Mihaela van der Schaar
Nick MaxfieldOn March 30, 2022, the van der Schaar lab held a clinician-oriented engagement session in which three of the world’s most prominent clinician voices—Eric Topol, Euan Ashley, and Geraint Rees—shared their own visions for the future of AI and machine learning in healthcare.
During her discussions with these three visionaries, Mihaela van der Schaar asked the following four questions:
1. What do you see as the biggest opportunities for ML to transform healthcare that are not being pursued today? Are there areas or aspects of healthcare that are especially suited to/in need of ML?
2. Where are you seeing ML making a difference today in healthcare?
3. What advice would you give ML for healthcare researchers like us to focus on, so as to make an impact on healthcare?
4. Do you have any advice on how our respective communities can interact and build successful teams across disciplines?
Since the answers provided were particularly inspiring and thought-provoking, we have provided them below. The full video of our March 30 Revolutionizing Healthcare session can be found at the bottom of this page.
What do you see as the biggest opportunities for ML to transform healthcare that are not being pursued today? Are there areas or aspects of healthcare that are especially suited to/in need of ML?
Prof. Euan Ashley, MD
Professor of Medicine and Genetics at Stanford University in California; Founding director of the Center for Inherited Cardiovascular Disease, Clinical Genomics Program; Co-director of Stanford Medicine Catalyst
I think the biggest challenge for machine learning and AI in healthcare is clearly implementation. We have a lot of amazing science happening. There are fantastic, effective, and accurate prediction models happening essentially all the time. Great publications every single day. But when I look around at my professional life in healthcare, and think about my patients or my own health, and then walk through an average week with a few different touchpoints in the healthcare system, I see AI and ML being used almost nowhere in the healthcare system itself.
I think the major barrier is to take this incredibly exciting science and turn it implementation science. I think it will take the science of implementation in order to make the execution effective, in order to move the wheels of the big machine that is healthcare—and the huge inertia behind that.
I think the opportunity is massive, and we all spend our lives thinking about the ways that this kind of technology and these approaches to quantitative data can really revolutionize care for our patients. I think those are all very valid and correct. But when I look even at our own work: how much have we translated? It’s pitifully few right now. We’re working to change that, but I think as a community that’s likely where we should focus. It’s very important that we get ahead of that and think about the science that needs to be done in that space as we start to really get better at implementation.
Prof. Geraint Rees, MD
UCL Vice-Provost (Research, Innovation & Global Engagement) and Pro-Provost (Academic Planning); Dean of the UCL Faculty of Life Sciences
The interesting part of that question, I think, is what’s “not being pursued.” I think actually quite a lot of opportunities are not being pursued at the moment. I’d pick three areas: drug discovery, time series diagnostics, and operations and logistics.
I picked drug discovery because at the moment AI is really being used in diagnostics. We think about how we can apply it to images in particular. AI is generally really good at images, and there have been major advances in recent years in the underlying algorithms and technologies. That’s fine, but I wanted to focus on therapeutics because therapies are obviously a huge part of what we think of as healthcare, beyond just diagnosis. There are some interesting opportunities in niche areas (for example, talking therapies in psychiatry), but drug discovery seems to be the space where the search space is really large. There’s been a huge AI advance in recent months in this kind of area.
Time series diagnostics are interesting because a lot of things in a healthcare system happen over time: people get worse, and they get better. The intervention needs to be timed for the injection in your eye to cure your eye disease. These are the kinds of things we have trouble spotting because they can get lost in a morass of data. But again, the search space for time series data is quite large, and the opportunities are abundant if we can use machine learning to get the right time to make the right intervention—for example, before a patient deteriorates or needs a more invasive type of treatment. Time series is something we find quite difficult as humans.
Finally, I chose operations and logistics. One thing that isn’t always apparently to those on the outside is that healthcare is a massive system of people and staff. There’s primary, secondary, and tertiary care, and these are all integrated—people and equipment are flowing. Hospitals have to cope with unknown influxes of patients, and (in the case of COVID) balance the demand on resources with treatment of other patients, in addition to the absence of unknown members of staff. All the equipment and staff have to be in the right place at the right time to do the right interventions for the right patients. This is done amazingly well, but it’s based on ancient technology. What would it look like if we could use all that data, all that flow, all that optimality and logistics, to transform healthcare?
