You are a co-inventor of the Human Patient Simulator (HPS®). Can you tell us a bit more about this technology?
Many people know the flight-simulator, which is a device on which pilots can train their skills but also learn how to respond to new or even emergency situations, for instance, due to unusual weather conditions or technical failures of the aircraft. These training simulators have enormous advantages over real-life training, obviously because they are much cheaper but also because it is not really possible to train pilots in dangerous circumstances in real life. Typical of these situations is that pilots have to assess the situation and take decisions with far-going consequences under time pressure. We all want pilots to make the right decisions in those moments where human lives are at risk.
The Human Patient Simulator is a similar device for training physicians on such "high-performance tasks."
It is called a patient simulator. Does this mean that physicians learn to operate patients by sitting at a computer screen and act through an operating panel, like those pilots?
No, no, the device consists of a mannequin with the looks, the sizes, and the responses of a real patient. Its fidelity results from a balance between training needs and costs, and it is used for training purposes in a real hospital environment and in life-like emergency situations, in which a team of medical specialists and nurses work together to save the patient.
With real blood?
Well, fake blood is used to simulate bleeding but one unique aspect of the HPS is that it consumes real oxygen, produces real carbon dioxide, and reacts to real anesthetic gases. From a physiological perspective, the patient’s respiration, circulation, gas transport as well as the interactions between these physiological sub-systems are represented.
What kind of training can you do with such a robot patient?
Different kinds of medical situations, such as to train a team in how to respond to a heart attack occurring outside or inside a hospital environment or to train an anesthetist in emergency situations occurring in the operating theatre, such as a ruptured aortic aneurism. The simulator responds to the blood loss and to the drugs and fluids given in these situations.
How does someone like you invent such a machine?
The human patient simulator has quite a long history, which started at the University of Florida in the mid-1980s, while the first release was in 1994. For my Masters research in electrical engineering at the Technical University of Eindhoven, I got a fellowship at the University of Toulouse in the then emerging new field, biomedical engineering. I worked on circulation and ventilation, and later on a heart-lung machine, on which I eventually did my Ph.D. Both projects involved modelling and simulation of human physiology. This experience was crucial in adding reactivity to the human patient simulator, in which I got involved in 1992.
The type of research that you did was not mere electrical engineering. How would you characterize the domain? Is it some kind of control and systems engineering, similar to current approaches in systems biology?
Well, I only came to agree with that view some three or four yours ago. As a student, I always thought that my predecessors in electrical and control engineering had invented systems theory, which is not true. Physiology in the 19th century is already a systems science, and later on, biologists expanded systems thinking when they started to think about ecology. But what electrical and control engineers did contribute is the much more quantitative side, which has led to a synergy between the systems view and the uses of mathematics in modelling, and of computers in simulation of electrical and other systems.
So, although I did not know it back then, we were applying systems biology, systems pharmacology, and multi-physics modelling, before those terms became popular. A major challenge was to couple the real gas exchange and drug administration systems to the mathematical models making the simulated patient evolve and react. Another challenge was the number of interacting physiological and pharmacological subsystems. What we did was to apply control engineering concepts and approaches to the modeling and simulation of physiological and pharmacological systems, which at that time was not as straightforward as it may seem now. And only very recently, I have begun to see how big a jump this is.
Why is that?
An electrical circuit and the model of an electrical circuit map onto each other in a straightforward and accurate manner, whereas we had to use analogies and approximations to match our models to the biological domain. For instance, we used the analogy of an electrical circuit consisting of typical electrical components such as resistors, capacitors and inductors to model ventilation or circulation in the body.
So, although we use the same techniques such as block-diagrams and circuit-diagrams, using similar or even identical symbols as in electrical engineering, we have to keep in mind - and I also teach this to my students - that we are actually making a few big jumps in terms of approximation and abstraction. We run into the question of the conceptual validity of our model.
The same seems to be true when using modeling techniques from mechanical or chemical engineering?
These models also match the man-made mechanical or chemical systems quite closely, albeit generally not with the same level of accuracy as for electrical systems. But when you start to use them for the modeling of naturally occurring physiological systems they often involve approximations, with which we must be much more careful.
How do you make the difference between models that are literal or close to reality versus those that are analogies?
Well, the point is that a model should be a good model, for which I use two notions. The one is conceptual validity and the other is operational validity. Conceptual validity has to do with the quality of the information that you put into the model – a model is considered to be conceptually valid if incorporated theories, assumptions and data are reasonable for the intended purpose. However, this does not tell you anything about whether the model will suit reality. Operational validity concerns the observed behavior of the model – how the model behaves from the outside. When you do an experiment in reality, for instance with animals or with human subjects, you obtain data, which you compare with the output of the same experiment that you run with the model. If the two agree in a number of different conditions sufficiently, the accuracy of the model is acceptable for the intended purpose.
