need the research evaluation, as all of us have said multiple times, to actually prove whether that, whether something works. It can't be an assumption.

Hear, hear.

Kylynn? I think for me, it's the same thing that AI is the be all and end all, but
for us really having good quality data for the AI to learn on. So we can't just rely on AI if AI is just feeding AI and that's a bit scary for me. Yeah, maybe similar in just the easy acceptance of AI and using it without thinking. And I think it's continued to be important to think about.

Okay, one last question for the panel. What's the most important thing organisations should focus on if they want to move the theme of today's webinar from AI promise to AI practice? And let's start with Denis. People. People, people, people. Right? It boils down to if AI is a...great equalizer, it comes down to individuals, their motivation, their drive, and therefore nurturing people to me is the key.

I was going to say consumers.
Look at that AI governance within an organisation and we want to make sure we're enabling the use of AI, the safe and responsible use of AI across organisations. So the governance has got to be fit for purpose and really supportive. I think it's remembering the problem we're trying to solve, so not just AI for the sake of AI, but actually what are we trying to fix here?

Hire more geeks. [audience laughs]

I bet it's Denis's point. It's around having the right people there, you know, the right technical people for technical problems. Where people who need people.

Thank you, Bevan. Okay, over to you, David. Any closing remarks?

I will, but I'm going to steal them off you. Thank you. Look, thanks. Thanks. Today's discussion really enforces reinforces a key a key point we're making in our report that AI is no longer emerging technology waiting on the sidelines. It's being deployed in real world settings and delivering value. And the challenge now is ensuring the value is realised responsibly, supported by evidence, strong governance, interoperability, of course, and collaboration. My thanks to Bevan, Liesel, Denis, Kylynn and Naomi for helping, doing a great job up here on the panel. Special thanks to Jess Hildyard.
from CSIRO's Industry Engagement team and the M3 tech guys we've got up there running the audio and video. Thank you to everyone who joined us online and in person. A recording and additional resources will be made available after the event. Don't forget to download our AI report.
And thank you very much for being part of a great CSIRO conversations. Enjoy the rest of your day.

Hit closing song! [Bad Medicine by Bon Jovi plays]