You might think the way to do machine learning in healthcare is to pick the biggest disease areas (COVID at the moment, or cancer, or cardiovascular disease). That’s fine—these are the areas of greatest need—but it’s also important to think about the capabilities of machine learning and AI, and where they might have most application beyond the areas of greatest needs. In these three areas, I think the search space for the optimal solution is really large. There’s a lot of data, there’s a lot of difficulty in working out the optimal solution, and the opportunities if when you find the optimal solution are quite large.
We need to look for the high dimensionality in healthcare, but this is often hidden. We tend to try to reduce patients to a single disease or a single phenotype—or we think of all patients as being treated according to the mean of those patients. We look at the disease categories, rather than the variability within those categories. This drives us to look for optimal biomarkers or optimal treatments for specific diseases, rather than looking at the full dimensionality. If you treat patients as intrinsically high-dimensional, what you’re really doing is delivering the ultimate in medical care: personalized medicine. Delivering that is sort of the holy grail of what health professionals are trying to do.
We also need to think about what constitutes optimality. For example, in the case of COVID: is optimality the ability to divert all resources to the urgent needs of the people at the front door of A&E? Perhaps yes, but equally there’s an urgency for cancer operations. And we see this struggle in the operations of our healthcare systems today. What I’m looking for, from a leadership perspective, is not for an AI solution that tells you what to do. It’s for a solution that gives you some possible combinations of resource allocation that may be optimal—and then lets you use those to inform and assist your decision-making. That’s also key to my vision of AI, which is AI in a world inhabited by leaders and healthcare professionals who are all extremely good at the jobs they do—the question is how to augment that in situations that are very challenging.
Prof. Eric Topol, MD
Founder and Director of the Scripps Research Translational Institute,
Professor of Molecular Medicine at The Scripps Research Institute, and
Senior Consultant at the Division of Cardiovascular Diseases at Scripps Clinic
It’s worth bearing in mind that each specialty really has different challenges. For example, cardiologists are very much segmented into different areas. Some are the electricians, some are the plumbers, some are general cardiologists, heart failure cardiologists, and so forth. So there are different needs even within the cardiology specialty—for example, the heart failure cardiologist is especially interested in the physiology of the person, whereas the interventional cardiologist is interested in what is ischemia, the stress tests, the angiograms, and so forth. What they concentrate on in so different.
To help them, obviously there are some common threads. No clinician wants to work at a keyboard. They all want help with NLP and ML features that will liberate them. They also want a better handle on all the data of each patient. That hasn’t been done—we all struggle looking at screen after screen, trying to get our arms around a patient, to learn what we can in a matter of minutes, because we don’t have much time. We desperately need help.
The other thing clinicians really need help with is keeping up with the literature. We don’t have anything like a daily download of what’s important for us, within our fields, to know in our care of patients.
These are a few things that are separate but also unifying.
Where are you seeing ML making a difference today in healthcare?
Prof. Geraint Rees, MD
UCL Vice-Provost (Research, Innovation & Global Engagement) and Pro-Provost (Academic Planning); Dean of the UCL Faculty of Life Sciences
I’m going to be a little provocative and say I don’t see AI or machine learning making a difference today in healthcare.
The word I notice the most when reading the related literature is “will.” Everyone talks about how AI and machine learning “will” make a difference and transform healthcare. And I agree with that—I’m incredibly keen on the potential. But it’s all “we will,” not “we have.” It seems still to be all in the future.
That’s kind of interesting, because I’ve sat here for almost two years talking to wonderful people through my screen, and it’s still “will.” It of interesting during the COVID pandemic, because AI has kind of vanished. That’s a contentious statement, but where are the AI solutions that are helping clinicians to predict hospitalization and deterioration in patients? They haven’t really emerged.
COVID has affected basically every person on the planet. It’s the perfect use case for very large-scale data, where we need to be able to answer questions like “should I go into work today, or is the risk too high?” or “what’s happening with infections in my local neighborhood?” That’s not to say that tech hasn’t made a big difference—an amazing job has been done by public health authorities in delivering large-scale data systems. In the UK, I can go online onto the government website and look at exactly how many patients were admitted in the last few days in my local hospital. I can look at the infection rates and vaccination rates in my local area. All of that is available via an API. That’s an amazing public health achievement.
