'We have the most to benefit, but also the most to lose': how AI could transform human health

Victor Pineda, Executive Director, Executive DirectorCenter for Independent Living: How do we ensure that AI is not leaving people behind, but rather is empowering people regardless if they have difficulty walking or seeing or hearing or remembering. People with disabilities have the most to benefit, but also the most to lose.

Robin Pomeroy, host, Radio Davos: Welcome to Radio Davos, the podcast from the World Economic Forum that looks at the biggest challenges and how we might solve them. This week: will artificial intelligence deliver a brave new world of healthcare, or could it marginalise ill and disabled people even further? This disability rights campaigner is both excited:

Victor Pineda: I remember seeing the deployments and thinking I've just witnessed my generation's moon landing.

Robin Pomeroy: And wary:

Victor Pineda: My values and the voices of 1.2 billion people in the world that live with a disability absolutely have to be part of this.

Robin Pomeroy: The promise for AI discovering new treatments is huge.

Daphne Koller, Founder and CEO, insitro: The ability to let machines discover things that people have not discovered. It's those discoveries that give us potentially the path towards new therapeutic interventions.

Robin Pomeroy: AIs is already helping speed up diagnoses of things like stroke where speed is vital.

Chris Mansi, CEO, Viz.ai: Taking that process from three, 3 to 5 hours down to 3 to 5 minutes.That patient can walk out of hospital, but time really matters.

Robin Pomeroy: Subscribe to Radio Davos wherever you get your podcasts, or visit wef.ch/podcasts where you will also find our sister programmes, Meet the Leader and Agenda Dialogues.

I’m Robin Pomeroy at the World Economic Forum, and with this look at AI, medicine and disability...

Daphne Koller: I think this is a tremendous opportunity for humankind.

Robin Pomeroy: This is Radio Davos.

Welcome to Radio Davos and to this special episode timed to coincide with the World Economic Forum’s Global Technology Retreat 2024 - we are looking at one of the most important areas where artificial intelligence is often said to present the most promise - health.

Later in the show, we will hear from the daughter of the late actor Christopher Reeve - the star of superman who was paralysed from the neck down in a horseriding accident. She helps continue the work he started with a foundation dedicated to seeking a cure for devastating spinal injuries.

We’ll also speak to people using AI to seek new medicines, and to diagnose strokes in a way that promises to save many lives.

But first, Victor Pineda, a campaigner for the rights of disabled people, he is president and founder of the Victor Pineda Foundation/World ENABLED.

I met Victor several months ago at a World Economic Forum AI conference. He is in a wheelchair and, as you will hear, is assisted in his breathing by a machine. I started by asking him how he saw the potential benefits and threats from artificial intelligence from his unique point of view.

And he started to answer that question by first giving a description of what he looks like and who he is.

Victor Pineda: Well, for those that can't see me right now, I'm a middle aged man, hispanic man, in a wheelchair and using a machine to help me breathe. And I have dark hair and I wear a scarf. I provide that visual introduction so that you get a sense of who's speaking today.

You know, I grew up in a world that was not designed for me. And with the fact that my body had increasingly developed weakness in my muscles, I kept finding barriers everywhere. Education system denied me education in my home country, Venezuela. But we moved to the US and there are rights and regulations are protected by right to an education.

I think now that we think about AI, we have a chance to make customised learning opportunities for a wide range of learners. We have a customised way to develop interventions in public policy or even in the way that we allow employees to show up in the workplace, with a variety of tools that allow them to be more productive.

So I think my contribution to the discussion is how do we ensure that AI is not leaving people behind, but rather is empowering people regardless if they have difficulty walking or seek our hearing or remembering. So looking at a use case of a user that might have functional limitations creates better innovations and better products for everyone.

Your question, though, is particularly important if we think about the both the benefits of the risks. People with disabilities, for example have the most to benefit, but also the most to lose. The most to benefit if their needs and their preferences and the ways that we interact with the technology, the cost and affordability, the availability of broadband. In terms of accelerating human potential human development, it is sort of the inherent human capacity to be a resource to your society. However, those same systems if they don't protect privacy. If they don't create the interfaces that allow a variety of users to interact with technology based on their preferences, if they don't create mechanisms for affordability or availability on an equitable approach that those communities that are already left out will continue to face a massive and exponential risk of being more and more marginalised.

