A look back before moving forward to 2023 keynote interviews
Prof Janaina [Jana-eena] Mourao-Miranda leads the Machine Learning and Neuroimaging Lab within the Centre for Medical Image Computing (CMIC) at the Department of Computer Science, University College London (UCL), where she applies pattern recognition and machine learning techniques to neuroimaging data. A major theme within Prof Mourao-Miranda’s research is uncovering the relationship between brain and behaviour.
At OHBM 2022, Prof Mourao-Miranda gave a keynote lecture on machine learning in neuroimaging and psychiatry. You can find a recording of Prof Mourao-Miranda’s talk here. Below is an edited transcript of an interview conducted with Prof Mourao-Miranda on June 17, 2022.
Yohan Yee (YY): Let’s start with a bit of background. Can you tell me a bit about yourself and how you got to where you are?
Janaina Mourao-Miranda (JMM): I started doing research during my undergrad in electronic engineering. I worked as a research assistant in a neuroscience lab in the north of Brazil that focused on electrophysiological recordings in animals. At the time I was fascinated by the similarities between brain circuits and electronic circuits. I was very lucky to have a great mentor, Prof Cristovam Picanço Diniz, who introduced me to neuroscience and really inspired me.
Once I finished my undergrad, I moved to the southeast of Brazil to do a Masters and a PhD. My Masters was in artificial intelligence and my PhD was in neuroscience, both at the Federal University of Rio de Janeiro. When I finished my PhD, I wanted to continue my academic career with a focus on research, but the opportunities in Brazil were very limited at the time, so I started looking for opportunities outside Brazil.
My first postdoc experience was in Germany, at Siemens, in Munich, and it was there that I started working with machine learning models applied to neuroimaging data. At the time this was a new field and the results from this work were published in NeuroImage as one of the first papers applying machine learning to neuroimaging. My second postdoc experience was at the Institute of Psychiatry, at King's College London, where I started working in collaboration with clinicians to apply machine learning to clinical neuroimaging data. After my second postdoctoral experience, I was motivated to become more independent so I started looking for fellowships. After applying for a couple of fellowships I was awarded a Wellcome Trust Career Development Fellowship in 2008, which enabled me to set up my own group at UCL—University College London— Later in 2013 I was awarded a Wellcome Trust Senior Research Fellowship to further develop my research.
Brains, behaviours, and things in-between
YY: That sounds like a long (but rewarding) path! So about your research, is the idea to take brain images of some sort and use machine learning to predict outcomes or treatment response?
JMM: My research changed over the years. In my first project using machine learning, we extracted features from some neuroimaging data—could be functional or structural data—and used that to train classifiers. Our aim was to develop diagnostic approaches for mental health disorders based on brain scans. This worked in the beginning, because we had very homogeneous samples. We could train the models and get some good accuracy. What happened over the years, with myself, my collaborators, and many other groups around the world, is that we started seeing that when the sample sizes increase, the accuracy of the models tend to decrease, particularly because of the heterogeneity that we know we have in psychiatry. So we started to rethink the models. We moved away from the classification problem and started treating the problem as a regression. So instead of predicting patients vs. controls we started to predict symptom levels as continuous measures from neuroimaging data using transdiagnostic samples. This approach works well, demonstrating that there is an association between neuroimaging features and symptoms regardless of the diagnosis, however it does not help us to find subgroups of patients for target treatment and intervention. More recently, we have been using multivariate approaches to combine brain, behaviour and other sorts of information, with the aim of finding some latent dimensions in the data that might be linked to mental health. We think that these might help us identify subgroups of patients (i.e. more homogeneous subgroups), and also, potentially, people at risk of developing mental health disorders.
YY: It's interesting that you mention there is a notion of (diagnostic) labels and predicting these labels is hard. So now you're going towards more latent dimensions. My impression of psychiatry is that seems to be the case across many different disorders. Is that fair to say?
JMM: It’s true: I think it's an acknowledged problem in psychiatry. This was one of the motivations for the National Institute of Mental Health to propose the Research Domain Criteria framework (RDoC) to find new ways of characterizing mental health disorders based on basic dimensions of functioning that might vary from abnormal to normal instead of categorical groups.
