#312 Can we Create an AI Doctor? with Aldo Faisal, Professor in AI & Neuroscience at Imperial College

Generative AI is transforming industries at an unprecedented pace.

But as AI changes how you or your team work, one thing is clear. Your skills also need to evolve. At datacamp, we offer everything you or.

Your team need to adapt and thrive with AI.

Whether it's business users looking to get the Most out of ChatGPT and Copilot.

Or developers and data scientists looking to fine tune models.

You can learn the entire AI skills spectrum on Datacamp. Power your AI transformation today. Start learning@datacamp.com real big wave that's going to wash over the whole field is now we call ambient intelligence and the whole space of AI for operational improvements. So that's not medical activity per se, but for example, transcribing the conversation between a doctor and the patient. So we have already a basic model that is multimodal. It combines different forms of data and different timescales already and it shows that it can predict outcomes from for certain patients much, much more accurately than reference language, model derived reasoning or other forms of, you know, conventional machine learning based predictions, for example, about the mortality of a patient.

Welcome to Data Framed. This is Richie. One of the most exciting use cases of AI is in improving medical diagnoses. The last decade has seen real breakthroughs with AI interpreting chest X rays and detecting breast cancer with with better than human doctor performance. This is just the tip of the iceberg and with the advances in AI, I'm expecting many AI related medical breakthroughs in the near future. Our guest is Aldo Faisal who runs Nightingale AI. It's a research project to assist doctors with making complex decisions. Aldo runs human and AI research at Imperial Global usa. He's a professor of neuroscience and AI at Imperial College London. He's a founding director of the UK Research and Innovations AI for Healthcare Centers. He's digital chair at the University of Beiruth and he's the editor of several journals. In short, he's one of the top researchers on AI for healthcare. I want to know all about Nightingale AI. I want to know how it's built, how it will impact public health and the role of doctors, and what happens when AI makes a mistake and gets a diagnosis wrong. So let's find out how close we are to getting that AI doctor from Star Trek Voyager. Hi Aldo, welcome to the show.

Hi Richie, delighted to be here. Excellent.

So in the last year or so we've seen a lot of quite ambitious efforts to replicate employees from customer service people through to software engineers, data scientists, even. So how close are we to being able to have AI doctors?

I think it depends on what you mean by an AI doctor. Right. So on the one side, if you think about, like, the doctor from Star Trek Voyager that autonomously treats people and has improveable bedside manners, I think we're quite a bit away from there. And there's also a question whether we want that at all. But I think what we're going to see is AI technology that helps doctors and nurses and carers and clinicians in general treat patients better, faster and more accurately. And there's a number of technologies we can talk about that are already there. And there's a whole range of things upcoming on the horizon.

Okay, yeah, I mean, I suppose even like to keep track with the Star Trek timeline. We've still got a few hundred years to kind of replace the Voyager or recreate the Voyager. Doctor, can you give me some examples of, like, real things that we can do with AI for healthcare now? Like what can you actually. What are you up to?

Yeah, so what's fairly mature now is AI and radiology. So this is you're looking at a brain scan and detect a tumor, or you look at an X ray and look for shadows that may suggest that you have an infection or something else going on. And these things have been developed, they've been already partially licensed, and some leading institutions and hospitals have already adopted them in their workflow. This is really what's. What's really supporting already AI and diagnostics. I think the real big wave that's going to wash over the whole field is now we call ambient intelligence and the whole space of AI for operational improvements. So that's not medical activity per se, but, for example, transcribing the conversation between a doctor and the patient. And then not just transcribing that, you can do that already with the voice capture software, but also fitting information automatically into the record of the patient so that the doctors don't need to spend ages writing up things afterwards, maybe already making suggestions based on treatment guidelines, what drugs should be prescribed to this patient or what referrals should be made. So I think we're going to see a lot of this operational improvement coming. The technology's there, it's been already licensed first. Institutions are really here using these, and we're going to see, I think, this widespread and not just in top hospitals, but all the way to your local general practitioner. You visit the town after you've had a cold or something like that.

Okay, yeah, certainly those operational aspects, I feel like it's not the kind of the, the, the cool stuff, but it is incredibly important. Although I'M kind of wondering because doctors famously have like terrible handwriting. Like I, I don't know whether the, the technology works to can it read a doctor's handwriting and accurately transcribe that.