So I'm here to ensure that we are intentional and conscious about the capacity to navigate away from the harms and towards the benefits of this technology.

Robin Pomeroy: For those of us who have just come across generative AI, I think it's become apparent that it's guilty of some of the biases that are already present in society in terms of racism, comes up. If you ask it to draw a picture of a doctor and a patient, sexism, the doctor will often be a man. And I'm sure there are, there must be countless examples of where that would come to people with disabilities. Does that continue to be a problem? Is it being addressed by the companies that are putting this stuff together?

Victor Pineda: Well, we have an open dialogue with various leading companies in the field. I think what's important is that Sam Altman from OpenAI is being very proactive in saying we need a variety of partnerships for these larger datasets from underrepresented groups in order to actually, you know, create models that are responsive and representative.

I think there's another issue, which is for these voices that are often drowned out, if you have a simple AI algorithm that does not really elevate the equity dimensions, those data sets are normalised, meaning they're averaged out and they're erased, because in a sense the dominance of the more present, more mainstream concepts can block out the more marginalised voices.

However, with the right approaches and regulations and the right set of values, you can actually create models that attempt to elevate those voices, in a way that they're not just present, they're becoming more valued or weighted in these models.

So we're in the process of experimentation. I can tell you that as a professor in the area of disability studies and urban planning, I had been a very early user of ChatGPT. I placed the title of my upcoming book, and nowhere in that title did I use the word disability or disability rights. But when I asked it to provide an outline of what this book could cover, it very effectively inferred that the title of my book, which is Inclusion and Belonging in Cities of Tomorrow: Governance and Access by Design, had a lot to do about people with disabilities and older persons.

And so I think what's interesting is there are definitely areas where the models are wrong, but we're also seeing like some snippets of successes, because I was surprised that the language, the terminologies, the frameworks, around disability were not 1990s, 1980s, sort of paternalistic or charity-based approaches, but the model understood what a rights-based approach to disability rights can look like in a city.

So I want to basically say that if we are conscious about the biases, if we are proactive about weighing the data in a way that elevates the representation of these groups, we can create outcomes that more adequately reflect our society.

Robin Pomeroy: So Sam Altman, who you have spoken with, is making some progress on that. But what about beyond that? We can't, I guess, rely on the goodwill of companies to do the right thing. A lot of the things that is being talked about here, about regulation or governance - what do you think needs to happen there, across industries, across countries, to instil those values and the importance of what you are talking about?

Victor Pineda: Well, for governance to work, it has to be based on a common set of values. And if we're going to continue to invest our resources in a way that protects our values - equity, fairness, transparency, representation of various points of view, a democratic decision-making - then we need to look at governance and structures that elevate those concepts in the design of this governance model. So let's get the values right, let's define the values.

But then I think what I can contribute from the perspective of disability rights or the movement of people with disabilities is a very simple phrase: Nothing about us without us, which is a concept developed to develop policies and plans not by able bodied people towards people with disabilities but actually promoting the participation of those groups.

Now, what's interesting is that that dialogue has now become much more universal. So now it's 'nothing without us', which is like every topic needs to have a way to have this larger consensus process.

Robin Pomeroy: You've worked for disability rights in politics, in big institutions and organisations. Is this kind of a new frontline for you now, AI? Because it's kind of where the future of humanity is going. Is that why it's so important for you to be here.

Victor Pineda: Absolutely. I mean, I remember seeing the deployments and thinking I've just witnessed my generation's Moon landing. The world will not be the same. We've crossed the threshold.

And so my values and the voices of 1.2 billion people in the world that live with a disability absolutely have to be part of this.

Robin Pomeroy: Victor Pineda, president and founder of the Victor Pineda Foundation/World ENABLED and Executive Director of the Center for Independent Living.

I met our next guest at the Annual Meeting in Davos this year. Alexandra Reeve Givens heads the Center for Democracy and Technology which is active in AI policy in Washington and Brussels. You can hear our conversation about that on a Radio Davos episode from February called: AI: Is 2024 the year that governance catches up with the tech?

Alexandra also happens to be the daughter of the late Hollywood star Christopher Reeve - for many of us, the original movie Superman. She helps continue his work campaigning for people with spinal cord injuries, and she told me more about that, and about her cautious optimism for AI.

