“It is precisely our plasticity, our long childhood, that prevents a slavish adherence to genetically programmed behavior in human beings more than in any other species.”
― Carl Sagan, Dragons of Eden: Speculations on the Evolution of Human Intelligence
I first learned about Ted Satterthwaite’s work when I started teaching about resting state fMRI and motion artifacts. His research showed how motion affects resting state connectivity measures, and I was thrilled that his group also compared the variety of effects with different preprocessing pipelines. In Mexico, every year we host a Neuroimaging Meeting where we invite neuroimaging researchers to visit the city of Guanajuato, [binge] eat Mexican food and talk to students, and so we were delighted to invite Ted to our 2019 meeting.
From our time together there, I got to know more about Ted and his research program. He is currently an Associate Professor in the Department of Psychiatry at the University of Pennsylvania Perelman School of Medicine, and the Director of the Lifespan Informatics & Neuroimaging Center. As a psychiatrist, he is highly interested in human development and building huge development datasets.
When I was asked to do this interview I knew it was going to be difficult to focus on a topic, but we managed to come up with a coherent chat, which I hope we can soon repeat with some beer and mezcal.
Eduardo A. Garza-Villarreal (EG): How did a psychiatrist end up writing these influential methodological papers such as the effect of movement on BOLD signal? And how did you go from psychiatry to methods? What was your career path?
Ted Satterthwaite (TS): I am a psychiatrist, and the reason I got into research was to try to develop tools that could be useful for the diagnosis and treatment of mental illness. That being said, I quickly learned if we ignore the methods, we probably can’t make progress on the ultimate clinical problems that we're interested in. I've primarily been working on large scale datasets, because to me, the only question in psychiatry is clinical heterogeneity. For example: someone comes to see me in the clinic and I diagnose them with depression. But clearly, depression is not one “thing” – depression is almost certainly many different things that we call one thing. It is pretty clear that to parse heterogeneity, we need large studies, because your ability to parse heterogeneity will be determined by both the noise in your signal and the size of your sample. Since we have noisy signals, we probably need very large samples. However, when having very large samples, although you get a lot of statistical power, you also become incredibly sensitive to confounding signals. There's a history in psychiatric imaging of being worried about medication confounds as well as other types of confounds. When we started doing developmental studies, that's right around the time my twins were born. Now, you don't need to scan thousands of kids to recognize that they don't sit still, you just need to come to dinner at my house. And it brought up this obvious question of “is movement going to impact our measures?”, and we started thinking about it because we could see the artifact in the time series. We were just very surprised. We assumed that this had already been solved. However, although there were papers from Karl Friston and others from a decade ago on motion in task-fMRI, there was nothing for functional connectivity. At that point it was just a practical question because we wanted to study brain development, we wanted to study psychopathology, and we know both age and illness are associated with in-scanner motion.
EG: Why are you so interested in development? And why do you think it's an important topic of research in psychiatry?
TS: The dominant paradigm in psychiatry now is that, when you see someone with their first onset symptoms like a severe mood disorder or psychosis, it's not like something went wrong in their brain right there that caused them to have the symptoms. It’s not like a switch that was flipped. Rather, there is accruing evidence for many years now suggesting that most mental illnesses are neurodevelopmental in origin. So, the goal is to understand how the brain develops normally, and then understand also how abnormal patterns of brain development are associated with different sorts of psychopathology. If you think about other fields of medicine, the way they've made clinical progress is by getting there earlier and unpacking heterogeneity. Think about cancer. We used to diagnose cancer only when we found huge tumors that had spread widely, and they were diagnosed on physical examination, like palpating the abdomen and saying, “you have a tumor”. Advances in oncology and in other fields, they have been predicated on both getting there earlier and unpacking heterogeneity; saying it's not just a tumor, but that it is a malignancy from this tissue, with this receptor profile and that genetics, and as a result it's going to be sensitive to this treatment. I think we're still at that “it's a tumor” phase in psychiatry. My hope is that by better understanding patterns of brain development and heterogeneity within the disorders we treat, we can get there earlier and ultimately achieve better outcomes for patients.
EG: As you say, psychiatric disorders are very heterogeneous, and there's a lot of overlap especially in symptoms. What do you think about the current disorder classification, about the Research Domain Criteria (RDoC), and about this overlap between disorders?
TS: There are two sides of the same coin: heterogeneity within disorders, and non-specificity across disorders. The RDoC framework is trying to map symptom profiles to brain systems, and I think that's a totally laudable approach. The one challenge of all this, though, is we don't have a ground truth. We don't have anything like postmortem pathology in neurodegenerative diseases, and so the challenge right now, for example in machine learning techniques, is that we don't know what the labels should be. You can have the best engineers and the best pattern recognition algorithms, but a lot of advances in machine learning have been based on supervised learning, and right now, we just don't have the best labels. Without good labels, it's kind of garbage in, garbage out. What I see is that the biggest challenge is using biological data like images to help us understand what those labels should be. I think it's a challenge and we're all still grappling with it.