But as far as I know, none of that is using multivariate analytics or any form of predictive analytics to make policy decisions, which is why I’m saying ML isn’t making a difference in healthcare.
On a positive note, there are areas where it’s making a difference behind the scenes. One thing I noticed in the business world is Microsoft’s purchase of Nuance Communications, who are behind Siri—clearly machine learning is used in voice recognition. That’s something that’s being used in hospitals—certainly in the U.S., where I believe three quarters of hospitals are customers of that particular company. That’s clearly using analytics to do voice recognition and ensure you get an accurate note. That’s an example of “AI in the background” which is important for operations and logistics.
Even more positively, if you look at the FDA, they maintain a list with about 350 AI solutions now approved for use. Lots of those are in radiology or cardiology, in areas like CT angiography or 3D echocardiography. I do think we’ll start to see an increasing proportion of solutions in image-based diagnostics coming to market and being embedded in systems and rolled out throughout healthcare.
What advice would you give ML for healthcare researchers like us to focus on, so as to make an impact on healthcare?
Prof. Euan Ashley, MD
Professor of Medicine and Genetics at Stanford University in California; Founding director of the Center for Inherited Cardiovascular Disease, Clinical Genomics Program; Co-director of Stanford Medicine Catalyst
I think you’re an exemplar of what should be done: the answer is partnership. We have to find the people who are capable of both funding and clearing the path for implementation. Those are potentially two different groups of people.
I think the easier group are the doctors—the ones who are in the trenches. We’re able to talk to them and see where this technology could have its biggest impact. But the challenge is that those doctors are often not in a position to actually “do” the implementation, because there’s usually an administration for a hospital or community healthcare. Especially if you want to do something at scale—which is where this technology lives—it absolutely requires a third pillar, which is an administrator who understands the opportunity and can really get things done rapidly.
There’s a whole expertise there, beyond (for example) the ethics of doing a research study. Even if we have a prediction model that’s ready for prime time, what does it look like to take that model and actually put it into practice? What do we need to do? That book is still very much being written: there’s no doctor handbook on the shelf saying what you need to do.
In some ways, we have something to learn from the pharmaceutical industry. It’s an industry that exists around inventing new ideas (molecules, targets, etc.). It starts in the lab, with cellular experiments, animal experiments, first-in-human experiments, phases two and three, regulatory approval, and then deployment. In deployment, there’s also marketing and much more. And each one of those steps is actually going to be required to really attain the potential that this technology has for healthcare. There’s an equivalent of each of these steps that we need to think about as AI researchers when we think about putting digital therapeutics or diagnostics into practice.
If we try to skip any of those steps, we’ll find that we come up against a large force: the inertia of the healthcare system. Put differently: there are so many opportunities to improve the care of patients, whether on the pharmaceutical side, the digital side, or the AI side. There are so many great ideas that are making noise for the attention of decision-makers in healthcare. So which one do they choose? If they’re going to spend some money on something like a phase three trial, where do they spend? I think this is as important as anything.
When there’s so much innovation in healthcare, the key is being able to get the attention of the right person who can help us with the equivalent of the multi-phase trial process, plus regulatory approval. My advice would be to take lessons from pharma and try to find the right people (maybe even people within the pharma industry who understand that pathway) who can help you along. We can also emulate pharma in terms of their great relationships with regulatory bodies—when they come to a regulatory body with a new drug, they know what they need to do, who they need to get in front of, and the timeline. We have to do similarly with our digital technologies, and they will reach prime time because they’re effective and in many cases much cheaper than a pharmaceutical product to bring to market.
Prof. Eric Topol, MD
Founder and Director of the Scripps Research Translational Institute,
Professor of Molecular Medicine at The Scripps Research Institute, and
Senior Consultant at the Division of Cardiovascular Diseases at Scripps Clinic
We need to be imaginative—we don’t want to have any compartments. We have to realize that with the help of deep neural networks we’ll be able to do things we never even envisioned. That takes imagination.
For example, who would have expected that you could look in the eyegrounds and determine a person’s kidney disease or risk of Alzheimer’s? Or who would have expected that you could look at an echocardiogram and determine someone’s hemoglobin and anemia? Sometimes there seems to be too much of a reach—for example, people saying they can detect COVID from a forced cough—people are doing these kinds of things, which shows they have imagination. Even if they fail, at least they’re thinking big and considering ideas with a lot of potential impact.