Alexandra Reeve Givens: I'm Alexandra Givens and I lead the Center for Democracy and Technology.

Robin Pomeroy: You're involved in spinal cord injuries, the [Reeve] Foundation is. Do you see any, maybe this is something which we couldn't have done five years or 20 years ago?

Alexandra Reeve Givens: Absolutely. So my father was the actor Christopher Reeve, and I've been very involved since I was a child in the kind of latest medical research to help people living with not only the spinal cord injury, but all types of physical disabilities.

And when you look at the potential of AI and there's real opportunity there. We're seeing it already, for example, for voice technology, for people who are locked in or people who have ALS, real ways to suddenly communicate with the world, which is, of course, just completely transformational and breathtaking when you think about the power of that technology.

In the space of spinal cord injury, there are increasing medical breakthroughs around what stimulating the spinal cord can do to actually trigger movement that a person can control, again, all through highly sophisticated assistive technology.

So that type of potential really is remarkable. And we need to harness that innovation. Of course, at the same time as making sure that it works well for everybody, and that that doesn't just become a smokescreen through which other, you know, less helpful uses are kind of greenlit.

But when I think about, you know, the future, what the rest of the 21st century looks like, there's an enormous story for potential there as well.

Robin Pomeroy: I do remember Christopher Reeve, your father, talking with such optimism about technology, he would say, the doctors are telling me this can never happen, that I will never walk again. Obviously, tragically, in his case, that was true. But if this was happening now, maybe this would be the breakthrough that he really felt would happen.

Alexandra Reeve Givens: And his belief was to name it. Right. To be provocative, the same way he took a lot of inspiration from President Kennedy, saying, by the end of this decade, we're going to put a man on the moon. And when you challenge a field to say the status quo is unacceptable.

I mean, I have a lot of pride looking back at my dad's legacy and thinking that the field of spinal cord injury research really was often referred to as the graveyard of neuroscience. People were not focused on it because they didn't think there would be any progress to be made, and he challenged them to get off the sidelines and try.

And now we see so many more generations of scientists really focusing on this. We're seeing medical breakthroughs. The Reeve Foundation continues to this day, funding that work and seeing transformational changes every day.

So yeah, of course you can be optimistic. I mean, even I think about it, I remember when I was on a school trip, when I was probably 11 years old, getting a fax, so this is going to date it, from my dad. And it was the first time that he had dictated a letter himself through technology. So he'd use drag and dictate software. Some of you may remember it from the 90s, to write his own letter. And, you know, it's very sweetly it was full of typos and all of these things. But the simple act of feeling like he had agency and could send something directly without having to have another person help them, was transformational. And that was for the most rudimentary version of this technology.

So again, there's a ton of potential there. We just have to unlock it and do it in a constructive way.

Robin Pomeroy: Alexandra Reeve Givens, head of the Center for Democracy and Technology. To hear more from her about AI governance in general, please do check out that episode of Radio Davos from February called: AI: Is 2024 the year that governance catches up with the tech?

Our next guest is a brain surgeon turned AI entrepreneur. Chris Mansi saw the potential of using AI to help diagnose stroke - something he says will be life changing and life saving.

Chris Mansi: Yes, my name is Chris Mansi and I'm the CEO of Viz.ai.

Robin Pomeroy: Tell us, what is Viz.ai?

Chris Mansi: Viz.ai is a healthcare software company that leverages artificial intelligence to increase access to life saving treatments. Just to be clear, what that means is we use the AI to automatically detect specific diseases that are serious in nature, like a stroke or a brain aneurysm or cardiomyopathy. And use that automatic detection of the disease to help make the patient care much faster and more consistent. So we get that information to the specialist who can treat you. And we do it much faster. And that leads to significant improvement in patient outcomes.

Robin Pomeroy: Give us some idea of how it works, because some of those conditions are quite difficult to detect, stroke or heart attack. Doctors or paramedics often have a hard job, but it takes a long time sometimes. And time's very important. So how does AI help them do that?

Chris Mansi: The innovation in the development of treatments for disease, that's a medical device or typically a pharmaceutical, a drug, is accelerating. And, I'd go as far as to say in the next ten, maybe 30 years, we'll have treatments for almost every disease out there.