Is the complexity of biological systems a problem for making these models? In a patient’s body, there are so many variables and possible interactions.
Not so much for the practical applications we work on, although careful approximations and simplifications have to be made. In electrical engineering, we often need to handle quite complex systems in terms of the number of elements and interactions. The biggest differences and difficulties lie in the variability between patients and in the limited available information, certainly concerning a specific patient.
Let’s go back to the very beginning of your career, to the question how you developed the original heart-lung machine. This was not yet a training-system, right?
No. It was a machine used in open-heart surgery to temporarily replace the heart and the lung of the patient. My contribution to that project started in 1987. The research question was to develop a better control strategy for this machine, and to get there we modelled the machine and - to a certain extent - the patient. So, we modelled the pump and gas exchange functions of the machine, and the circulatory system and gas uptake and delivery of the patient. We expected that with such a model we could at least test the safety controls that could be built into the machine to help the perfusionist (the person who knows how to operate the machines that keep a patient alive during open-heart surgery).
What were the challenges of modeling this machine?
The number of processes involved, of which we did not have enough knowledge, like blood-flow, gas exchange, acid-base balance, drug delivery, and so on. I needed to learn about many different kinds of physical and physiological processes and use knowledge about these processes from different scientific disciplines. We now call this multi-physics modelling, which is certainly different from modelling a system in electrical engineering.
In 1992, after your PhD you moved from the biomedical engineering center in Toulouse to this group of professor Good in Gainesville, Florida. He is specialized in anesthesiology.
This group in Florida worked on medical training simulators for acute care, which are the areas, in which if something goes wrong the patient could die in a few minutes. Anesthesia, intensive care, and emergency medicine are the typical areas of acute care. When I arrived, they were working on an anesthesia simulator, and they wanted their simulator to respond to the patient. So they needed someone who could transform their prototype simulator into a fully interactive, model-driven system.
Over the years we have developed a number of full-body model-driven training simulators. Behind the observed behavior of the mannequin and the signals on the monitors are physical and mathematical models of physiology and pharmacology.
The number of monitored variables and options for therapeutic interventions are huge. For example, we modelled the effect of 55 intravenous agents, all independent or input variables. Dependent or output variables included 5 different monitored blood pressures and - at the time - 4 electrocardiogram leads. Most of my former education and research had focused on single input-output models. In this research project, our modeling challenge was to connect these inputs and outputs in a meaningful way, so that the model would allow you to simulate various patients realistically and in various conditions.
This sounds like new challenges?
Different from the project in Toulouse on the heart-lung machine, the challenge of this training simulator was that it depended on even more interconnected systems, which increases the model complexity. We also had to build a mannequin that looks and behaves like a human patient. It has to interface to its environment, so that a team of physicians and nurses can perform procedures like injecting medication into the blood-stream and administering nitrous oxide via the inspiratory gases. Furthermore, the simulator should generate data on the state of the patient. For training purposes, different patients and scenarios must be programmable.
How do you validate such complex models?
Initially, we had a quite intuitive approach to validation. We preferred to use models published in the scientific literature but we had to design a few of our own. We coupled 14 different models of physiology and pharmacology. We made extensive use of expert opinion, with whom we discussed the design of the models as well as their output. These experts could tell us whether the output was plausible or not. They played an important role in assessing what I now call conceptual and operational validity.
Usually, we believe that we have a good model, a model that we can trust, if it contains all the processes involved, including the kinetic data for quantifying the dynamics of the system. But one of the difficulties of such modelling is that very often we don’t have those data, especially on interactions of different processes, as these may have been only measured in isolation.
We do try to structure that a bit more by first validating models of subsystems such as the circulation, pulmonary gas exchange, or specific reflexes, sometimes with animal data, sometimes with human data, sometimes via expert opinion. But the integrated response is something we can only assess with the help of expert opinion. Anyway, you have to keep in mind that strange things can happen when you connect these components together. New phenomena may appear that we also have to validate.
Does that often happen?
I can give you a funny example. One of the first demonstrations with our anesthesia simulator was in Vail in Colorado. I stayed in Florida but I got a very worried call from my colleague, who told me that the machine – this mannequin that was shown to our clients – had started to hyperventilate. So, I asked him, what is your altitude? It turns out that Vail is at 2400 meter and room-air oxygen is much lower than 21 %.
Would this happen with a real patient as well?
If a real person all of a sudden goes to 3000 meters, he may indeed start to hyperventilate. But a real person will adapt to this height over time – a phenomenon we had not modelled. So, we had to re-program our simulator to reflect adaptation to altitude.
Is it also possible to discover unknown physiological phenomena by simulations on such a machine?
When we did chest compression on the simulator, we observed that we created a blood flow, which was to be expected. But we also started to realize that by chest compression alone we actually create enough ventilation to keep the patient alive. Some 10 or 15 years later, the American heart association reduced their requirements for mouth-to-mouth ventilation, not always popular with rescuers at a heart attack, because chest compression alone appears to be enough to keep blood-flow and ventilation going – which is what we had already discovered on our simulator.
