Marjolein Fokkema: ‘My algorithms produce increasingly flexible decision trees for mental-health professionals’
Making predictions about emotional problems or the effects of air pollution: Marjolein Fokkema’s algorithms are getting better at this all the time. She is making her algorithms increasingly flexible, so they can predict not just characteristics at one particular moment, but also how skills, for example, improve or deteriorate.
The largest children’s mental health centre in the United Kingdom, the Anna Freud Centre, has a wealth of data on more than 400,000 children. Marjolein Fokkema met the founder of the centre’s evidence-based practice unit at a conference, and she was very interested in Fokkema’s algorithms. Together they began applying these to the Anna Freud Centre’s data.
Fokkema: ‘My algorithms can help predict an improvement and deterioration in children’s mental health. They take the data and produce a clear decision tree or flow chart, which a mental-health practitioner can easily consult.’
The key predictor of a developmental delay is the first split in the decision tree.
You begin by ‘feeding’ such an algorithm a large number of characteristics of a large number of children, for instance their home situation and certain test results. Then the algorithm looks at which characteristic is the key predictor of a developmental delay, for instance: this characteristic represents the first split or branch in the decision tree. The algorithm then looks at what the second most important characteristic is, and so on. The decision tree tells the practitioner the order in which they should look for relevant characteristics when estimating a risk or choosing the right treatment.
Change over time also relevant
This is similar to the algorithms that Fokkema’s colleague Elise Dusseldorp is perfecting. The big difference is that Fokkema is working on increasingly flexible algorithms. ‘My algorithms can also take into consideration characteristics that change over time. Say you’ve measured a child’s reading skills 20 times and want to take the course of time as a characteristic. You then need a different statistical model than for static characteristics or ones that you have only measured once.’
Once she is able to include the time dimension in her algorithms, much more will suddenly be possible. ‘The algorithms are already fairly good at modelling linear growth. I’m now trying to get them to model non-linear growth. Non-linear models would be more help with predicting the effects of air pollution on plant or animal behaviour, an application that I’m working on with biologists. Or the curve at which a person’s reading skills will develop, or cognitive skills in the elderly. In analyses, for the sake of ease, we often assume a linear development of such skills even though we know that isn’t true.’
A curve as result
Fokkema wants to develop decision trees or flow charts that produce not only one prediction of skills, symptoms or risks, but also different curves. Practitioners would then be able to estimate whether a client will improve quickly, slowly or in a strongly changing pattern, for instance, or whether they will be at risk of a sudden relapse.
There is one form of flexibility that Fokkema would like to develop in her algorithms. ‘The data now determine everything. Imagine you’ve done multiple tests on children but know that test A is more reliable than test B. You can’t yet include this prior knowledge in the algorithm. It looks for the characteristic that makes the most difference and gives this as the first branch in the decision tree. If that is the result of test B, the algorithm is missing the point. I want the algorithm to also be able to learn from this kind of knowledge and practical insight from experts.’
Text: Rianne Lindhout
Photo: Patricia Nauta