But the problem is the vast majority of patients don't get access to those treatments. So how AI can help is by automating the detection of that particular disease and making a guideline directed workflow consistent that works every time.

Put simply, the biggest problem in health care today is the variation of care. If you show up to a specialist hospital at two in the afternoon on a Tuesday, you're highly likely to be diagnosed to see a specialist and get the relevant next test or treatment that you need for your particular condition. But you know, medicine is a hyper speciality game, meaning that if you treat brain aneurysms, you're probably one of three people in the city of San Francisco that does that. And so a patient with a brain aneurysm who might show up to a different hospital needs to get to that specialist somehow.

AI works consistently, no matter the time of day, the day of week, or where that patient shows up to a hospital in terms of geography. And so it's really the tool that allows us to democratise the detection and flow of patients to the right place list democratising health care.

Robin Pomeroy: What does that look like if I have a patient showing up with those symptoms?

Chris Mansi: Yeah, if I give you an example of stroke. So, you know, we got the first FDA approval for AI and that was for our first indication, which was in an ischemic stroke.

Before our technology, a patient would show up to a hospital, they'd have weakness in their arm or leg or difficulty speaking, and the doctor would arrange a CT scan of the head. That CT scan would go into a radiologist's work queue. And once they got to it, they would read the scan. Sometimes it was a specialist type scan. So they need a neuro radiologist to take a look. That would also take time. Then they'd have to communicate the results to the emergency room physician, who would have to then get hold of the neurologist, a neurosurgeon, often in a different hospital, in order to make a diagnosis and treat that patient.

That process used to take between 3 and 5 hours. With AI, the patient still gets the scan, but within minutes, the neurologist, a neurosurgeon at a specialist hospital is alerted. They communicate with the emergency room physician and allow that patient to be transferred and treated much faster. Taking that process from three, 3 to 5 hours down to 3 to 5 minutes.

Robin Pomeroy: Is that so important?

Chris Mansi: Well, in the case of acute disease, you know, we have a saying, time is brain. In the case of stroke, every one minute almost 2 million brain cells die. And so if you imagine over 3 to 5 hours, you get a lot of brain cells that are dying, which leads to stroke being a devastating disease. It's the number one cause of long term disability in the developed world. If we can get patients to treatment where we remove that clot so we unblock that vessel, that patient can walk out of hospital. But time really matters.

Robin Pomeroy: How does an AI learn how to diagnose something like that?

Chris Mansi: So what you do is you take thousands of scans that show the particular disease and you get an expert, in this case it might be a neuro-radiologist to ground-truth those scans, saying this patient has a particular type of stroke and where that stroke is, and you use that as your positive set. And then you have a negative set where there's no stroke and you train that model again and again and again. And what essentially it is, is a combination of statistical models that each time it's being told what the right answer is, it's adjusting the statistical weight to start to learn, just like the neurons in the brain do, that this pattern suggests that this patient's having a stroke. And over time it gets to human expert level of performance, which is what you need in order to get your product FDA approved.

Robin Pomeroy: So tell us your story. How did you come into this?

Chris Mansi: So prior to this, I was a neurosurgeon in the UK. I came out to California. I spent some time doing graduate studies at Stanford, and I realised the technology which had only emerged in 2012, had huge potential in health care. Back then it was mostly being used in imaging to do things like tell the difference between, you know, cats and dogs or for facial recognition.

Robin Pomeroy: How did you come across it? Was it just being in Stanford and surrounded by...

Chris Mansi: I think it was being in Stanford and being surrounded by amazing computer scientists, statisticians, engineers. It was really just like a generative AI is today. It was the hot technology back in 2014, 2015.

But I had the experience, the firsthand experience of the problem that it could solve and the clinical perspective to realise that if we were able to make that detection of disease consistent, we could use it not just to improve disease detection but improve patient outcomes by organising better, faster workflow so the patient gets the place they need to every single time.

Robin Pomeroy: These advances have happened here in California. But we know in the U.S. and in many countries, health care provision is quite unequal, depending on where you live, how much money you have, and around the world the inequalities are enormous. Do you see AI improving that in any way or is it just going to get worse and worse?

Chris Mansi: I see AI as part of the solution.

