AI in Healthcare: Are we ready for data to examine us?
Joachim Schultze (00:04):
Yes. So many good ideas about applying AI to medicine out there. A lot of startups and very cool young people that are in academic institutions. There's absolute amazing things happening, but what's missing all the time is the good data.
Michael Bird (00:22):
Okay, so you've heard it from us before on this show, but we'll say it again. AI is transforming our world. In almost every sector, humans are being supplemented with tools which release them from some of the more burdensome tasks of their job and allow them to make the most of their human skills for creativity and lateral or strategic thinking. Basically, AI is fast becoming a way to elevate human expertise.
Aubrey Lovell (00:48):
And that help couldn't come at a better time. Research from the American Psychological Association suggests burnout is in an all time high, and that's especially true in the medical profession. After years of disruption due to COVID and ever present cost saving measures, many health services are struggling to keep up with their workload, and healthcare professionals need all the help they can get. And that's what we're going to be looking at in this episode, the ways in which AI is revolutionizing healthcare, both in terms of diagnosis and treatment, but also in terms of bringing together and sharing human expertise.
You are listening to Technology Untangled, a show which looks at the rapid evolution of technology and unravels the way it's changing our world. We're your hosts, Aubrey Lovell ...
Michael Bird (01:46):
And Michael Bird.
Now, depending on which market research you look at, AI in healthcare is already somewhere from a $14 to $21 billion industry in 2023. Those stats, by the way, are from MarketsandMarkets, and Statista, respectively, and you'll find links to their reports in the show notes alongside the American Psychological Association Burnout Report and many of the other stats that we quote in this episode. Either way, that's around double what it was worth just two years ago.
By 2028, it's set to be a $100 billion global industry, growing some 40% year-on-year. That is astonishing, even in the already skyrocketing AI sphere. So in this episode, we'll be looking at a wide spectrum of expertise to get a sense of where the field is right now, what the future looks like, and some of the cool technologies which might fill it. So let's start with the present. AI in healthcare is an exploding field right now. So where did it all start? Well, Mike Woodacre is the Chief Technology Officer for High Performance Computing and AI at Hewlett Packard Enterprise.
Mike Woodacre (03:06):
We've just come through the COVID pandemic over the last few years, and actually in a way it was almost perfectly timed for taking advantage of high performance computing and AI capabilities to accelerate our ability to understand COVID, and then create treatments and vaccines. People were probably willing to take bigger leaps of faith in trying things. We've now got HPC simulations and AI models very tightly coupled together. So we've used generative AI to create new molecules and which one is potentially helpful to attack the COVID challenge.
We've seen the revolution in large language models, which again allow you to bring together huge amounts of text information. Obviously during COVID, there were tens of thousands of papers published in the medical field around COVID. No one human can analyze all that, read all those articles and make the connections, but we can use natural language processing techniques to analyze that that potentially would otherwise take years. So I think this has just been a huge acceleration that COVID has pushed forward that we can now apply in many other domains.
Michael Bird (04:35):
I don't know about you, Aubrey, but I remember at the beginning of the pandemic in those early days when the world basically changed overnight, I remember thinking to myself, I wonder what good will come out of this. Because in interesting and strange times, that's when innovation really comes out.
Aubrey Lovell (04:53):
You have these moments in time in history where it's obviously crazy what's going on in the world when we had COVID and things are closing and businesses are closing. But at the same time, we did have this time to really look at our technology and say, how do we figure this out?
Michael Bird (05:08):
There was this rapid technology growth in basically every sector because they had to figure out how can we make technology work for our organization? So I suppose when we're looking at healthcare like we are here, obviously hospitals are open, but they were really having to think, how can technology help us to be more efficient? How can technology help us to do this in ways that we weren't able to do before?
Aubrey Lovell (05:30):
The gap between humans and technology was minimized. And we talk about that almost like revolution. You think about our industrial revolution, this is almost kind of like a period where we have a technology revolution. It's kind of been kicked off and it was a mixed sense of trying to kind of embed that help in public urgency. So what's the state of the field right now?
