In the fourth edition of AG Connect this year, Andre Marquand was interviewed about his research. This article is a translation of the article that you can find in Dutch at agconnect.nl.
André Marquand, professor of Computational psychiatry at Donders Institute, uses his scientific background in both computer science and psychology to find a solution to this. It is quite a challenging task, though. "This is not about discovering a problem on a single cell or protein. These disorders are - besides a biological predisposition - also the result of experiences in development, in interactions with other people and a lot of environmental factors. It is a very complex interplay."
Fortunately, a lot can be measured these days. Demographic, geographical, economical and even medical data are increasingly available for research. Marquand: "In some ways, we are swimming in an ocean of data but the signals we are looking for are often very subtle and scarce.”
One example Marquand cites is that someone who has been followed for some time often suddenly gets up much earlier and engages in fewer social interactions. These can be early signs of depression. They involve detecting deviations from normal (anomaly), a kind of 'early warning system'. These signals may be evident in the data, but not yet immediately noticeable as symptoms in everyday life. Environmental factors come into play, such as that living in a city brings more daily stress. Someone with a certain sensitivity, may develop a condition as a result. The same applies to poverty.
Past
Marquand started research at the Donders Institute in 2014, after completing his PhD in brain scanning in England and working for some time in London's financial services industry on data science projects. He sees many data specialists regularly switching between different fields of work. "They can make good use of the knowledge they have gained about things like probability distributions and what to do with correlations. Many of those things translate very well. But I would stress that having subject matter knowledge is really very important. You need to know what the underlying aspects are that drive systems in a certain direction."
Present
So, Marquand's team tinkers with algorithms, but does not make everything themselves. That would be impossible. "There are so many steps in the analytical process. There are people who spend their entire careers trying to get rid of artefacts from brain scans. It would be nonsensical to replicate the expertise that is already there. We can often link a lot of existing software tools, libraries, and APIs. This allows us to focus more on new ways to analyse and interpret the data coming out of that pipeline."
Marquand thinks how well you can explain the technology used is very important for patient confidence that the doctor is making the right decision. But transparency is also important for properly interpreting the outcomes of a diagnosis. For example, an MRI scanner is often used to take an image of the brain. Such a scanner is very sensitive to the smallest movements of the head. "But someone with ADHD often has difficulty lying still in the scanner for a long time. When you then very naively start classifying the data, you can get seemingly good predictions from the analysis, but in fact measure nothing of the abnormality. You've just created a very expensive motion detector for the head. So, if you actually want to make a meaningful claim, you need to be able to demonstrate a direct relationship with the disease or abnormality."
One of the challenges facing Marquand's team is the difficulty of integrating different data sets to get a complete picture of a particular individual. For example, an MRI scan might show that an individual has a certain abnormality in some brain circuits. But it is also known that that person has had negative experiences in childhood and lives in a poorer neighbourhood where there is a lot of crime. This causes a lot of stress in daily life. Machine learning may be able to help integrate these different insights in a good way. That won't be easy, Marquand expects. "Nobody reacts to certain environmental factors by default. Those responses are all encoded in our brain in one way or another. If we understand how people differ from each other, then we can look for how a psychological condition can have an effect on that variation."
Future
Marquand sees many opportunities for the further development of this new field. There will be new algorithms and different types of data. His big question is whether all this can eventually lead to predictions that are not yet possible with the current approach. And whether basic scientific advances can also eventually be used for clinical research. "The holy grail in much medical research is to predict problems early and avoid the onset of symptoms. This not only prevents the patient from having to go through a lot of unpleasant years, but also saves the taxpayer a lot of money on healthcare provision, in the long run. Achieving that is still extremely difficult. It does raise interesting issues and as a data analyst, that's one of the things I really enjoy."