So this video, there's a fun study there actually that people did on doctors recording the record of a patient. And there's been evidence that when doctors went from handwritten notes on a paper card to the electronic record, that they have more difficulty recalling that specific patient if they're not looking at the picture of their handwriting that they did on the day. So there seems to be something with memory coming into place there that is quite interesting on the neuroscience side of it, so to speak. But I think in general what the AI can now do is provide you with the relevant context that you're checking on a patient and it helps you memorize people much better. It also has of course, handover between different doctors, so you're treated by a team of clinicians and experts to, to, to link things up better.

Okay, that's absolutely fascinating. And I think, yeah, in general, no one really likes writing notes about what they had a conversation about. So outsourcing this to AI seems like a very useful idea. I'd love to know about your Nightingale AI project. So what does this involve?

So Nightingale AI is meant to become a very, very large health model. So just like we have large language models that are trained on literally all the English texts that you can find on the web, and we do want to do that the same for electronic healthcare records and biomedical research data and medical information in general and the entire medical literature. And the challenges here is of course the medical data. How do you get hold of large amounts of medical data? Well, you need to cooperate with healthcare providers, you need to work with patients. And here, because we're based in the UK with Nightingale AI, we have of course the national healthcare system. So that's a one stop shop that provides all healthcare in the country, literally all doctors and nurses in the country work for the nhs. And it's really a support that you have from cradle to grave that three is the cost of at the point of care. So literally everyone is in it. And this system generates immense amounts of data on a daily basis. And there's of course a question which what are we doing with this data? And of course patients are concerned about their privacy and citizens don't want their data to be abused. But the idea is that we want to create an AI that learns about medicine in a different way from other things you may have heard about. May you've heard of large language models applied to medical data or medical information. And to me that's a bit like an English literature major who knows about text and language, reading about medicine and then answering questions about medicine. While we want to build models that are trained on multimodal data so electronic healthcare record, an X ray, a brain scan, genetic information, some lab test results that were done, some pathology that was done on slices of, I don't know, some blood test or something like that can ingest all of this information, understand and combine it and then reason out of the biology. So reason like medical doctor in some way should out of the science about the science. So that's the goal of Nightingale. And we're making some major steps towards we're bringing in data donations from partners for giving us data. We are working actively with NHS about how we can train and build such a system in a way that preserves the privacy of the data. Here we're talking about what we call federated machine learning. And so there's a whole range of development that you need to do if you want to use large foundation model technology in a highly federated private sense. This is something that the mainstream is not interested about because they're doing it on the freely available text data. But we are basically now developing a lot of the underpinning foundational technology for privacy insecurity to be able to train this model in secure data environments within healthcare settings where we can absorb this data but not take it with us, but just learn the knowledge about that. What the Nike becomes is a platform, right? And we're developing this as a non commercial platform. So it's going to be run and developed by academics, universities, institutions that are not for profit in that sense. And this gives us a different access and a different trustworthiness with the patient population. We've done surveys with patients with the thinking about that. And this is something that they find actually an acceptable mode of rendering the data useful. But just like in other big projects, Linux as an example here you can have corporates that support or sponsor work in the project without owning or controlling the project. And that's how we're seeing Nightingale AI to develop it becomes a platform. And then of course once you have a platform, you don't need to access the medical data behind that, you don't need to access the patient information, you just need to access the knowledge that is condensed in Nightingale. And so then as a company, as a corporate, as another researcher, anyone can just build on top of the platform and something on effective for them, I. E You know, build their own product using our platform below.

Okay, this is very cool. Also sounds incredibly ambitious, and I'd like to get into some of the technical details shortly, but it sounds like the big goal of this is to be able to answer questions in the way that a doctor would answer them. Can you give me an example of, like, what are the sort of questions you're going to answer? And like, how is this going to have an impact on public health?