GE Healthcare is one of the leading providers of medical systems and services in the world. Being general manager for Northern Europe and living in the UK, Andy Cachaldora works heavily with Britain's National Health Service or NHS. Though I think it's safe to say the core pressures and demands of the UK healthcare system are mirrored elsewhere in the world. So what's driving them?
Andy Cachaldora (06:19):
First of all, product development is one of those in terms of making sure that we're market leaders, but also the demand in the marketplace. Everybody's well aware of the pressures that there are in the NHS and in healthcare systems where there's just not enough diagnostic staff to cope with demand that's coming. And as a result of that, we're having to think of more creative ways on how do we speed up diagnostics with the same resources that NHS have today, and more importantly, with the eye of making sure that we can improve patient outcomes.
When we originally looked at AI in diagnostics, we're looking at image interpretation and assistance in terms of trying to make clinical decision supports on whether patients have cancer or not or other specific diseases to now actually as a result of COVID and backlog, et cetera, AI is more embedded into the workflow to speed up that workflow. So it's not in one piece of where GE plays, it's actually everywhere now even right down to the MRs, the CTs, the ultrasound, the x-ray equipment that you use today. So there's AI there in terms of image interpretation or understanding how equipment is being utilized and when it might need a service before it breaks down so that you can keep the equipment running as long as possible.
Michael Bird (07:35):
And incidentally, we've linked a British Medical Association article in the show notes, which backs up what Andy was saying about the shortage of diagnostic stuff. Now I'd love to pretend that was some AI driven MRI scanner beeping in the background, but it was actually the office dishwasher, which if it had a little more intelligence, wouldn't have interrupted us anyway. It won't happen again.
What I thought was really interesting with what Andy said was how AI was sort of augmenting rather than replacing people, healthcare professionals. In particular I think it's fascinating, the concept of AI analyzing images, flagging stuff to a professional and the sense that actually with the overstretched and understaffed healthcare systems we have across the world, that can massively help with that and actually allow the stuff that these professionals are really good at, let them do that and let the AI deal with the grunt work, which I think is really interesting.
Aubrey Lovell (08:32):
Yeah, I totally agree. I think my spouse is a healthcare professional in the emergency room. So from a personal perspective, I can see how much they have to get through day by day and how quickly they need to make decisions. And there's so many processes and steps that need to be taken in order to ensure that things are done right. So if you can have an extra set of eyes or almost like a additional support as you go through that process to make it easier for you to do your job, absolutely. It's a huge benefit, not only to the provider, but also to the patient. You're going to get faster results, there's going to be faster decisions being made to help them in whatever that they need.
That's something Mike Woodacre is keen to point out. The drive here isn't to replace, but to augment people's expertise.
Mike Woodacre (09:24):
I think many people are familiar with things like image analysis. Imaging is one of the great technologies that healthcare had been advancing over many decades. But then suddenly with AI, you can train the image analysis to actually have accuracy that's in a sustained sense, more accurate than a human expert looking at images. Now, one of the key things to remember is AI does not replace humans. It enables humans to be more efficient. So for example, in that image analysis scenario, the AI can be very good at recognizing patterns it's seen before, but if there's a new combination that's showing up, something new, that's where the human can really be used to steer AI in a different direction to pick up new signals.
Aubrey Lovell (10:14):
And that brings up something interesting because the melding of AI and human expertise doesn't just stop at AI taking out the easy cases before pointing to something it doesn't understand and alerting the expert. It's actually bringing together countless decades of medical experience and providing it at your fingertips. After all, any AI requires training, and the training in this case is provided by medical experts for use by their peers.
Michael Bird (10:44):
It turns out that the reason AI imaging works so well is that the expertise of a dozen doctors are better than one at analyzing images, a kind of swarm intelligence or swarm learning experience, if you like. And that drive to train AI has spawned some really cool research all on its own.
Rutwik Shah (11:03):
My name is Rutwik Shah, and the work that we'll largely speak about today, which is the application of swarm intelligence, is from work that I did as a researcher, as a physician scientist at the Center for Intelligent Imaging, largely focused on building machine learning and AI solutions in the medical imaging space. But one very key difficult challenge that we encountered is that there could be a very high degree of disagreement on especially challenging diagnostic cases. And this is where the idea of swarm intelligence as an application came in, where you can actually form human forms to find better and accurate answers to any set of questions. So when we kicked off the study, we started with creating a data set of knee MRI scans and we recruited two different cohorts of radiologists. So one were senior radiologists with at least 10 to 15 years of experience.
