News Report / Profile
The subtleties of the brain a step away from being unveiled by a Startup
We meet at the terrace of the IMM, while she has lunch in a hurry, something healthy she carries in her bag. It's November, but the sun is still there. Spontaneous and communicative, she describes herself as optimistic and tenacious, but it was stubborness that brought her to Portugal and has kept her here. She came three years ago from King's College in London, where she lived for twelve years and lectured for a while, but her true vocation is research. In London, she looked up images of the brain with schizophrenia, here she studies that feelings go through molecules and that all our behaviours have a biochemical explanation even when we are not aware of it. Diana Prata is 38 years old and is a Cognitive Neuroscientist. After winning a prestigious European Union Marie Curie Fellowship for her first project in Portugal and an FCT research contract, she came to the IMM with the Social Cognition project. As an independent lecturer at the Faculty of Medicine of the University of Lisbon, she was immediately invited by the Director of the Radiology Service, and lectures in several PhD and Master Degree programmes of the University of Lisbon and of King’s College London, and, as a Guest Professor, at ISCTE. What made us talk to each other was the startup in which she is involved, which has already won some awards. The latest is partially sponsored by the Presidency of the Republic and took her to the Web Summit in search of new investors.
It's called NeuroPsyCAD and is a project that aims at the early discovery of neuropsychiatric diseases like Alzheimer's and Parkinson's using a software that analyses images of the brain. The cognitive subtleties, which the medical eye does not detect, are brought together and a statistical result is drawn with a diagnosis.
All is being done to reach the goal in 2020, but that will require funds that only a large investor can afford.
How did this idea of creating an artificial intelligence system that knows how to early detect diseases such as Alzheimer's and Parkinson's come about?
This was already a PhD project of Ricardo Maximiano (another of the co-founders of NeuroPsyCAD).
It was a matter of opportunity, because I was already working with the same techniques to respond to other things. I was studying schizophrenia at King's College, at the Institute of Psychiatry, and I learned genetic and neuroimaging techniques, applying them to the study of schizophrenia. My last study was to try to distinguish which young people suffering from mild paranoid delusions would develop schizophrenia or not. To try to predict, I looked at a patient's brain, as well as his family and psychological antecedents, and put all those elements into an equation, an algorithm. The same applies now to Alzheimer's. I was studying these techniques while Hugo Ferreira (CEO and co-founder of NeuroPsyCAD) and Ricardo were creating these techniques and this is how our collaboration started. I was used to collecting patients' brain images and analysing them with software created by engineers. So Hugo, who in turn developed this kind of software, seemed like an ideal partner. He came up to me and asked: "Don’t you want to do the Cohitec programme?" We went on a six-month entrepreneurship course and we won a hundred thousand euros from Caixa Capital, in 2016. In May 2017 we were able to set up the company and create employment for four people.
What is an algorithm?
An algorithm is an equation in which several variables coexist. An example: when we are children, we learn what a table is without anyone having told us how to recognize one, that it, because it has 4 legs, or a top. Even because none of these characteristics would be enough to identify a table (and, for example, distinguish it from a chair). We learn by ourselves that it is a pattern of features. This pattern is like an equation of variables, including the fact it has a top, leg height, the context in which the supposed table is... that is, it is the combination of everything that makes up our idea of a "table". What our brain does is automatically and through pure exposure, as a child, to examples of tables (and chairs) discover this discriminatory pattern. It is our pattern recognition algorithm for tables, and although it needs exposure to environment examples, this capability is encoded in our genes. We try to build these algorithms with characteristics from within the brain to make the most of the information the person has inside, about the presence of Parkinson's disease and Alzheimer's. The magnetic resonance gives us the image and we have to interpret it in a statistical way, which is not possible with our eyes.
Are we going to be able to delay Alzheimer's and Parkinson's diseases?
These diagnoses do not heal, of course, but they help showing doctors which treatment most effectively alleviates the symptoms and as early as possible, and thus bring greater quality of life to these patients. It's a way to delay the disease. We want to give doctors more confidence so that they diagnose the disease sooner and from there decide the best treatment.
In Parkinson's, 10 to 17% of cases that are diagnosed as such and that look like Parkinson's are not actually because they do not have the dopaminergic deficit that is the cause of Parkinson's. These patients are not being treated well and there are side effects, in addition to more costs. Nowadays, there is even a type of test that distinguishes these diseases in the beginning, but it is five times more expensive. It is called DATScan, a technique in which it is necessary to give an injection of radiopharmaceuticals and the requires a type of scanner that is much less accessible (because there are few). In the case of Parkinson's, with less cost, we can make a conclusive and cheaper diagnosis, and start treating the patient according to what the condition he/she really has.
In the case of Alzheimer's, the question is whether the slight cognitive decline that the person has is of the more natural and gradual type compared to Alzheimer's, which has cerebral changes that can only be seen post mortem and where the decline is much faster, crippling and severe. At an early stage of the disease, doctors cannot distinguish between these two types or predict the evolution. In this case, what doctors do is look at a scan and compare it with a memorized image that they have, and have learned as being a sign of Alzheimer's. If the disease is already in an advanced stage, it is easier to see the brain changes, but that is not where we want to help. We want to help the patient who presents the first complaints, where the brain changes are still subtle and very scattered.
This the reason why we use artificial intelligence algorithms that are especially good at detecting small change patterns. They are pattern recognition or automatic learning algorithms. This software is able to learn from the data we give it. If we give it images of patients with Parkinson's disease and of others without this dopaminergic deficit, it begins to learn which small changes pattern can distinguish the two conditions.
What we do is to point out a number with the likelihood of disease risk. In future, will be able to analyse the brain with regard to the study of diseases such as schizophrenia, autism or depression.
Why the choice of these two diseases?
We realized that these diseases were giving us good results, and doctors are already more accustomed to doing scans. The techniques we use are precisely for that, because we do not start from prior knowledge, we let the machine compare everything and do not condition it, it will conclude something based on the data and not on what I am saying. It may even discover more diseases or other diseases that I had not expected. There are many diseases that present the same symptom, the machine identifies a pattern that can distinguish three different types. The physician's role will be to analyse the correct use of the drug in the exact type of the diagnosed disease, as in one type the medication may work in one way and in the other it may not work, for example. Another advantage of these artificial intelligence techniques is that we do not need to restrict ourselves to prior knowledge because algorithms learn from the data rather than from what we are telling them to choose. They may even discover more illnesses or other diseases beyond the one we imagine. For example, with some diseases with the same symptom, indistinguishable through the medical eye, the software can find a pattern that distinguishes the three different types. Or it may find a pattern that distinguishes, before prescribing treatment, two groups of patients with the same diagnosis and symptoms but responding to different drugs. In one type, the medication may work one way and in another it may not even work, for example.
How does the doctor respond to the findings of a scientist?
Doctors will have more confidence in their decision if more reports, such as ours, are made available to support their therapeutic decision-making, making it more objective. It is the doctor who asks for this exam and then it is he/she who does the follow-up after the exam, so we do not replace the doctor.
Where does the human factor stand in the midst of the progress of Artificial Intelligence?
The doctor needs to advise and to follow-up. The doctor has a placebo effect through his empathy and this is irreplaceable. A software or a robot do not know how to create complicity. They can be programmed by us to show empathy, return a smile if we do it too, but it is a false empathy, as it is very difficult to convince a human being as well as another human being. In my opinion, the metacognition that occurs when I feel or know that you know that I know, etc.. is not possible to be reachable by a robot.
Image by Tiago Figueiredo