Okay, so I think it starts with a simple question you could answer. For example, I don't know, you can upload your X ray and you can upload your blood work and ask the system what you predict my next X ray is going to look like. Say you have a chest infection and I'm prescribing you with type 1 of penicillin or type 2 of some other antibiotic. How will this evolve? And the system will predict how your X ray is going to look like. It's going to predict how your next blood work may look like, how you may feel in terms of, you know, breathlessness or other things. It will then reason about that out of the biological understanding that it has. And it then can, you know, give you explanations in textual form and so forth. So that's one example of how, in a practical, so to speak, patient doctor type interaction, you may ask questions. It could be also used by medical doctors who are basically now more interested in understanding what are the side effects? Say you have problem if you're in your liver and I'll give you a drug to cure something that you have in your kidney. What are the side effects, given that you have this liver effect? And how do all these drugs interact with each other positively, negatively? We call it the polypharmacy problem and reason about that. You could also be, for example, a pharma researcher and wanting to know, okay, where's the best population to run my drug trial for this specific drug so that I have a good chance at seeing the best results possible coming out of this drug trial. And then the system would reason based on that, where and what and how to recruit and select patients into the best trial. So this is an area of active AI research in pharmacy. How do you recruit the ideal patients? And this system should be able to do that internally. And then public health really is, when we're talking about healthcare, a setting where, so to speak, the government or the state or, you know, your city tries to manage health. Just think about the pandemic, you know, how, how that was managed. But you can also simply ask them you know, you want to do something about obesity and the challenges to obesity. You want to limit the spread of not just infectious diseases, but, but other bad habits that, that are unhealthy for you. So here you can now start using the AI system to, to make it reason about things and crucially, what the system does, because it reasons about, effectively how a patient is and how the patient evolves. The system has, so to speak, a digital twin of you, a healthcare digital twin of you, a description of how your body is or your physiology is doing. And so you can do that for literally every patient. And if you can do that, then you can run basically forward predictions and see, okay, how do different people evolve? So, for example, if you're an asthmatic patient and you know that's going to be very hot and there's going to be high levels of heat and dry air, then maybe you should not go out on the streets where there's a certain amount of pollution there and the system can predict you and can tell you don't do that. Every year, thousands of elderly die in heat waves or they die because they are asthmatic and get breathing problems, for example. So these are ways by which you can use these technologies to do what we've done before, or what we tried to do before, conventionally to do it with AI. But what we are really interested in is thinking about how can we use AI to do things that we couldn't imagine doing before. And that's where it becomes interesting. That's very cool.

It seems like there are so many use cases for this. Once you've got this model in place, actually, do you envisage it mostly being used by doctors or public health officials, or is it going to be something that could be used by anyone to ask questions about their own health situation?

Yeah, that's great. I think, I think the challenge is at the start it's going to be a research tool. So it's going to be used by AI researchers, digital health researchers and clinical researchers that will interact with it. Also, because the user, our focus is not on the user experience or user interface, it's getting the intelligence in. I think as the system evolves, you're going to get a proper chat and a proper prompt and then you can start conversing with the system. Now, again, I would say the first steps would be to have this for professionals, because the system, you know, doing a proper alignment on the system, that it doesn't say wrong or crazy things that have nothing to do with medicine but simply are in the language Model is a non trivial problem. You want to have some control over what the system speaks in general if you lay it out to the general public. Right. Because if it's seen as an authority, you need to be very mindful about what type of authority. So I think, you know, at the start we can show that it works and transforms things differently from what you would do with conventional AI. I think it's going to be for scientists and researchers, then it's going to be for trained professionals. And then hopefully by then we will have a scale and interest in the project that will allow us to then also develop a proper interface for that. But maybe that's a product that somebody wants to develop on top of our system.

Okay, so for professionals, first of all. And once you've sort of ironed out the kinks, that's when you can sort of gradually roll it out to everyone else. I'd love to know a bit more about this healthcare foundation model. So can you stop me through how does this differ from a large language model?

So a large language model is typically trained on sequences of text. So the quick brown fox dumped over the lazy dog, right, that's these words. And what a language one learns is then quick brown fox. And then it tries to predict, to speak the next word, jumped or ran or whatever other word there is in the English language. And just by reading a lot of text, this model can then very accurately predict what the next word is and can fill it in. And so this next word prediction is what basically drives this technology. And the amazing thing about English language or language in general is that that is enough for the system to complete sentences, actually write answers. So you start with a question and the system starts writing the text that's part of the answer. And so we do this through what we're calling transformer models that are models that can capture very long sequences and reason about them in an efficient way. So that's a language model. Now, a healthcare model is based not on a description of text or language in general, it's based, for example, on your physiological information. What's your blood pressure, what's your heart rate in this second minute hour, how much oxygen is in your blood. And so for, and all this information is collected, say when you're in the hospital. We know this from the movies when people are plugged in and there's this beep, beep. All this is data that's flowing out of you, that's describing how you're doing. But it could be also other information. Say you went to your General practitioner, and he took your weight and he did a blood test and measured your blood glucose level to see whether diabetic and so forth. This is all information that describes you, but it describes your different moments in time. And so now, instead of basically reasoning over a poem or an essay that you find on the Internet, you have now information about you that appears in different moments in time. So it's again a sequence of information, but it's clumped together. Say we're in your hospital and you have an X ray and you had other stuff, all this information clumps together and then maybe, you know, takes you six months till you see a doctor again and he will take your weight and blood pressure and add a bit more information. It's different in the format and that's not a continuous stream of data. And so we need to build and think around the foundation model aspects, the, the core technology, do the computer science around that to be able to deal with this vast mismatches in the timescale of information. You know, some information is updated every second, say your heart rate, and some information is updated every few months because you've done a blood test every few months. And so this we need to bring together and what you then get is a model that can then not predict the next word, but can then can predict your next medical information. So I can say, how's your next X ray going to look like, how's your next blood pressure going to look like? And so forth. And the critical thing is this thing can also reason if I do something to you, because that information is also in the record. So if I give you a drug, how does that drug affect you? I can see that from other people. I can then predict how that drug will affect you and so how things should improve or not improve and so forth.