And then the other cohort was a set of junior doctors, if you will, who were resident radiologists or doctors in training. The senior physicians were a group of three physicians and then the junior physicians or radiologists were a group of five physicians. With having done this, we asked them a very simple question on the set of MRI scans that we provided them with while participating on an active swarm session on a platform that was built out by one of our collaborators. And you have the swarm or the human swarm collectively move a central part, which is on the user interface to the answer of choice that is decided by the collective intent of the group in itself.
What we realized towards the end is that the swarm answers for both these cohorts outperformed individual answers of all the individual participants. Both of these cohorts also outperformed a state of the AI solution that was built in the space. And more interestingly, again, we also found the fact that if you have this experience to junior physicians, but if you have them in enough numbers, you could actually get answers of the same accuracy and caliber as that of the senior physicians with the use of experience.
Michael Bird (13:17):
Gosh, that is absolutely fascinating.
Aubrey Lovell (13:20):
It really is. If you think about the response time to be able to identify something that early, it could literally save someone's life. But there is a question around training and accuracy, and it's something that we've covered to a degree in episode one of this series of technology entangled where we talk about bias in AI, right, bias in bias out. And that is that if certain factors are included in an AI's training upfront or it doesn't have a particularly diverse dataset, it can lead to inaccurate outputs.
Michael Bird (13:52):
Absolutely. And you'll find a link to that episode along with a link to Rutwick's research in the show notes.
AI trained on a poor dataset is bad news in any application, but it's particularly important in healthcare. An AI solution trained using data or expertise gathered in one part of the world may not work in places where different diseases are prevalent or where people have different risk factors. It's something that HPE have been thinking a lot across the AI space, and healthcare is of course no exception. Fortunately, swarm learning is finding a use in this space too. Here's Mike Woodacre.
Mike Woodacre (14:33):
If you think you've got, say a hospital in London, we've got one in Chicago, we've got one in Bangalore. If you look at the people going in to get scans of lungs, you'll see very different rates of different diseases showing up in those scans in those different places. So if you just trained your image analysis model in that one location, you'd obviously get very accurate on the high frequency cases you see in that location. But occasionally when you get that one that doesn't show up very much in Chicago, that shows up a lot in Bangalore, you would miss it.
But with swarm learning, what we can do is train models with the local dataset in London, Chicago, Bangalore, but then we can combine the metadata information, the model parameters, the weights that go into those models without sharing that base data. So you're protecting the privacy of the patient data, but by combining the model information together, you're effectively getting a model that can be deployed globally that's trained on that global data set.
Aubrey Lovell (15:44):
There's a small stipulation though. Data and in particular healthcare data, is sensitive and valuable, so moving it around the world presents a significant ethical and legal issue around data protection. It's also potentially not very secure and risks someone collating all that publicly gathered data and then hoarding it for their own benefit. There's a fascinating and elegant solution in the works though.
HPE has been working alongside a global group of researchers and providers specializing in leukemia to develop a blockchain-based method for moving insight and learning around the world. It can be used to accurately train AI on a global scale. Mike Woodacre was involved in the project alongside Joachim Schultze, Professor of Systems Medicine from the German Center for Neurodegenerative Medicine.
Joachim Schultze (16:33):
Of course, leukemia is the cancer of the blood. We collected the data from many colleagues and many groups in the world to learn whether we could apply such machine learning approaches to help us to classify these diseases. And after all this enormous work with very dedicated young scientists in my group that collected the data and curated the data and so on, we really could show that it is possible to use this very modern data in medicine to better diagnose, differential diagnose these diseases and that's how it started. But it was very clear that the biggest problem was it took much too long to collect all this data to a central space. We had data protection rights and was not easy to get data from one country to the next one. Data protection for me is a good thing. I think medical data should be protected.