Your examples suggest that your simulator is not only important as a training device in hospitals, but also as a device that can be used to experiment on?
Definitely. New medical devices and procedures are already tested on our simulators. The underlying models are accurate enough descriptions for such purposes. Once we are able to predict outcomes on individual patients, many other clinical applications come within reach, such as selecting the best drug and drug dose for this particular patient in this particular condition.
That would be an important contribution to the so-called personalized medicine?
Something that strikes me is your trust in the knowledge of physiological, pharmacological and physical processes built into your simulator, whereas in my own field – chemical engineering – I used to be quite skeptical about the reliability of knowledge produced at idealized conditions in labs for modeling specific industrial systems.
An important difference may be the limited time-scales at which we work, which during an acute situation is very short. So, our simulator only needs to reflect fast processes, which are relatively basic, such as blood flow, oxygen transport, carbon dioxide elimination, a bit of the brain-stem control of these processes, and fast drugs such as fast anesthetic agents. I am not saying that our descriptions are perfect, but we can get close enough for the training applications.
This seems to be an important aspect of control and systems engineering: that in your modelling you only take into account, you assess, what is relevant to the specificities of the situation. You are not interested in some kind of complete description.
That’s right. We rarely look at the tissue level, the cell level or the molecular level. Although I know a bit about systems biology, I don’t consider myself a true systems biologist. Because that is really a bottom-up approach, whereas I take the opposite, top-down approach, from the person, to the organs, and sometimes to the tissues but not much deeper.
But part of this is also what systems biologists say. So what would you call specific of the control engineering approach to the modelling of these systems?
To model our system, we start by looking at three different types of variables at three levels: carriers, solutes, and signals. Through those variables, we can distinguish between the associated interconnected macroscopic processes in human physiology such as ventilation and circulation for the carriers; oxygen, carbon dioxide, drug and hormone transport for the solutes; and for the signals, mostly the control by the brainstem of ventilation and circulation. This approach, this distinction between levels and subsystems is based on what we know of physical, chemical and physiological processes.
What is so interesting in the distinction that we make between carriers, solutes and signals is that we map knowledge and approaches commonly used in engineering disciplines on physiological phenomena – we take a fluid mechanics approach to gases and blood (the carriers) in the body, a chemical engineering approach to the transport of solutes between - and metabolism within - organs, and an electrical engineering approach to nervous signals in the body.
Also specific to the modelling mind set of control engineers is that we look at what the model requirements are, in terms of relating relevant input to output variables, rather than striving for absolute completeness.
In the minds of scientists and also philosophers of science, engineers work very intuitively, but when I hear you talking, it involves much more systematic thinking about what you are doing?
I must admit that it usually starts very intuitively. But this is probably also true for many scientists. If we were very systematic all the time, there would not be any creativity. At the same time, systematic approaches are intuitive themselves, they have become second nature, and we are not even aware when using them.
Nevertheless, you have taken up the effort to think more systematically about how you did your research projects. You have even written a textbook that aims to teach students about modelling and simulation in biomedical engineering, by which you have contributed to the methodologies of systems and control engineering.
Well, thanks. Yes, I do think that systematic ways of analyzing models, which I put forward in this textbook, can help in finding structure. It helps with more precise ideas about "good" models and simulators, and an easier and faster research and design trajectory of the next generation; we still have some way to go before we will be at the level of realism of flight simulators.
My last question. You started saying that philosophy of science has helped you in getting your own ideas clearer, especially on explanation. I would be very interested to hear a bit more about it.
Recently, I have started to ask myself if and how models and simulators could help clinicians understand the physiology of their patient better in the clinical environment. The question is then what we mean by understanding and explanation, and how to apply this in acute care situations. Reading philosophy of science gave me a more precise vocabulary, and the limits of the traditional concept of explanation when applied to complex systems became clearer to me. A more specific concept I read about is that of ‘explanatory relevance’, which turned out to be very useful in my thoughts about designing models and simulators for explanatory uses apt for these specific training contexts.
Think, for instance, of clinicians who have to decide and act very fast in an emergency situation. Not everything can be covered by standard protocols. Certain situations would benefit from explanatory models that help them diagnose and think through those situations, communicate effectively about them, and intervene on the patient. But these models need to apply to the particular patient and situation while not overburdening the clinician with unnecessary detail. How this can be achieved is my next challenge.
Interview by Mieke Boon. On the topic in this interview also see: Boon, Mieke (forthcoming). “An engineering paradigm in the biomedical sciences: Knowledge as epistemic tool.” Progress in Biophysics and Molecular Biology (special issue: Validation and Models in Computational Biomedical Science: Philosophy, Engineering and Science).