My sort of understanding of large language models, and apologies to any AI engineers listening, is that basically you're sort of scraping the whole Internet, cleaning up the text a bit, throwing it into a big neural network and then magic happens. Maybe you need to tune it a bit at the end. So in this case you've got very heterogeneous data types. I mean, you mentioned the idea of like X rays, of like numeric data about people like their body weights and whatever. You've got time series data, it's a lot more complex. So is it the, the same kind of structure? Can you just kind of bundle it all into a big neural network and magic happens, or is there something more complex now?

I think you've got exactly Right. What the challenges are. And that's actually where we're now going from something useful to sort of basic questions about how do we develop the machine learning that enables that? And these are open, active questions. How do we deal with massively multimodal data, say, images, text, genetics and blood work? How do you combine them? And the challenges here are that in naive approaches, you build a model for each one modality and at the end each one gives you an answer and then you just combine the answers. We call that late fusion, fusion for bringing together and latex. You're doing it, the end. And that's not very smart because of course, say a doctor will look at your X ray and hear you coughing and then deduce that you may have a chest infection in that moment, not six months after he's treated you with antibiotic. And so we need to build AI systems that can ingest this information at the same time. And we're looking here at a number of architectures, so literally how to design the neural network communication when what is combined so that the system can reason across these modalities. And there are a number of competing approaches. It is not entirely clear to us what the best way forward is, because people have worked on vision to text transformers, for example. So you show a video and it tells you what happens, or you write text and it generates your video. We know how to do that. But what happens when this information is not in one chunk, but changes in a year's time? How do we deal with the sparsity of the data? So these are the two big science questions that were basically computer science questions that we're answering by developing Nightingale.

Okay. Yeah. So it sounds like there's still very much like a lot of open research questions in order to get a model that performs well. Sounds like a lot of it is about just capturing the knowledge there. Can you talk about. I know you've got techniques like retrieval, augmented generation, very popular for just retrieving facts. Is that a component within your model or.

Yeah. So here's the funny thing, right? When you're an AI researcher, you are at the moment completely data driven. So ignore all of human history and you just take the data and learn it from scratch. But of course, we now have. Medical research is a bit older than computer science and there's knowledge that we have accumulated in books and articles and papers. And so the idea is that we're going to use a standard language model to read all scientific literature ever published that we can get hold of and basically learn something that we call a knowledge graph. So some representation that links together how different entities stay in organ as a failure because some cells are not functioning, because some chemical process in the cell is not working, because some environmental agent has anticipated that and so forth. So you can build an entire ontologies like a graph, a network of how everything is linked up with everything by reading the literature. And so you can then use that with rag. So the, to combine the two, what the transformer of health says, together with the knowledge, information from human knowledge. But the trick is, of course, human knowledge in academic papers, I'm an academic, are, are not free from contradiction. You may find two papers that say two opposite things. So therefore I'm very careful about calling this a graph because there may be two links going in different directions, linking two entities together. Is the gene responsible for this disease or is the gene not responsible for that disease? So we need to think about how to model these relationships. That human knowledge is not precise, it's not logical at every single step. But it's many, many millions of people coming together, millions of studies being brought together to define knowledge. And you know, in cancer, for example, the areas of cancer research, they're publishing 1,000 research papers a day. A week. Sorry, a week. But it's accelerating, so it may soon be a day. No single cancer, cancer doctor can keep on track on that. Right. Just reading all this information. And so the idea is that with this knowledge graph we can extract information. And the idea in science is, of course, if something is systematically reproducible, you will see it come out over and over again. And with that we can get some certainty in this knowledge ground.

Okay, so yeah, certainly having access to the entire sort of corpus of knowledge. But yeah, no one can read hundreds of papers a day. So that does seem very useful. Okay, so from there, how do you deal with these contradictions? And have you made any progress in working out, like, what happens when we don't really know the answer? Like, it seems like a lot of the cases you have these generative models that just will tell you, oh, I believe this. And they're essentially hallucinating. So what happens if you don't know?