That was one of the motivation, not the only one. There's others as well that we said what it would be good that if we could push the algorithms to the data into the hospitals and then locally, we can already calculate things because that would automatically respect GDPR regulations. It also would reduce data transfer, data application, so there's other issues that would have been solved by that. And we had done a catalog of these aspects, what we would like to have technology look like, that it fits to medicine rather than medicine has to change that it fits to technology, and that is the system of being completely decentralized, having access but not needing to share data, but rather having an algorithm that could be applied locally and then somehow the information can be integrated.
This is when we came into contact with our colleagues in collaborators at Hewlett Packard Enterprise, the AI team there had actually an idea about something like that out of a different motivation. They were more like about energy consumption and saving energy. And that's why we said we have use cases. You might not have thought about it because you're a tech company, but they're very interesting because this is a big, big, big point that we have to resolve in medicine. How can we access the data but not sharing the data and certainly not expose data privacy and actually generate accuracy of machine learning that is translatable into clinical practice. And that I think is a big achievement.
Michael Bird (19:03):
That's an incredible idea. A decentralized global network of swarm learning nodes, sending their insights to one another to better train AI systems. And we've linked the team's research, which was published in the prestigious journal Nature, in the show notes. But there is a chink in the armor because in order to get good insights, you still need to provide good data. That's true across all AI applications in healthcare and beyond.
The thing is though, good record keeping means standardization. Now in the UK for example, most GPs or family doctors are now registered with any one of two or three systems that mean patient records are kept digitally and are unified. But whilst some data can be shared between the different systems, they aren't exactly the same. And that creates issues for AI solutions, and it's something that companies like GE Healthcare are trying to overcome.
Andy Cachaldora (20:06):
You need data that is collected correctly to create the algorithm. And also you've got to think the NHS as a whole and social care are very tribal in terms of the way that they work. Every single NHS trust, every primary care, every GP has different methods of collecting data. They don't see the wider picture initially, and that's why it's absolutely crucial that the awareness of collecting that data accurately from the first point of interaction with the patient.
Now, what we have done in GE is in standardization of protocols. So people collect the same data in the same way that adds value to the algorithm or the development you're looking for. One of the key things is that GE now also provides services around data assessment. So we can actually look at the data and understand whether the data is labeled, curated and structured, ready for AI, and if it's not, then we can provide services to help them do that before they make that investment. That's why the business cases for these transformation and the commercial models need to be designed upfront before they're implemented so people can understand what they're buying into and what are the benefits downstream.
Aubrey Lovell (21:22):
That's something that's core to Joachim and his team's efforts. The idea that data should be private and secure but also be available, and it's something their decentralized blockchain driven solution accounts for, while also potentially changing the way the medical industry views collaboration in the financial value of their data and insights. And that's a solution which could be better for everyone in the long run.
Joachim Schultze (21:46):
The technology looks complicated, but there's a couple of very simple principles. The first one is the blockchain connects hospitals or medical institutions, and it's not like a public blockchain. This is completely different to the one that we are using. Ours is called privacy permissioned. So there's private because only the institutions that should be part of such a network so that they have to have some credibility, they need to be medical institutions that know how to handle medical data.
So that makes the situation private and permissioned. You're not allowed to just enter that blockchain. You need to say, I as an institution want to contribute to make a certain diagnosis better. I think that's important to say because then we can keep these blockchains very small and we can have those people that really want to make a change and make the things better rather than anybody that might have different opinions or different ideas.
The blockchain allows us to exchange information between institutions in a very secure way because everything that is exchanged is locked and is immutable, meaning you cannot change the information anymore. Now, on site is the nice thing that every institution that would join such a small network would install very important software that you can run. On the one hand, the machine learning algorithms, and on the other hand, there's a whole stack of software that makes sure that everything in that calculation is super secure. On top of that, this swarm board is fit within the institution, but it's completely separated from the data of the hospital because we don't want to tap into any data that might have data that could be linked back to the person. Every trace of private data is deleted from that data. It's then aggregated so that you really cannot go back to a single patient.
And then it's used for this machine learning and only the results of the machine learning. These are the ones that are shared with the other hospitals to see how are the different hospitals doing and which parameters actually are the best ones to distinguish, for example, between a leukemia and a healthy person. So what we are sharing is not data. We're sharing only the insights, the results or the information that comes out of our local learnings. And what we see is that the hospitals understand that they might become a share of the market again.
