Building the Star Trek Med-Bay

Dr. James A. Glazier is working to build a digital twin of the human system. A roadmap has been published and the long-term implications for tackling diseases like cancer are immense. We discuss what this all means, what the timescale might be, and why he thinks "we may wind up being cyborgs in ten or 15 years".

Below is a transcript of this podcast. Some parts have been edited for clarity. 

Michael Hainsworth: Digital twins aren't just for factory floors, next generation airplanes and municipal traffic control systems. At Indiana University Bloomington, Professor James Glazier is among the scientists working to twin our immune systems to better fight infection and disease. But unlike the world around us, getting big data shoehorned into an AI is next to impossible until humans make the leap to cyborgs. He tells me his work is nothing short of changing the way medicine is performed in the future.

James Glazier: At the moment, medicine is an open loop. You wait till you get sick, you go to the doctor, the doctor makes a prescription, you go home, you take it, you wait some time, usually weeks, you report on what happens, you go back, and you try again. Almost every engineered system, even simple ones, has built into it a concept of control. You don't turn your thermostat, your heater system on, your furnace on, leave it there and then go and turn it off. There's a thermostat.

You don't hand adjust things in the engineered world, they're controlled for you. That makes a huge difference. We have autopilots on airplanes, we have cruise control on cars these days. Modern cars are getting better and better at doing collision avoidance on their own. And that kind of control is something that's missing in medicine today. Now, it's not completely absent. If you have a pacemaker, the pacemaker will monitor your heart condition and an emergency will apply a shock to your heart.

If you have an artificial pancreas, it will measure your glucose and insulin levels and deliver insulin as needed to maintain them appropriately. This kind of closed loop system is beginning to come into medicine. Another place you see it would be an Automatic Positive Airway Pressure (APAP) machine. If you wear it at night for sleep apnea, that machine will measure your breathing and will adjust itself to accommodate what you need. But most medicine is still reactive rather than proactive. It's an open loop rather than a closed loop.

Now you could say that the doctor himself or herself is a digital twin. The doctor has some opinion about what your state of health is, how it will change, what needs to be done to change that health state in a desirable way. But the timescale on which that's done is very slow. What digital twins for medicine are trying to do is to speed up that cycle so that you can diagnose problems faster, ideally before they become serious, predict what will happen, and ultimately, design therapies that are more effective for individuals. We hear a lot of words about personalized medicine and precision medicine, but we're still very, very far from actually delivering personalized and precision medicine, because the element of time needs to be included to deliver true personalized medicine.

MH: Well, help me understand how you build a digital twin, because I can understand the concept of a digital twin on a factory floor. You've got Internet of Things (IoT) sensors for the assembly line, you've got cameras that monitor individuals, you've got an incredible amount of data that's coming into an artificial intelligence machine learning system to build that digital twin. What are the data points that allow you to build a digital twin from the human body? Are you going to have to put machinery in me?

JG: Well, that's an interesting question. I think we may wind up being cyborgs in ten or 15 years. If you ask what the pieces are you need for a digital twin, you've identified them very well. One of the big differences between an industrial digital twin and a medical digital twin is that you don't have a design, you don't have a blueprint for a human being. And there's a lot of uncertainty about how human beings work.

There's also a lot of stochasticity. I can't rerun time, but if I could rerun time, the way your body plays out would be slightly different every time. Your reaction to the same treatment won't be identical every time. We're a little bit more in the world of weather forecasting, where we say that there's a probability that the tornado or the hurricane will go in a particular way, and we keep updating our forecast based on what we see.

Ideally, what we want to be able to do then is, if we have a lot of data points, we have a lot of data for many individuals. We know that our models work some of the time and they don't work others of the time, and we'd be able to refine our models to make them better and more predictive. That worked very well in weather forecasting, but it was also a hard job. Weather forecasting still can't go out for more than about a week, even with all the modern computing, and it's taken a long time to get there.

But if you ask what it was like 30 years ago, basically you could predict the weather a day or two in advance. Now they could go out for about a week. I think the thing to think about is what we can do to predict short times and then ask the question, how do you extend the time you can predict? And the other key questions, as you point out, will be instrumentation. What can you measure?

A third thing, which is important, is what kind of model you use to do the prediction. You mentioned machines, machine learning, artificial intelligence, forward prediction. And that works very well in many cases, and it's used a lot today. But if you look at industrial digital twins and you look at the history of digital twins, for example, General Electric, they started out with detailed mechanistic models of how the devices that they were going to predict worked.