YY: Are there variables that seem to come out stronger than others? I mean—when I ask this question—I'm thinking back to a 2015 paper by Stephen Smith, on the positive...
JMM: positive-negative mode? Yes, it's very interesting. We replicated these results of finding brain-behaviour associations—this positive-negative dimension—in the HCP data, with different approaches. To answer the question: I think it depends on the sample. We used a sample of adolescents with 345 subjects between the ages of 14 and 24, with 10% having depression. There, we found two latent dimensions, one that captured a developmental aspect in the sample, and the other that captured a depressive aspect. So it really depends on what you put in the model.
I have one exciting study in particular that we're just about to publish. We applied a similar dimensional approach to a sample from the ABCD study. We used multivariate machine learning approaches to find these latent dimensions between the grey matter voxelwise features and psychosocial variables, including behavioural, clinical, and other types of information. We found six latent dimensions, three of which were related to mental health and three of which were related to IQ (and income) and other cognitive features. We're really excited that these dimensions capture both cognitive and psychopathological aspects.
YY: And even the fact that you're finding some of these latent dimensions in healthy participants is really cool, because...
JMM: ...it shows that mental health and cognition are variable across a healthy population.
YY: Where is the field going—will psychiatry use more machine learning? Or will there still be a psychiatrist in the clinic?
JMM: I think machine learning will be more used, but a psychiatrist will always be needed. I think whatever we do, it's just to add information, and help with diagnoses and clinical decisions.
Building models: from assumptions to complexity
“You should always start with the simplest possible model and add complexity as needed”
YY: It sounds like there's a lot of work going into this in terms of predicting brain and behaviour and the relationships between them. Can you speak a bit more about the models that you're using? From what you've described, it seems like you're using some sort of PLS (partial least squares) or CCA (canonical correlation analysis)?
JMM: What we've been using for these types of approaches are regularized versions of PLS and CCA, which can work with very high dimensional data. CCA in particular has the limitation that if it's not regularized, then you can only use it when the number of variables are smaller than the number of samples. When you work with brain images you have hundreds of thousands of voxelwise features. In addition, the item levels from assessments we use to track behaviour are also on the order of hundreds. Not all the samples have the scale of the ABCD study’s 11,000 subjects, so we use regularized models such as regularized CCA and sparse PLS. A lot of the work we have been doing as a group is developing new models because, as we use the currently available models, we realize that they still have some limitations. I can say that there is actually—as far as I know—no proper sparse CCA approach that works with very high dimensional data, in the order of hundreds of thousands features. Usually, the sparse CCA models make the assumption that the covariance matrix is an identity matrix (which means the features are uncorrelated) in order to be able to work with high dimensional data, with the implication that they become a PLS model that maximize the covariance of the latent variables instead of a CCA model that maximizes the correlation. It means that the associations we find with these models are driven more by the variance in the brain and the behaviour data than by the relationship between them. This is a limitation. We are now working on developing better models: a PhD student in my group is developing a new framework for regularized CCA which will enable using different types of regularization and sparsity.
[Note: a tutorial on PLS and CCA for identifying brain-behaviour relationships can be found here.]
YY: I see models lying on a spectrum where you have simple linear models on one end, and then as you go the other way, you get more complexity, like deep neural networks. Can you comment a bit on how these models are being used and what's the value of models on each end of the spectrum?
JMM: Yeah, that's a good question. I think there have been a number of studies in the past few years showing that there is not much advantage in using deep learning models as compared to linear machine learning models for neuroimaging data. We have a similar experience with that; we took part in the ABCD challenge in 2019 to predict fluid IQ from structural MRI features. Our team tried many different approaches from linear machine learning models to deep learning models, and our team won the challenge. Surprisingly, the model that gave the best performance was kernel ridge regression, which is simply a regularized linear regression. As deep learning further develops, we might see more examples of the benefits of applying it to neuroimaging data. But my take-home message is that for many predictive tasks we don’t need complex models in order to get very good performance. I suggest that you should always start with the simplest possible model and add complexity as needed.