In medicine for all time, those that produce the data, which are the hospitals and the medical institutions, have been kind of exploited. Of course, they've been working with industry together to do clinical trials, but then the medical device is sold by another company. Despite the fact if you go for data and AI, the data is actually still in the hospital. I think if people understand that it's better to have a network of equal institutions of high quality and you improve your medical test based on AI in a continuous fashion, even if you're not having the full share of that thing, you gain more than not engaging at all.
Michael Bird (24:58):
So what's quickly becoming clear is that sharing information is key here. Be that insights, expertise or just simply data. When that happens, AI can transform from a helpful aid to a system-wide platform to exponentially improve your performance.
Aubrey Lovell (25:16):
That's not just in reading scans either. It can fully transform the way a patient's journey through the healthcare system works from the moment they book an initial appointment or get in the car to go to the hospital, which sounds pretty futuristic, but it's something which GE Healthcare has been working on and testing for some time. Here's Andy.
Andy Cachaldora (25:34):
AI is not just in the diagnostics. Diagnostics tend to happen when patients become ill and then they turn up to A&E, or the problem statement's already happened. In other words, there's aches and pains and the patient has to be admitted to hospital. So there's multiple benefits, but actually AI is being used more at the front end. What we're trying to do is maximize the amount of scanning time that imaging staff can do. So we have some smart algorithms around predicting whether a patient's going to turn up for an appointment or not, and that takes different fields like anything from carpark capacity at the hospital and whether they're going to be running late for their appointment, to taking localized data feeds on traffic to see how far the patient lives away from the hospital and whether they're caught up in a traffic jam, to behaviors of patients, whether they've turned up for appointments or not, and then gives a specific profile or whether a patient is going to attend their meeting. And if they're not, then maybe we can reallocate that slot to another patient that could be waiting.
So you are maximizing your capacity in terms of its workflow and actually how diagnostics is and AI is being used now is more around the prevention rather than when a patient's already needing some form of intrusive therapy. So how do we stop people from getting late stages cancer and actually can we use AI to predict whether that's genomic sequencing or looking at historic data and looking for clinical signatures to identify, this patient's already potentially at risk of having some form of disease and what that kind of disease might be. Then as a result of that, then we can even stop late stages of cancer even happening. So that means survivorships for patients massively improve. It means that cost to systems and treatment paths are massively reduced. Even we just capture a patient from stage two to stage one means that just for that one patient, it's £100,000 saving in treatment.
Michael Bird (27:42):
So what's starting to become clear is that AI is an amazing tool in healthcare when it's trained with a spirit of collaboration. That's something that's not always in high supply when dealing with countless disparate organizations and individuals. But fortunately, there are teams of people working to break down those barriers. And when they do, remarkable things can happen, as in the teams of junior radiologists outperforming existing AI and experts in the Center for Intelligent Imaging study, as Rutwick Shah explains.
Rutwik Shah (28:13):
The AI performance that we were measuring against, in terms of accuracy was performing as the best radiologist in the group was. And then the swarm performance that we analyzed and assist was at least two weeks better than the state of the RTI performance. Why this happened is because the AI model that we benchmarked our solution against wasn't itself trained by just two individual radiologists working at two different time points versus the swarm, as I mentioned, was either a collective group of three senior physicians or a collective group of five junior physicians.
Then just because we had a quantitative advantage over the original AI training system, this was a huge reason why and why it was able to outperform the state of the RTI solution. So if the human swarm also is able to outperform AI trained by individual humans, it stands to reason that there is a much better approach in improving your AI performance, which is to have a swarm, human swarm train the AI such that there is a swarm AI to go with and that should be able to then have drastic improvements in its accuracy, reliability, and consistency.
And so that's the next step one, of where we want to take this research going forward. And then the second is a direct clinical application without creating an intermediate layer of EIO machine learning in between, which is if you have certain clinical decisions, especially diagnostic decisions. Now, one good example of this is especially in the radiology setting, is when interpreting indeterminate breast mammograms where you see a curious looking lesion on a breast mammogram. The current guidelines, at least in the European setting, is the fact that you do have to have at least two radiologists review the scan independently.