There are models like that for parts of the human body, for aspects of human physiology, but because there's a lot of scientific uncertainty and we don't have blueprints, those models are pretty incomplete. And so, one of the things that we face when we build medical digital twins is a choice about whether we're going to use machine learning database methods or mechanistic methods to predict. They both have advantages and disadvantages, and ultimately, we're going to have to put them together. There's no way that you could use either one by itself.

The big problem with machine learning is that it can't handle what are called contrafactuals. Machine learning is wonderful for interpolating, but it's not so good for extrapolating. Sometimes you're lucky and extrapolation works, but in general, if you have two individuals, and the person you're talking about now is in between them, machine learning will predict pretty well. But the problem with medicine is that, usually the cases you're trying to predict, you're going to be treating somebody who's never been treated before.

The illness is going to be different from the illness in other people. And if you want to personalize, you face the problem that you don't have that in the training set. And therefore, pure machine learning tends to do poorly in these situations, because it doesn't have the experience. By definition, if it's a medical case, you can't treat and not treat the same patient, and therefore, you cannot have data in your database about what that patient would do. And that's why mechanistic models have some advantages, because they represent the best scientific understanding and our best predictivity of what that body would do under a particular situation.

MH: It sounds like what you're saying is that we wouldn't necessarily need to build a digital twin for each individual person, because that would just be unwieldy. We need to build digital twins as a general concept and then apply them to individuals, and AI isn't necessarily the most effective way to accomplish that?

JG: Let me be clear about this. There's a term that the Environmental Protection Agency developed which I like, which is called the Virtual-Tissue Modeling System (VTM). And that is the idea of representing, in a computational predictive way, as much mechanistic information as you can about a biological system. And they use that to make prediction, especially developmental toxicities for drugs and chemicals, industrial chemicals.

One of the problems we face, for example in current medicine, is that we have a lot of nice ways of getting drug candidates. You could do rational drug design with molecular dynamics; you could do high throughput and you could identify candidates very well. But what we can't do is predict the outcomes of those drugs. The success rates for clinical trials haven't gone up in the past 25 years despite everything. We're better at generating the candidates, but we're not better at predicting what they're going to do.

And the reason for that is that one of the things in engineering as a principle is that when you're going to do control, you want to measure as close to the point of control as possible, and you want to apply the control as close to the outcome as possible. In medicine, especially with drugs, you're applying a molecular perturbation. And modern molecular biology is tremendous. We know all sorts of things about the molecular states of cells.

But if I ask the question, given a particular set of RNA-seq data or particular molecular composition of a cell, what is that cell doing? I can't tell you in general. Maybe if P53 is elevated, the cells dying through apoptosis. But in general, cells have complex internal organizations, they're machines. A parts list doesn't tell me how a car drives. And so, the key thing is that we have levels of organization in the body, which means that our control, which is at the molecular level, is very far from the outcomes we care about.

And that means that linear prediction, which is what machine learning is really good at, doesn't work very well, because the same control could have very different outcomes. There are molecules that you could take that are cytotoxic and that will kill cells in your liver, but that don't cause any long-term problems. There are other molecules that seem to have at the molecular level, very small perturbations, but at the system level, cause big problems.

And that's because you have within the cell complex organization and between cells, complex organization and between tissues, complex organization and between organs, complex organization. And just the way, when you drive a car, your experience driving the car isn't only of the car, it's of the city and the environment you're in. You have what technically we'd call emergence at multiple levels between the level of control, which is the molecule, and the level of the whole body, which is what you care about medically.

And that means that direct prediction is very difficult. I'm not promising that you'll have medical digital twins or immune digital twins in the next five years, I'm afraid. It's a hard problem and you've identified, I think, one of the killers, which is instrumentation. At the moment, we can't measure cytokine levels. If you're in the hospital, you're lucky if they do a cytokine profile once a week. If you ask the question about infectious diseases, by the time you have symptoms of the flu, your viral load is already peaking. That's why antivirals generally aren't useful. They work very well, but by the time you know you need an antiviral, it's too late to take them.

MH: I want to come back to your point about that you believe at some point we'll all be cyborgs, but let's put a pin in that for a moment and talk about that complexity, specifically because you've published a roadmap for building out the digital twin of the human immune system. What are those key milestones and when can we expect them?

JG: That's a great question. If you look at measurements today, we run this National Institute of Health (NIH) working group and do these seminars, and we try to get industrial representation. And the other day, we had a talk with a company that's essentially doing monitoring, biosensor monitoring, attached, a glue on biosensor. They are measuring heart rate, body temperature, blood oxygenation. The simple things that a Fitbit can measure essentially, although it's glued on so it's a little bit better than a Fitbit.