So that's a great question. So I mean, to be honest, we started training Nightingale on the national AI factory of the United Kingdom, that's called Isambard AI eight weeks ago. So we have already a basic model that is multimodal. It combines different forms of data in different timescales already, and it shows that it can predict outcomes for certain patients much, much more accurately than Reference language model derived reasoning or other forms of, you know, conventional machine learning based predictions, for example, about the mortality of a patient. That's the good news. I think the challenge that we're seeing is once you're starting to do the rag, you really need to be very systematic about that. So what we're looking here now is to find suitable language model to read all medical papers. So we tried that with a simple model that we can literally run on our local computer cluster and it's not very good at extracting medical relationships out of standard processing. Now if you use a very sophisticated model, it can do that, but it becomes very expensive and we're just an academic endeavor. So we're looking here actively for, for support and sponsorship. So again, this is just providing the compute so we can run a large language model that's sophisticated enough to reason about medical papers and extract the knowledge. The first steps were okay, but it's clear that it's not so easy. And that's ultimately because scientific papers are written a specific way and you need to know how to read them. And this is something that really when you do a PhD, you learn over a number of years how to, how to pick the tricky bits out that they were carefully left out or basically bits hidden behind the curtain. And that is not easy to convey to a language model per se. It's not a simple problem. So we're actively working on that.

Okay. Yeah, certainly a lot of academic papers, they're not easy to read for humans, like even the target audience. So yeah, I can see how it's going to be difficult to pull out the right information for an AI, particularly at scale at least. A larger question though is that if you make predictions about people's like the safe of their health, if you get it wrong, that's got some pretty serious consequences. So have you thought about how you're going to handle what happens when the AI makes mistakes?

So that's a huge problem. Right. And I think what we need to be fair here is to ask what is the human chance of mistake if you look at a whole range of diagnostic decisions? So do you have the disease, yes or no? It's a fairly easy thing to say. You can look at a whole range of fields in medical diagnostics. So for example, for sepsis, which is infectious disease, the biggest killer in hospitals, actually the doctor's decisions is two out of three. Right. So every third patient is, they get wrong on, on, on first attempt, so to speak. For dementia, it sits around nine out of 10. So one in 10 you get wrong with the diagnosis. We have AI system that reach or beat this value in dermatology. Diagnosis of a skin cancers by human doctors around 98.5%. And the best AI system now achieves around 99.5% accuracy. So machine systems can be better, but they cannot be perfect. And I think the realization is, if you're talking say about self driving cars, we're expecting self driving cars to drive thousand times more miles before having an accident than a human driver would. But human drivers don't have an accident every couple of miles. They drive a lifetime and maybe have one or two accidents. So that's a very high bar to breach. Now in medical decision making if you're getting 2 out of 3 right and 1 out of 3 wrong already. Going from 1 out of 3 wrong to 1 out of 5 wrong or 1 out of 8 wrong would be huge impact on patients. So therefore if the eye system can be better and we need to very systematically evaluate them, they can be huge bonus in picking up things. And it's not just that, you know, doctors make mistakes, human makes mistakes, but it also that as humans we're focused on specific things. So say you're going in for did I have a broken rib? You have an X ray and they look at the rib and say, no, the rib is fine. But actually if they would have paid attention to the shadow in the top right of your lung, they may have spotted a tumor. And that's a personal case that I know about, where they didn't see because they didn't pay attention to it. They looked at the bones, not at the shadows. And then much later the person was diagnosed with lung cancer. And, and the AI, this was a colleague who had access to the system, put it in and says, the AI said yes, there's a tumor there, but the human doctor didn't see it because they were focused on something, things, other things. So it's here where the AI system can really boost the accuracy of the human decision making. And so AI can get things wrong and it can get very dangerous if it gets it wrong, especially if it's built badly, if it's trained on bad data, if it's trained on too little data, and if it's trained on biased data. And so we need to be very careful when we're designing these systems. Fortunately, we know that these problems exist, but it's not a new discovery. And we need to make that extra effort to develop what we call patient ready AI, AI that you can really let out of the corral. And work on patients. And that also requires that you understand how you work with, how doctors work with AI.

Okay, so is the intention then that the AI is going to provide information to the doctor and then the doctor makes the decision based on that, rather than just AI says this and kind of over like you don't need to bother with the doctor in the first place.