If there is a disagreement between their reads, then it goes to a third senior radiologist for arbitration nano final decision. It's still highly time inefficient because from the point of biopsy to the point of final diagnosis could be a few days just through the process that it has to go through, versus if you were to implement a process of swarm intelligence to begin with, an answer to this could have been obtained in the very first instance of the diagnostic read if there was more than one radiologist reading the scan.
Aubrey Lovell (30:46):
So AI and healthcare is a huge growth area where established players, laterally thinking researchers, and plucky startups are working together to build and train the next level of medical aids and operational efficiency drivers. So what's next for our guests? For Mike, it's about squaring the division between the positive of using AI to help people and the potential issues of using predictive AI to isolate people who may be high risk for certain conditions.
Mike Woodacre (31:14):
We can use AI techniques to really benefit the global population, but we have the ability in our hands to determine how we protect information, protect individual privacy issues. I think there have been movies made on that sort of topic where it's essentially when people are born, they get a DNA analysis is done and you can learn a lot about the trajectory people may have, and that's the big challenge is there may be certain probabilities of based on that DNA, how things turn out, but it doesn't exactly determine the future.
So what rules are we going to apply? And I think again, this is where, as a society we need to discuss these things and decide what's the right answer. It is a very tricky area, but this is one of the areas where technology and society really needs to work hand in hand as to what we want to do with this great capability we have.
Michael Bird (32:24):
For Andy, it's about building the business case for AI and making sure the right solutions are being provided for the right problems.
Andy Cachaldora (32:31):
Also, we've got to think about the change management because the technology is one thing, but the adoption by the users and the confidence in adopting it is absolutely crucial. There was fear originally about AI possibly taking over imaging staff as the main clinical decision support in terms of that first triage part. I think we're way off from that. But in terms of adoption, I think as a result of the backlog of COVID staff are just so stressed with the backlog that actually the only way is technology now. Nobody really wants to be working a 12-hour day, seven days a week. The only way you're going to be able to cope is actually using the technology, but we need to improve the confidence rate in that.
More importantly, does it commercially stack up? It could be really good AI, but actually if the cost per an exam is more expensive than a human intervention, you're never going to get it through your board in terms of approval. So I think there's some aspects in terms of making sure that it's financially viable. And also the other aspect of it is the ethics around that. How's it going to be implemented? Is it going to be in real world? Is it going to be in a test environment so that you can try and measure the outcomes and the accuracy of it? I think that will evolve over time. There are companies now that provide services in assessing AI and not the development. They actually test the algorithm for its quality and its accuracy.
Aubrey Lovell (33:58):
And for Joachim, quite simply, it's about sharing knowledge for the betterment of everyone.
Joachim Schultze (34:03):
We want to unite the hospitals of this world, and I just love this sentence. That's a dream, that's a vision, and we're working hard to contribute to that, but it needs a lot of people to make that happen. But it is an opportunity to really make our worldwide health better by having more people having an easy way to collaborate and still protecting their own rights.
Aubrey Lovell (34:30):
That's a lofty goal. And while there are still some issues to iron out around standardization of data and the ethics of responsibility when things go wrong with an AI assistant, it's clear the field is in a hugely exciting phase. And ultimately, that's good news for all of us.
Michael Bird (34:50):
You've been listening to Technology Untangled. We've been your hosts, Aubrey Lovell and Michael Bird, and a huge thanks to our guests, Mike Woodacre, Rutwick Shah, Andy Cachaldora, and Joachim Schultze. You can find more information on today's episode in the show notes. We're going to be taking a short, mid-season break, but don't worry, we will be back very, very soon. This is the sixth episode in the full series of Technology Untangled, and next time we are exploring non-traditional routes into tech careers. Do subscribe on your podcast app of choice so you don't miss out, and to check out the last three series.
Aubrey Lovell (35:24):
Today's episode was written and produced by Sam Samuel Datta-Paulin, Michael Bird, and myself, Aubrey Lovell. Sound design and editing was by Alex Bennett with production support from Harry Morton, Alicia Kempson, Alison Paisley, Alyssa Mitry, Camilla Patel, Alex Podmore, and Chloe Sewell. Technology. Entangled is a lower street production for Hewlett Packard Enterprise.