future view

From earthquake model to predictive model for IC admissions

31 July 2020 • 3 min reading time

As part of the Brains4corona initiative, TNO, in partnership with Amsterdam UMC, Erasmus MC, and Leiden UMC, has developed what is known as the SEIR model, which can predict the number of IC admissions. The surprising thing is the inspiration behind it – a numerical-based method for predicting earthquakes.

Solutions are sometimes found where they are least expected. After all, a model for predicting how changes under the ground in Groningen are translated into earthquakes would at first sight appear to have little in common with addressing the coronavirus crisis. But nothing could be further from the truth.

From earthquakes to IC admissions predictions

Jan-Diederik van Wees, who is an earth sciences professor at Utrecht University and a researcher at TNO, was convinced in late January that the new coronavirus would reach the Netherlands. ’Given the information from China – spreading under the radar and infections doubling every two days – it was more or less inevitable. When the RIVM (National Institute for Public Health and the Environment) stated in a briefing to the House of Representatives that it had no model for estimating IC demands, I decided to look into the matter in greater depth. After all, there was no reason why such a model could not be devised. I first started working on it in my own time, and then as part of a Brains4corona project,’ explains Van Wees.

‘I quickly realised that the “earthquake model” could be converted to predict IC admissions

Dealing with uncertainty

Van Wees quickly understood that the type of mathematical model used for predicting earthquakes could also be suitable for dealing with the coronavirus situation. ’Below the ground, there are changes taking place, due to gas extraction for example, which spread out. But the situation is very uncertain. The same type of basic characteristics are a feature of coronavirus too: there are only a few parameters that offer any kind of structure. Much still needs to be learned about how the virus is transmitted, after all. However, even with limited information, we can make reliable predictions using data assimilation. Together with the Amsterdam UMC and Erasmus MC, we devised the open source “COVID-SEIR model” in mid-March. And it correctly predicted the rapid increase in demand for IC beds in late March and the peak in early April,’ says Van Wees.

‘Many countries are easing their restrictions and then tightening them again when things go wrong’

How does the SEIR model work?

SEIR stands for Susceptible (the vulnerable part of the population), Exposed (the part of the population carrying the virus), Infected (patients showing the initial mild symptoms) and Removed (patients who have recovered or died). The model takes account of uncertainties for parametrisation – expressing COVID-19 in parameters – and social distancing measures. The model shows the drastic effects that small changes can have. ‘Just a minor change to the parameters can have a far-reaching effect on the predictions of new cases, hospital admissions, and deaths. It illustrates how crucial accurate data can be in tackling the virus. It can be calibrated daily and therefore give an up-to-date picture based on the most recent data. This means it provides some comfort among all the uncertainty,’ says Van Wees. The model and Dutch predictions have been made public.

Preference for adaptive control

Van Wees continues, ’Many countries are easing their restrictions and then tightening them again when things go wrong. Using the model, we have studied different exit strategies and carried out stress tests. We have observed that when restrictions are eased gradually, you need to take small steps to prevent IC units being overwhelmed. You also need an emergency stop in order to act whenever IC admissions increase rapidly. This results in a strong wave-like effect. An adaptive strategy, in which you can alternate easing measures with more restrictive ones, is better. This ultimately allows for more easing measures, because you do not get large up and down movements. No handbrake is needed and at the same time, you learn better what does and what does not work.

‘An adaptive strategy, in which easing measures can be alternated with tightening them, works best’

Other applications

At present, the focus is on healthcare capacity, but other applications are possible too. ’In fact, you could apply this model to anything that logistically progresses in small stages. That includes infections, hospital admissions, and therefore IC admissions. The advantage is that you can make daily predictions with little, but nonetheless up-to-date, data. It means, for example, that the Red Cross is able to use the model in Africa, where there is no extensive contact tracing.’ 
We are also working in partnership with the University of Amsterdam on a public web version for daily healthcare capacity predictions. Users will be able to carry out interactive simulations involving easing or tightening measures, and thereby gain a clearer picture and a better understanding of COVID-19.

Care to know more? Then please contact Jan-Diederik van Wees.