And interestingly, they started with mechanistic models, and they went to a pure AI model, because the kind of prediction they're trying to make doesn't need a bigger model and it's faster. The key thing that they try to do is they're trying to back out the unmeasurable things about the body from the things that you can measure. For example, if I look at body temperature, every human being's body temperature is different. Doing a population average on body temperature doesn't tell me much.

On the other hand, if your temperature is going up, that may tell you something significant about an infection. There were some very nice studies that were done in 2020, just at the beginning of COVID, that asked the question, not is heart rate elevated or is temperature elevated, but is heart variability greater? Or when this person is exercising, is their heart rate going up faster than it would have under normal conditions?

They're looking for predictive markers that would tell you that, your cytokine profile, your IL-2 and IL-6 in particular, which are the things that kick in that fever response, are beginning to be elevated before the person notices. In the case of diseases like Sepsis, conditions like Sepsis, where the immune state changes very, very rapidly, over periods of 15 or 20 minutes. Even an hour earlier notification that a person's going to get a septic shock could save thousands or even billions of lives in the hospital.

And I think the place you're going to see this kind of thing done first will probably be in intensive care wards. You're going to see people who have a central line. You're going to see the development of sensors that can measure cytokine panels that can be put in there, and that will give real-time measurement of the immune state to warn you when you're having a septic shower and these kinds of things.

Historically, companies like Eli Lilly had Xigris as an antisepsis medication, and that was a famous failure, because they couldn't give it appropriately. You gave it too much, the patient died of immune suppression. You gave too little, the patient died of septic shock. You couldn't adjust the dosing fast enough to make it effective.  

And I think that those kinds of medications, if you imagine some future world, you could imagine cytokine regulators. You have your panel of interleukins in an implantable device, the way you do for an artificial pancreas. You monitor immune state, and you regulate immune state in order to keep your immune system functioning optimally to deal with infection. You don't want to suppress it early, because you want to get the infection down. Diseases like flu are very good at interfering with interferons, because if they didn't, you wouldn't get the flu. It's a problem with things like steroid inhalers.  

MH: But let's come back to that roadmap idea. You just pointed out that a lot of these advances that'll get us to a digital twin of the immune system will start in places like the ICU, where you've got people already in a bed, they're already hooked up to machines, that's what they're accustomed to. So that's how we can build that digital twin. But having said that, where do we get to that next level?

You sort of hinted about the cyborg thing and we've talked about the limitations of building a digital twin being tied to the lack of data, and we need that. Is the Fitbit that thin edge to the wedge that gets us that digital twin? How far off is it? As you say, it could be more than five years before we have a digital twin. What needs to happen next that gets us down that road?

JG: Well, my colleague, Gary An, who's a surgeon and also a great scientist at the University of Vermont, likes to say that "just because something is difficult doesn't mean we shouldn't do it." I think that if we don't have a vision of where we're trying to go, I do believe in scientific progress. I believe that we will get there eventually, but I think that having a coordinated vision of what needs to be done could get us there a lot faster.

Now, is it fast enough that a typical VC in an era of high interest rates will want to fund it today? Maybe not. Three years ago, maybe in the era of low interest rates. But what do we need to be able to do it? We need to be able to measure the immune state. We need to be able to have models that take the immune state and predict what the immune state will be going forward. And that won't be a single model. That's going to be like weather forecasting, where you'll have an ensemble forecast of probabilities of the forward state.

Once you have that forward state, the question is how far you can predict forward. You need to decide whether the model prediction and the measured state agree or not. That sounds trivial. With temperature, the thermostat could say, "It should be this temperature. The actual temperature is this." We could make a difference. With things like the immune state, it's a little bit harder, because we don't know how to combine our signals in a way that gives us what's called a residual, a measurement of the error of our result.

And that's actually a hard problem. That's a hard scientific problem. Once we have that error, we need to be able to adjust our models. We need to be able to design self-improvement systems for our models, so this won't be for individuals only, although that'll be part of it. For the individual, you'll tune the parameters of the model. You'll adjust the model parameters as you run it, to make it more and more like that particular individual.

Your APAP machine works that way. It has a generic model of how a person works, and then it tunes it to your particular body as it sees how it works. And so, you'll have a template, and that template will then be customized for the individual. But when you see failures in the population, then you need to adjust the model structure. You're understanding essentially the science of how the system works, and we need better ways of doing that. The self-driving car people are doing that really well. That's the place where that technology will probably develop first.

MH: Every science fiction world has a medical bay (Med-Bay) where a robot or a bed or a gadget does all the heavy lifting for the doctor. Building a digital twin of a human, I believe, is the foundation of science fiction becoming science fact. And I've already tried a few times to pin you down on times and dates and milestones, and I can understand your reluctance to do that. But talk to the science fiction, wide-eyed boy in me. How long before we get the Star Trek Med Bay based upon the work you're doing today?

