Method for better prediction models

Predicting the extent to which a person is at risk of a particular disorder or predicting the course of an individual’s disease; the use of prediction models is increasing in the healthcare sector. The more patient data you have for such a model, the better the prediction. However, the data of individual patients from different hospitals in different countries is often not entirely comparable. Valentijn de Jong developed a method that uses the differences in characteristics of patients, hospitals and environments to develop better prediction models based on different datasets. He obtained his PhD on this on November 8.

Take the COVID-19 pandemic as an example. It is very important to know how likely it is that you will die from this. Ideally, you would want to put all the figures that are available worldwide into one model. But there are all kinds of factors that make these figures difficult to compare. In a country where the standard of healthcare is not so high, that mortality rate is probably higher. If the average age in a population is higher, the mortality rate is also higher. Valentijn: “The method developed now allows you to create prediction models that take all these things into account.”

Care for the individual
Prediction models also help to improve individual care. Valentijn mentions thrombosis as an example. A model can be used to predict the risk of thrombosis for a patient. The doctor will then know whether additional testing is necessary to detect thrombosis in good time. If the model shows correctly that the risk of thrombosis in a patient is very low, the patient does not need to undergo unnecessary invasive testing.

If such a prediction model has been developed in an American hospital, it is probably less suitable for predicting thrombosis in Dutch patients. This is because these patients and hospitals can differ greatly from each other. In order to make the model more widely usable, you need to collate data from studies in different countries. To be able to draw the right conclusions from these collated datasets, you need the advanced statistical methods I have been working on over the last few years. This leads to a prediction model that can also be used for Dutch patients.

If a new prediction model is to be developed based on multiple datasets, anyone can use the new method. They can be downloaded free of charge here or here. “I’ve already heard from several investigators that they are happy to use it,” says Valentijn. He has also worked out a way in which a model takes potential measurement errors into account. This also allows the use of data from hospitals and studies where the best measurement method was not used.

The results of Valentijn’s research are a step further in being able to work safely with prediction models. “The prediction models are mushrooming, and I’m sure they’re going to play an increasingly important role in future care. However, this development is still slow because researchers and doctors find that too often they can’t rely on it. Only if the prediction models produce really accurate results for different groups of people will they have an increasingly greater impact on patient care. My research contributes to that.”

Valentijn de Jong obtained his PhD on December 8 at Utrecht University. The title of his thesis is: Methods for individual participant data meta-analysis in prediction research