So within a health system, and each country has their own different set up. You know, I'm from the UK where we have the National Health Service. It is very different system to the US. But each of those health systems are trying their best with the resources they have to serve their population.

And what AI can do is make it so there isn't a zip code or a postcode lottery. And wherever that patient shows up, they're going to get that level of expertise equivalent to what they would get in a specialist hospital.

And that's very important because, even across London, where I used to work, there were huge variations in outcomes. Right? Hospital that might be five miles down the road where those doctors have trained, you know, in great centres and, you know, in theory should all have the same knowledge. But what happens is the specialist centres for a particular type of cancer, they send those doctors to the specialist conferences and so they learn about new treatments, new diagnostic techniques. And so they are much further ahead than the hospital down, down the road.

And if you take it to a global scale, I think the you know, AI, generally technology, has the ability not just to detect more patients, get them to the right place, but also to help educate.

You know, one programme that we have Viz is we allow a a doctor, typically a specialist in the US hospital, to sponsor a hospital in the developing part of the world. And so they will be on the same system as that hospital, often in a country that they have some connection to, and they can provide that expertise and education to their hospital system.

Robin Pomeroy: Chris Mansi, the CEO of Viz.ai.

Our last guest in this episode is a computer scientist who is one of the founders of Coursera, the online education platform, but also the founder of a company that uses AI to find new medicines. This is Daphne Koller

Daphne Koller: I'm Daphne Koller, the founder and CEO of Insitro, which is a machine learning-enabled drug discovery and development company.

Robin Pomeroy: When people talk about the massive positive impacts AI could have, it's one of the, if not the first thing that people say: medicine. Why is that? Why should we be optimistic about advances in medicine from AI?

Daphne Koller: I think this is a tremendous opportunity for humankind.

I've been working in the field of AI and machine learning for biomedical data for well over 20 years now, and it is absolutely one of the most complex fields out there. And people just are not able to understand biology and human biology specifically. There's just too much complexity for the human mind to understand.

And so I think that we have a unique opportunity, a unique moment in time to measure biology in unprecedented scale and quality, using using new tools, new measurement approaches, new devices, feed that into an artificial intelligence and let it actually tell us stuff about human biology that we will never be able to figure out on our own.

And I think that has repercussions that range all the way from the most immediate delivery of care and what medicine do we give to which patient and how do we track whether they are or not responding to that, all the way to the very earliest of biological discoveries that might give us the next generation of medicines for these diseases that currently remain largely with enti rely unmet need, whether it's Alzheimer's disease or the people who still die from cancer, from metastatic cancers and so on.

Robin Pomeroy: We must have been using computers to work things out for a few generations. What's so different when it comes to medical research? What does machine learning give us?

Daphne Koller: So I think when we've been using machine learning for medicine, it's often been utilised as a way of substituting maybe with somewhat higher reliability, maybe with somewhat lower cost things that a clinician already does. So there's platforms or algorithms that will say, here's what a clinician does, I'm just going to do it faster and maybe with slightly less variability.

The thing that we're able to do today by collecting enough data about human biology all the way from the sub-cellular data of proteins and genes, all the way through to organismal changes, is the ability to let machines discover things that people have not discovered. And it's those discoveries that give us potentially the path towards new therapeutic interventions.

Robin Pomeroy: Do you see a way, when it comes to governance or regulation in the field in which you work, to ensure we get the right outcomes but avoid the potential risks. Have we managed to kind of conceptualise where governance needs to go in that field?

Daphne Koller: I think we haven't fully conceptualised that yet and I think there is clearly an opportunity to ensure, for example, that models that are used in this type of application are trained on high quality data, that there's potentially an appropriate way of assessing tools via randomised clinical trials, which is something that honestly AI driven tools have not been subjected to that level of appropriate scrutiny and and assessment.

And as a consequence we've had tools that have been deployed in health care that have been quite flawed, that have actually made outcomes worse, that have exacerbated, you know, health disparities versus helped address health disparities.

And all of that has been happening because there hasn't been a governance on those tools in ways that we're now starting to think about, maybe because the tools have become more powerful. But frankly, the simple tools can do as much harm and maybe even more harm than the than the more sophisticated ones.