JG: I think it depends on a decision about funding and effort. If you look back at the Human Genome Project, it looked like it was never going to happen. Once it happened, it didn't happen primarily through classical science. There were big projects and there were commercial projects that did it.

Once you have an initial digital twin, even if it's a crummy one, the bootstrapping will be really fast. Improvement is really fast. The key thing is to get something out there. If it's based on a Fitbit and the data that we have, that's already better than nothing. And the question then is, can we build a model of how the immune system works, whether that's an AI model? And the problem with AI models is that they don't have understanding built into them. Or a mechanistic model, based on our scientific understanding of how your immune system works.

Once we have that, we can improve rapidly. I would say that you're already going to see the 'Fitbit type' and AI-based predictors coming in the next two or three years. I think you're going to have models of the immune system that are possible in a mechanistic way within three or four years. The sensor technology is harder, because getting under your skin is going to take more time. But I have colleagues, for example, at Notre Dame University, who are trying to do transdermal sensors based on sweat to quantify cytokine levels.

It's not quite there, but I think in five years there will be wearable sensors that will give you basic cytokine panels in real time. The prediction side of it and the computing power is also hard. There are these technologies on the computing side. How do we use the predictions? How do we build ensemble models? How do we improve ensemble models?

Now, that will take a little bit more time. Your point is an excellent one. If we cannot predict, we can't control it. If we have no idea what's going to happen when we give a drug, we won't give the drug. Everything we do in control is based on prediction. Your point about prediction is critical, but prediction is not enough.

If I want to do control, I'm in some state, I want to get to some state, and this is done in robotics, especially for robotics, for swimming robots, where there's a lot of fluid flow, so the robot can't necessarily get from where it is to where it wants to go. Maybe not be able to do it. And so, the robot will evaluate strategies for swimming from where it is to where it's trying to go. And there's noise. There are fluctuations in the environment. It doesn't know everything.

Solving those kinds of control problems, which are very analogous to medicine, is hard, because I don't just have to predict one course. I have to evaluate tens of thousands of possible treatment options and decide which one's the best. Just computationally, the forward prediction now is already expansive. If I want to evaluate tens of thousands or hundreds of thousands of control strategies, I need a lot more computing.

Now, that's probably five to ten years away, I would say. And then the third thing to bring medicine into the modern world, to have that vision of control, is you need actuators. You need something that can deliver the treatment, the dose, the drug, in a controlled way. And that's a little bit further off. The pumps are there, but as you know, nobody likes having an implantable insulin pump. It's not pleasant. We're going to have to go further with those. That's probably ten to 15 years away.

And there's a regulatory issue, which is at the moment, we depend, except maybe for pacemakers, on a human in the loop system. And initially, digital twins will be used by a doctor, to inform a doctor, to inform a patient. But to really achieve the maximum results, you want them to be automatic. You don't want to wait for the doctor to read the chart and make a decision. If you've got a real-time prediction, you want to have real time control. There's also going to have to be a decision about giving up some autonomy.

We're going to have to allow these systems to be able to make decisions for us. And that's threatening and it should be something we think about carefully. It shouldn't be something we do right away. Trust, understanding how these systems work when they fail. This comes up, of course, in self-driving cars, a lot. Came up with autopilot.

They're much safer overall, but they're going to fail at different times from the way things would've failed if a human were driving. What do you do about liability? Who's responsible? I don't want to minimize the difficulties. There's a whole stack of difficulties here, but I think within five years, you're going to see simple AI-based 'Fitbit type' digital twins.

And that as soon as those are there, you're going to see rapid improvement, because they learn from use. And again, the improvement in self-driving cars is so tremendous now that there are a lot of them out there. As soon as they're used, you'll see that improvement. It could be faster than I say, but I would say five years for a really effective initial one. And then I would say we're talking about 15 years. And medicine will change. The world will be different when we have them.

MH: Which jives with your analogy to the Human Genome Project, which started in 1990. It took 13 years before we had mapped the human genome, and then there was that massive explosion as a result of that knowledge that came from it. Are you optimistic that within that timeframe we will revolutionize medicine the same way as a digital twin of the human immune system?

JG: I think that if somebody today, Bill Gates and Mark Zuckerberg decided to put two billion dollars into developing an immune digital twin today in a serious way, in five years we would have one. I think it's achievable. I don't think it's achievable with the current model of dispersed funding. It would have to be a moonshot project, but I think it's possible.

<< Go to previous episode