I’d also like to see more emphasis on the patient side—patient empowerment. As clinicians and scientists, we always think about the scientists, the researchers, and the physicians. We don’t think enough about the patients, who are desperately seeking more autonomy. We can provide neural network support for them, to make diagnoses that are not life-threatening or serious (urinary tract infections, ear infections, skin rashes, etc.). The fact that we started with arrhythmias through a smart watch suggests that these things can get pretty sophisticated. Something I’m especially enamored with is smartphone ultrasound. The idea is that you could coach anyone—even my six-year-old grandson—to get an echocardiogram through a smartphone if they just listen to the AI and follow its instructions about how to move the probe. It automatically captures a really high-quality loop. That can be done (and will be done) for all parts of the body except the brain. In the future, this kind of “medical selfie” via smartphone could be part of a telemedicine visit.
We use the term “patient-centric” a lot, but we don’t give it the regard, respect, and priority it deserves. We keep coming up with tools that can help doctors. And of course we need help, too. But there are a lot more patients out there than doctors and nurses.
What isn’t clear to many people is that as clinicians, we see patients for an infinitesimally small part of their life—maybe a few minutes once a year. But we now have the ability to capture data continuously, passively, and at high frequency. That data can be highly useful. There’s the idea of having a virtual medical coach (for those who want it) where we use deep learning and hybrid models to help coach that person and prevent illness. This is something we can do as clinicians—we can’t be with patients and assimilate all that data. This is another dimension of where the field will go over time—dealing with all the data so we can actually prevent diseases. That’s the dream. But the point is that we have to understand the limits we have. We have an icepick view of any patient—very tiny in time and real data (vitals, lab tests and scans at that particular moment), but there’s a whole world out there.
The clinicians have this sense of ruling the roost and knowing everything—we have to be more humble and have humility, since we are only such a tiny part of a patient’s world. As a patient, I can see the other side all too well.
Do you have any advice on how our respective communities can interact and build successful teams across disciplines?
Prof. Euan Ashley, MD
Professor of Medicine and Genetics at Stanford University in California; Founding director of the Center for Inherited Cardiovascular Disease, Clinical Genomics Program; Co-director of Stanford Medicine Catalyst
It starts with a single person or relationship and a single conversation, and grows from there. A lot of it is about understanding the other’s position—especially if it’s going to be a long-term relationship.
We need three groups to be involved if we’re talking about implementation. We need the machine learning experts. We need the people in the trenches doing the medical work. And we need the people on the administration side who access and manage the IT systems that are going to have to host the new algorithms.
I’ve been in the room with those three groups many times, but it’s not something I see happening a lot. Often it’s two of the three groups. But until you have all three speaking each other’s language, things aren’t going to move forward as fast as we’d like.
There’s another stage to this – something I think of as a human intelligence-artificial intelligence cycle. As we start to put the AI into practice, the humans will start to respond to it, and of course their behavior will change. And then their judgments will change, which will change the training data for the next generation of the AI. I’m very excited about learning from and understanding that cycle, and having both the AI and the “HI” get better as a result.
Prof. Eric Topol, MD
Founder and Director of the Scripps Research Translational Institute,
Professor of Molecular Medicine at The Scripps Research Institute, and
Senior Consultant at the Division of Cardiovascular Diseases at Scripps Clinic
First, I’d love to see a really good online program or syllabus built for clinicians that doesn’t take too much time, but gives them an intro to AI and ML. The aim is to give a grounding.
On top of this, I’d like to see a meeting place for discussions, where people can get together with others and access some kind of synergy.
I applaud these kinds of efforts, and I think they’re very important. At the moment, I think isolation drives us to go after problems that don’t exist. We leave all the unmet needs unmet, and we come up with ways to predict things that shouldn’t be predicted, because they’re unhelpful, or things that could actually be hurtful. We can do much better.
Full recording of our March 30 Revolutionizing Healthcare session
Join our community!
If you’re a practicing clinician would like to be part of our ongoing discussions exploring the future of machine learning for healthcare, please sign up for our Revolutionizing Healthcare roundtable series.
Tim Schubert
Andreas Bedorf