So I think it depends. For Nightingale, definitely we will start with it's a decision support system, it's information system, it's a, it's a what if, what happens if I do this or this with the patient? How will this affect them? But we have already the world's first fully autonomous medical device in skin cancer detection that can fully autonomously make a skin cancer diagnosis. And effectively you can refer people without them having to see another patient. That's the 99.5% accuracy that I mentioned. This has been already accepted by the regulators. Interestingly, in Europe, the US has not yet allowed such a system to be licensed, but they're coming soon. So in very specialized applications where people have a very narrow focus and therefore can watch for the quality in a fairly easy way, I think we will see more and more of these autonomous decision making steps because they're just faster and often cheaper and maybe you can do them even at home. And. And so we are going to see more of that.

Okay, that's fascinating. And it's very cool that you've got a completely autonomous system just for the skin cancer case. I'm curious. It seems like the performance from one disease to another is very different. So do you need different AI systems for individual health conditions or is it better to have a broad platform like it sounds? Nightingale's pretty broad in this case.

Well, so that was part of the reason behind Nightingale. You can spend a whole career. And I've been doing this medical AI for now. Yeah, well, 12 years now. 15 years. So over a decade. And, and that's what you effectively do. You build a specific system for each disease and you know, you get faster or you get better. But ultimately there seems to be some common patterns. And of course, two diseases that are not related to each other are related to each other. I give you an example. So stroke, that's a cardiovascular disease because blood's not coming to parts of your brain, but it's also neurological disease because your brain is damaged. And you can look at it from the perspective of a neurologist or from a perspective of a cardiologist, but actually these are just silos of human thinking where you're schooled in a specific way of thinking about a problem. And once you bring an AI system that's not encumbered by these human limitations, it can just look at the patient as a whole. So it's really esoteric in some ways. We can talk about holistic healthcare by using AI systems that can think across boundaries and we're just talking fields of medicine. We can think about, you know, environmental impact on people, social impact on people, and so, of course, about how we can really think about treating and helping people, but most importantly, preventing that people get sick in the first place.

Absolutely. I mean, it certainly has pretty incredible potential. And it sounds like in this case, having something broad, it means, at least in theory, the whole is sort of greater than the sum of the parts then, because it can reason about multiple different things from different angles. I'd love to talk a bit about the data. So we've had a few guests on the show talking about trying to work with electronic healthcare records and talking about having to do, like, really incredible data engineering and using AI to sort of clean things up. But those, those are all the guests from working in on US healthcare records, where I think the situation a lot more fragmented. So you talked about the NHS patient data being the main source of data. Can you talk me through, like, how you go about collecting all that, how you go about processing it in order to make it ready for AI?

So the being a national healthcare system, it's as much a tool for medical benefit, it's also for performance optimization of the system. And so these systems are fairly up to date. So the way this information is collected, it's. It's literally the. The routine operational data that you have in your IT systems that gets stored and logged, basically messages that are being passed around or being stored in a database. What's then being done is that you typically anonymize this database. It's going on what we've done in Britain for a while now, in different parts of the country. So also in London, where my institution is built this, based it, we've built what we call secure data environments that follow very rigorous standards for data privacy, that take this patient data, be it from the hospitals, but also being it from the GP practices from the pharmacies, because everything is an integrated system and you can link this data up first and then you anonymize it and then you make it available in this research setting. And that's a very exciting thing because it means, I can see after, say, you had surgery what's happening with your disease, what's happening with you 10 years afterwards. Something that typical surgeon in a US hospital would not see because patient comes and they go and that's it, goodbye. And you don't have a follow up. But you can also look at what happened to the patient before they came to the hospital. So we just ran a big project on strained on 1.1 million patients and learned to predict from this data whether a patient needs to go to hospital unexpectedly in the next three months. And turns out we can do that with over 80% accuracy. And what, what surprised us even more is that we said we want to make the problem harder. So If I get 10, if I get your blood work every week, probably you can be quite accurate with what's happening with you. But that's very expensive. So what we said is we wanted to just look at administrative data. So just what happens if I have admin data about you, but not actual medical data that needs to be collected by a medical professional? Turns out with this medical data we achieve 81% accuracy of 4 out of 5 predictions correctly. And so that was surprising. So this just shows you, once you have this data what you can do. And so the challenges that we faced when I spoke to colleagues in the us this was the chief medical officer of a major US healthcare system is they say that, you know, we have patient, they're on the world's best cancer treatment, they get so sick they can't work anymore, they lose their health insurance and we don't know if they survived or not. So that's very tricky to think about building an AI for cancer treatment if you don't know whether the patients arrived or not. And so aside from the human drama that sits behind there, this is one of the advantages of this integrated care systems that we have in Europe generally, but also the UK specifically, that it gives us this cradle to grave data trajectory for patients. And so with that, you can do very systematic studies on people, not just for one specific thing.