I would, however, caution one thing, which is one of the risks in the deployment of a new technology is to hold it to a standard that is unreasonably high and throw away, you know, have the perfect be the enemy of the very good. Where I think the right bar to hold is not, is the solution, is a technological solution perfect? But is it better than what humans do today? And I don't think I need to tell you that doctors today make many diagnostic mistakes. There's a lot of biases in how they apply health care to people with different groups. And I think there's a lot of issues. And not to condone or excuse that, but I sometimes worry that by saying, well, the machine learning the AI has to be perfect, we're going to continue to perpetuate the status quo. And the status quo is deeply flawed.

And so the bar should be, are you doing better than the status quo? Not have you achieved perfection?

Robin Pomeroy: There seems to be a risk, doesn't there, that technologies, because of the training data that you mentioned, kind of amplify existing human and societal flaws and that those are huge in medicine around the world.

Daphne Koller: There are very clear risks in this area that would absolutely need to be assessed and and appropriately safeguarded or mitigated. That's where and an appropriate protocol process for assessing that technology is the right approach.

But like what I was trying to suggest was that the bar of when is something good enough to deploy, isn't that it makes no mistakes, but rather that it makes fewer mistakes or is more equitable than a human clinician is. And that is that should be the bar because people are very far from perfect.

And your average clinician today went to med school 15, 20, 25 years ago. Whatever they learned back then, they have largely forgotten or has become obsolete. It's impossible given the volume of publications that are happening today for an average clinician to keep and to keep on top of even a 100th of a percent of what's being published.

And so the bar of what current practice is really not as high as we needed to be as a society. And so, yes, we shouldn't just launch random tools into this ecosystem and and not have any oversight or quality checks, but having overly high quality checks that try to aspire to something that is effectively perfection and otherwise you keep people doing what they're doing today, what people are doing today is not great and there's a lot of people dying because of medical mistakes. There's a lot of people not getting access to clinical trials that could save their lives because their clinician doesn't even know that those trials exist. Many, many examples of places where there's significant human limitations that should be part of how we compare whether a AI system is actually better or not.

Robin Pomeroy: For people who don't follow this that closely, for a normal person, where will they see the impact of AI when it comes to medicine? First, it maybe is already happening and we're not seeing it. But when will I as a patient, think: ah, AI did that.

Daphne Koller: So I think that the most obvious and immediate use case will be when we go to our doctor and rather than the doctor relying entirely on their limited recollections of what what might be relevant to my own personal situation and in a very limited amount of time in the doctor's office, remember that and be able to provide a course of treatment, that instead the doctor will be able to look up on the whatever on the AI what is known about the condition, maybe pull up the most relevant case studies and such and be able to provide even within the limited time constraints of a doctor's visit, a much better, more tailored recommendation in the course of treatment. So that, I would say, is number one.

I would say a second one is as we have better and better ways of assessing our physiology in high content, ways of reading it out, whether it's MRI or blood or very extensive blood biomarkers or other ways of measuring biology, that we will move away from more traditional, one size fits all definitions of disease and into something where diseases are defined via the underlying biological mechanisms that drive them rather than by some symptomology that dates back a hundred years before we understood the underlying biology.

I think everyone now realises that Alzheimer's is not one disease, it is multiple diseases driven by quite diverse biologies. The same is true for type 2 diabetes, the same is true for most neuropsychiatric diseases, and yet we treat them as if they were a thing. And if we were able to more finely parse the subsets of patients, I think there would be a much better ability to target course of treatments individuals.

And then I would say the third one, which happens later, is that understanding of biology is going to give us a better ability to identify new intervention, new therapeutic treatments, and that usually is a longer process. It might take, you know, ten years to get there, but there will be better therapeutics because of our ability to better understand human biology.

Robin Pomeroy: Daphne Koller, the founder and CEO of Insitro.

To find out about the work of the World Economic Forum's AI Governance Alliance, please visit our webiste - links in the show notes to this episode.

And we have many episodes on AI on Radio Davos and on our sister programme Agenda Dialogues - find them all wherever you are listening to this, or at wef.ch/podcasts. If you have any comments on this or any of our podcasts, please join us on the World Economic Forum Podcast club on Facebook.

This episode of Radio Davos was written and presented by me, Robin Pomeroy. Studio production was by Taz Kelleher.

We will be back next week, but for now thanks to you for listening and goodbye.