NEuropean scientist Karl Friston, from University College London, built a mathematical model of the functioning of the human brain. Lately, he has applied his modeling to Covid-19, and uses what he learns as a suggestion Independent Sage, a committee formed as an alternative to the UK government’s official pandemic advisory body, the Scientific Advisory Group for Emergencies (Sage).
How the model you use is different from conventional epidemiologists rely to give advice to the government in this pandemic?
The conventional model basically matches the curve for historical data and then estimates the curve in the future. They see the surface of phenomena – observable parts, or data. Our approach, which borrows from physics and in particular works Richard Feynman, enter under the hood. It tries to capture the mathematical structure of phenomena – in this case, the pandemic – and to understand the causes of what is observed. Because we do not know all the causes, we must conclude. But that conclusion, and implicit uncertainty, is built into the model. That’s why we call it generative model, because it contains everything you need to know to produce data. As more data comes in, you adjust your beliefs about the causes, so that your model simulates the data as accurately and simply as possible.
Could you give an example of what you mean by uncertainty, regarding Covid-19, and how you included it in your model?
A common type of epidemiological model in use today is the SEIR model, which assumes that people must be in one of four countries – vulnerable (S), exposed (E), infected (I) or recovered (R). Unfortunately, reality does not break them down very neatly. For example, what does it mean to be restored? We know that with Covid-19 you can be infected but show no symptoms, so does that mean recovering from symptoms or recovering from an infection? And that question hides a number of other questions, including questions relating to national testing strategies. The SEIR model starts to fall apart when you think about the fundamental causes of data. You need a model that allows all possible conditions, and assesses what is important to form a pandemic trajectory over time.
This is the first time a generative approach has been applied to a pandemic. Has it proven itself in another domain?
These techniques have enjoyed tremendous success since they moved from physics. They’ve been running your iPhone and your nuclear power station for a long time. In my field, neurobiology, we call this the dynamic causal modeling (DCM) approach. We cannot see the state of the brain directly, but we can conclude it with brain imaging data. In fact, we have pushed that idea even further. We think the brain might do its own dynamic causal modeling, reducing its uncertainty about the cause of the data fed by the senses. We call this the principle of free energy. But are you talking about a pandemic or brain, the fundamental problem is the same – You are trying to understand a complicated system that changes over time. In that case, I didn’t do anything new. Data is produced by Covid-19 patients rather than neurons, but if not, it’s just another day at the office.
You say generative models are also more efficient than conventional ones. What do you mean?
Epidemiologists now handle the problem of inference with numbers on a large scale, utilizing high-performance computers. Imagine you want to simulate an outbreak in Scotland. Using a conventional approach, this will take you a day or longer with today’s computing resources. And that is just to simulate one model or hypothesis – a set of parameters and a set of initial conditions. Using DCM, you can do the same thing in one minute. It allows you to print different hypotheses quickly and easily, and thus quickly get the best.
Are there other benefits?
Yes With the conventional SEIR model, intervention and supervision is something that you add to the model – tweaks or disturbances – so you can see its effects on morbidity and mortality. But with the generative model these things are built into the model itself, along with everything that matters. Our responses as individuals – and as a community – become part of the epidemiological process, part of a self-regulating self-monitoring system. That means it is possible to predict not only the number of cases and deaths in the future, but also the response of the community and institutions – and to attach the exact date to these predictions.
How well your prediction has been proven in this first wave of infection?
For London, we are predicted that hospital admissions will peak on April 5, deaths will peak five days later, and occupancy of critical care units will not exceed capacity – meaning Nightingale hospital is not needed. We also predict that improvements will be seen in the capital on 8 May allowing social steps to distance them – which they announced in the prime minister’s announcement on May 10. Until now our predictions have been accurate in one or two days, so there is a predictive validity for our model that is lacking.
What is your role with Independent Sage?
I am a member with special responsibilities for modeling. When they first approached me, I did not see “Independent” … I was joking, but only partially. I consider Independent Sage as the main exercise in public involvement; how would it be if you and I and everyone could sit in a real Sage meeting. I have heard self-defending politicians say that its existence greatly weakens the original Sage, but as a scientist I cannot subscribe to it. In my view, there is never anything wrong with transparent, information-based discussions. The other committee’s role, which is just as important, is to provide an alternative hypothesis to the British government – to provide more room for maneuvering.
What does your model say about the risk of the second wave?
The models support the idea that what happens in the next few weeks will not have a big impact in terms of triggering a rebound – because the population is protected to some extent by the immunity obtained during the first wave. The real concern is that the second wave could erupt in the next few months when the immunity is gone. We can test a series of hypotheses, based on a very short duration of immunity – from the common cold to the immunity that lasts for decades. For each duration we can calculate the probability that a second wave will appear, and when. These are the earliest days for this work, and I hope with sincere excitement for new data about immunity to become available, now reliable antibodies test there is. But the important message is that we have a window of opportunity now, to get the test-and-trace protocol in place before the second wave is suspected. If this is applied coherently, we can potentially delay the wave beyond the time horizon where treatment or vaccine is available, in a way we could not do before the first.
After the pandemic is over, can you use your model to ask which country responds best?
It has already happened, as part of our efforts to understand the latent causes of data. We have compared Britain and Germany to try to explain the relatively low mortality rate in Germany. The answer is sometimes counterintuitive. For example, it seems that German death rates are low not because of their superior testing capacity, but rather because of the fact that on average Germans are less likely to be infected and die than British people in general. Why? There are various possible explanations, but what seems increasingly likely is that Germany has more immunological “dark matter” – people who are immune to infection, perhaps because they are geographically isolated or have some sort of natural resistance. It’s like dark matter in the universe: we can’t see it, but we know it has to be there to explain what we can see. Knowing its existence is useful for our preparation for the second wave, because it shows that targeted testing of those at high risk for Covid-19 may be a better approach than non-selective testing of the entire population.
What are the future generative models of disease modeling?
That’s a question for epidemiologists – they are experts. But I would be very surprised if at least some parts of the epidemiological community did not become more committed to this approach in the future, given the impact that Feynman’s ideas had on so many other disciplines.
Finally, a Wired Interview the word You like smoking, don’t talk to anyone before noon, don’t have a cellphone and regret meeting one on one. HaThat’s one of thchanged during lockdown?
I do not think so. It is true that this can be considered a one-on-one meeting, but my default mode is to share ideas in groups – the Independent Sage style – and normal service will resume soon. Right before I talked to you, I declined the invitation to talk on the radio in the morning, and now I go to smoke.