Okay. It just seemed like a pretty incredible data set. And that's cool that you get 80% accuracy on predicting whether someone's going to unexpectedly go to hospital in the next three months. I feel like I would kind of like to know that about myself. Am I about to need a trip to the doctors? I think data privacy is going to be a big issue here. Certainly healthcare data is considered incredibly private. There are lots of laws around it. How does it work, making sure that if people start passing their healthcare data to the AI, it remains private, super important.

So I think there are two things That I want to say first, I've been working with patients for, well, almost two decades now. And in no case that a patient who was suffering from a disease that I could help, that I was working on, of helping with machine learning and AI, they were not concerned about the data, they wanted to give us the data. People who are very concerned about data are often healthy people. And so in a mean way I could say there's an egoism of the healthy not to want to give their data, but there's a huge generosity of the, of, of the sick because they want to get better. So we need to talk about that. But we also need to talk about how healthcare data can be abused, can be abused by insurers, for example, that they pump up your premiums or by employers that they may not employ you if, if you're likely to be a sicklish candidate, for example. And that's of course a no go. And so there different countries have different laws in protecting you, not just your privacy of your data, but how it can be used, say by insurers or employees, employers for your recruitment. When we're talking about the protection of information, we need to see how we can build systems. So the typical model, now a lot of the European data sits around is that it sits in what we call a secure data environment. So this is literally a remote desktop environment where you can remotely log in, you can work remotely, you cannot download anything, you can run your code there and you can see on a screen with limited resolution plot results. And if you want to, then to get the actual plot back to you, it needs to be exported in this human reviewers of the data will look at that and say, okay, you can take this out, you cannot take this out because you can re identify patients, for example, from it and so forth. So that's a typical mode of work. Now we have technologies, privacy preserving AI technologies, federated machine learning. So that's a bit like a model going on an apprenticeship journey. You send them to a secure data environment, it trains locally, it learns stuff and then it leaves that, it comes back and then you send it to somewhere else and it learns stuff. So all that you're integrating is effectively the knowledge accumulated in the parameters of the neural network. Now of course you can set up a neural network to effectively just memorize the data and regurgitate again that. And so there are a number of tests you can do, data leakage tests that you can run, so to speak, from a computer science side to evaluate how much data leakage can exist there. And a Fantastic. Colleague of mine just ran a study to look for what books have been read by certain language models, and he used that leakage analogous to identify which books should and should have not been read by the AI based on copyright. So you can do these assessments now. It depends very much on the policy and the governance of the secure data environment, whether these algorithmic satisfactions are enough. And quite a few of them are not satisfied with that. So they will not allow parameters of a neural network work to be exploited. And that's a challenge. So one thing that you can do is you build a model that is a bit superficial, right? You. That doesn't learn a lot about the diseases, but just a bit. And you build loads of these models and everyone learns superficially about something without giving out too much, without learning too much about a patient, so to speak. And the idea is then that you have to speak, you know, wisdom of the crowds. Then you bring these many thousand little models together, and the hope is that that is sufficient to build a smarter model. So that's one strategy by which you can overcome this federation challenge, that you need to export the parameters of the neural networks. Another approach, and that's what we're pursuing for Nightingale, is what we call building a meta secure data environment. So a secure data environment of secure data environments. So that basically allows these things to connect and communicate without human review, because you have some mathematical guarantees on what information can be exported and not. And so that's another approach. But it's not easy. And it's okay that it's not easy, but I think the important thing is if you care about it, you can find a way to make it work.

Okay, that's good. Yeah. Certainly you touched on a lot of the ethical issues there. Like, I'm kind of happy for my medical data to be used for research, but I don't want anyone to be able to like, type in to the AI, oh, what's Richie's health situation? And have it spit out. Like all my past history of things, and certainly I wouldn't want insurers or employees or just strangers being able to know things about my health situation. But that idea of grouping lots of models together, having your wisdom of crowd situation, it sounds a little bit like the idea behind random forests, where you've got lots of weaker models joining together into something stronger.

Yeah, it's. It's swarm intelligence, basically. Yeah.

Okay, cool. All right. So I know it's kind of early stages for your project, like you just training the model at the moment, there's still A lot of research questions. Have you had any success stories yet? Like what, what are the big wins?