Advice to trainees and the community
“Choose an area of research that you’re really passionate about and that you really find, very, very interesting, because this will help you move forward during challenging times”
YY: At OHBM, there are many trainees. Do you have a message for the young folks in the crowd on the things they should do as they start their careers in science?
JMM: I think, for young researchers, my first advice would be to choose an area of research that you’re really passionate about and that you really find very, very interesting, because this will help you move forward during challenging times. Also, finding good mentors that can give you strategic advice throughout your career and help establish a good network of collaborations is essential for success.
YY: You've mentioned mentors several times. Are there specific people that have been good mentors to you?
JMM: There are four people who have mentored and inspired me throughout my career. The first one was my first supervisor, Professor Cristovam Picanco Diniz from when I started doing research as a research assistant. He was a great mentor and educator. He used to discuss my career path with me and what I could achieve so this really inspired me. The second one was my former PhD supervisor, Professor Eliane Volchan from the Federal University of Rio de Janeiro, who is also a neuroscientist, and she is one of the most meticulous researchers I've met. She taught us the importance of understanding our experimental design and all the steps of the analysis approach we're using. The third one was Professor John Shawe-Taylor from University College London. As head of the Computer Science department, he was the sponsor of my Wellcome Trust fellowships, and he has been a person that I can go and discuss the challenges we face when applying machine learning to clinical neuroimaging and other types of data. It's really great to have discussions with him and think about, okay, if these models are not working, what can we develop to address this challenge? Finally, a person who has inspired me in the beginning of my career is Professor Karl Friston from UCL. I read all his early papers during my PhD and I was really fascinated by all the different approaches he was proposing to analyze fMRI data.
YY: Have you encountered any challenges in your career? And if so, how did you overcome them?
JMM: As a female researcher from Brazil I faced many challenges in my career. I think the biggest one in particular, early in my career coming from Brazil, was a financial one. When I started doing research and until I finished my PhD, we had very limited resources in the lab to do the experiments and to attend international conferences. But at the same time, we had this feeling that we could do it. We had a strong sense of community and supported each other. We would find creative ways to do the experiments and to go to conferences—we even used to write letters to airlines to get discounts to go to international conferences! And we did get discounts and we were able to go! So I think that if you have a supportive environment and are really motivated you can find ways around challenges.
JMM: One of the things OHBM has is a travel award [Merit Abstract Award and Merit Abstract Travel Stipend] which helps people from low-income countries attend OHBM conferences. I think these awards should continue because they increase the diversity of the community, which is very important. They can also have a huge impact on early career researchers: putting them in touch with mentors, and increasing their networks of collaborators. It is really, really important.
[Note: OHBM also offers reduced registration costs for members from developing countries]
YY: Finally, is there a moment during your career that you were proudest?
JMM: There is a short story that also tells how important it is for young researchers to take part in OHBM. One of the proudest moments of my career was during my first OHBM actually, back in 2004. Karl Friston was one of the people who inspired me, and he had a talk at the educational course on fMRI. The title of the talk was something like “the three new methods I like the best”. So I was really excited to go to the talk and see what the three methods he was going to talk about were. He started talking about the three areas that he thought were promising, and one of them was machine learning. So I thought, great, I'm on the right path. Then he gave examples of a few conference abstracts that were within this area. One of them looked like the title of my abstract, so I thought, oh no, I started doing something interesting and someone already has done it! But then he showed the abstracts on the screen, and at that moment, I realized that he was talking about my abstract! I couldn't believe it; I was really proud!
YY: We've spoken a lot about science and a bit about OHBM. What are your interests outside of science? And how do you balance that with being a scientist because, as I'm sure everyone knows, it's a very time consuming career.
JMM: Outside science, I love being a mom—I have an eight-year-old son. I also love being active: going to the gym, running, cycling, and swimming. It's really difficult to balance all of these, and I don't think there is a recipe for that. We're always trying to find ways to balance things, but sometimes you need to push more on one side, and other times, push more on the other side. I try to have a balance between being a researcher, being a mom and being active, and it is not easy. A strategy I found useful is to be selective. In terms of research, this means focusing on projects that I find really, really relevant. I think it's more efficient to focus on a few important things, so that I can have a bit more time to spend with my family.