So I think the big wins is that we've, we basically again we've been literally working at it for three months since we launched it and, and, and we've been training it for two months now. But I think the real success stories are that we've inspired others to contribute. So we have, I think it's been already officially announced so the Children's Hospital of Orange county and Grady Health so that Southern Californian healthcare providers have provided us with data to support our AI activities at Imperial and they want to actively contribute their insight in medical data to build Nightingale. And so as we're going along we are building more and more recruitment, we're ramping up the funding. Again we're not corporates and we don't want to be for this purpose of getting the patient buy in. So we've joined a 29 million fund research grant from the European Union that allows us now to work together with partners in Switzerland, in Spain and Germany and France towards multimodal AI. And so Nightingale is part of that. We have launched in the UK something that we call the Generative AI hub. So that's a, it's a multi institution center for working on generative AI. And as part of that hub we have now integrated Nightingale and activities of there. So it's basically we're in the ramp up and build up phase. I think results are going to come. I said first results look good. It's not published so I'm keeping a bit behind there. But we're getting the results that we don't see with the benchmark models because the reasoning is simply better. And that's very, I think comforting. Yeah, but I think the key thing is that we're opening up avenues of ingesting more data and more people want to be part of this endeavor to support what we can do for people.

Absolutely. That's very cool that even like three months into the project you're seeing some results and sort of beating some of these benchmark models actually. So since you mentioned money, like I know a lot of the big AI foundation model companies, they're building like multi billion dollar data centers and they're incredibly well funded. So how do you go about competing with that for tens of millions of pounds?

So we can't. Right. But so let's put it that way. I don't think these companies will get access to the data in the same way we can get access to the data. So that's our advantage. I think we have also slightly different motivation behind that. I think the key thing is that we're going to see a whole range of academic research AI factories coming up. Isembard AI in Bristol is the one big one. The half billion investment, close to 6,000 GPUs being already deployed there. There's new investments announced in the UK of building larger AI factories in the eu. So Europe, Germany, France and so forth. We're seeing a whole range of AI factories coming online specifically to supporting academic research. And I have to say as excited I am about technology and have also spin outs in that sense. So I'm very much for corporate activity in that space. I think there's this neutral ground advantage of doing things in academia and with academia that especially on very sensitive data gives you a very different buy in with patients and doctors, for example.

Absolutely. I think just the huge computer resource requirements. There's been a big shift from research happening in universities to research happening in companies. Like it's been a tremendous change over the last decade or so. So it is nice to see that there are some still like big academic projects going on in AI. Just what are the timelines for things being released, things happening? When can we see nightingaleye going live?

So I think Nightingale AI so we want definitely a first proper paper with model comparison and so forth coming out in the next 12 months. I think we want to have also version that expert users can use in around 18 months. Probably there's going to be intermediate versions before then that we're making available slowly to the community. Again because it's an AI system and you can use and up use it. We want to release it carefully so you will not be able to download it as such in that way. But we want to make it accessible to clinical users. It's a bit with when you develop a new drug, you can't just buy the drug over the counter, but you need to make it available to the right people so they can evaluate it and play with it and improve it of course, or build something on top.

Okay, so 18 months for the sort of initial release. Looking forward to that. All right. So finally just to wrap up, I always want recommendations for people to follow, new work to look into. So whose work are you most excited about at the moment? Who should I be looking at?

I think there's a whole range of amazing people and if I name check one then the others will be unhappy that I don't name check them. I think there's been a number of very interesting developments. One, there's been a fantastic paper by Apple AI. So Apple publishes as well. It's not all secretive and they had a wonderful paper about the limiting reasoning capabilities of language models. Another great place to look at for stuff in that space is the AI Safety Institute that was set up by the UK government and where we have fantastic colleagues looking at exactly the limitations of AI systems in that space. I think where the exciting things are happening is around multimodality. And here I would just recommend Googling on Arxiv, where all the newspapers end up, what's happening there. But I think what I'm really excited about is efforts about thinking about how to assess AI. And I think the trend is going there that we're thinking about AI system more being assessed like humans, not like, you know, a screwdriver for safety and operationality. And so there's a whole slew of work coming out in that space because ultimately we want to build AI that helps people, not replaces them. And so we need to think about how they can safely work with us.

Absolutely. I mean, all very important topics there, AI safety, certainly multimodal AI. Especially important if you want to look at X rays and other. Other medical data. Yeah, a text description probably not so good. And yeah, I certainly agree, like, the interaction between humans and AI is going to be just like a hot topic for the foreseeable future, I think. All right, wonderful. Thank you so much for your time, Aldo.

Richly. All the best. Thank you very much, Sam.