EG: Do you think machine learning will have some influence on the future of psychiatry, or you think it's one more tool in the bag?
TS: I think both. Machine learning is going to have a big influence, especially as we have larger and larger training datasets, and the ability to generalize things across samples. However, part of the problem right now is that we don't have a lot of datasets where we can link multivariate patterns from machine learning tools to outcomes of interest, which I think is an essential step. Datasets like the Adolescent Brain Cognitive Development (ABCD) Study that follow 10,000 kids over 10 years are actually really important starting points because longitudinal prognosis is a very important clinical outcome to be able to predict. However, there is still a lot of room to make advances in terms of incorporating these methods using health system data where we have medical records on medications and hospitalizations, but we have a long way to go. Also, I think machine learning won't solve everything. Dani Bassett, one of my closest collaborators, makes a very cogent case that machine learning, while great, alone will not do it; you need a combination of good machine learning and sound theory. I agree. We should not forget that we know a lot about the brain from decades of basic neuroscience – we need to incorporate that knowledge as priors to inform these tools and help interpret the results.
EG: To you, what is your lab’s most interesting project and why?
TS: Right now, the one I am most excited about is a super ambitious project that I lead together with Michael Milham at the Child Mind Institute. We're trying to get a lot of the larger studies of brain development - around 11,000 samples in total – and make sure they are pristinely curated, processed, and QA’d ahead of a public data release on the International Neuroimaging Data-Sharing Initiative (INDI). The project is called the RBC - “Reproducible Brain Charts” – project. One thing that I've learned is that a lot of the sexy stuff in science is actually easier than the non-sexy stuff like data organization, curation, the really low-level things that are necessary for reproducible neuroscience. These things are often really time consuming. What makes me excited about this project with Mike and his team is that if we can do those things well, the dataset will be much more useful to everyone else, and everyone can just run faster by not having to recreate the wheel. I think that's super exciting. But it's been a huge challenge, because it's very heterogeneous data from different studies, which meant we've had to build new tools to handle it.
EG: Can you tell us a bit about the tools you're developing?
TS: Sure-- there’s a couple, which are at different stages. Sydney Covitz and Matt Cieslak have a really cool tool that is being presented as a OHBM poster at the meeting called “Cu-BIDS” for “Curation of BIDS”. It is a tool for curating Brain Imaging Data Structure (BIDS) formatted data at scale, designed for very large heterogeneous datasets -- we needed it to be able to handle the RBC data. I wish I had a time machine and we had that a couple years ago, because it makes life so much easier. And then once you get that data curated into BIDS, and this is all building off from Russ Poldrack, Chris Gorgolewski and Oscar Esteban’s work with BIDS and fMRIrep, we're focusing on running everything in containerized pipelines. For example, xcpEngine is our containerized post processing pipeline that was developed in the lab by Azeez Adebimpe (now) and Rastko Ciric (originally); it consumes fMRIPrep output to produce derived measures for studies of functional connectivity. Now, a lot of efforts are about moving beyond fMRI. For example, I am super excited about Matt Cieslak’s works to build QSIPrep, which is a fully containerized and highly generalizable BIDS-app for diffusion images. Lastly, Azeez is just finishing up ASLPrep for ASL MRI data. In the end, the goal of all this is to make it all super reproducible and open. If you talk to other neuroscientists, one of the biggest reservations they have about imaging is that it's so complicated, and there's so much data processing... that they just don't believe it. Over time, I’ve come to really agree with other people who have been doing open science for many more years than me -- I think the only answer that will get people to believe us is to just show them everything; sunlight really is the best disinfectant.
EG: In terms of your research achievements, what are you most proud of?
TS: You know, it's funny. What makes me most excited is not the individual projects, but the people I work with. I've just been super lucky to have amazing students who do really incredible work. This year’s keynote leverages work from many awesome people in the lab. The title of the talk is from a review from Valerie Sydnor, one of my grad students, who just dug into complex literature and produced a work of incredible scholarship. Similarly, some of the developmental data I am most excited about is from the hard work of another grad student, Adam Pines. What I'm most proud of is the trainees and how much heart they put into their work, and how much they teach me. That's the fun part for me, the people, not any individual finding.
EG: To me one key discovery was about motion artifacts and how it affects the signal. Even right now there is a huge debate on things like the global signal. I don’t think we have it figured out.
TS: Yes, it remains contentious. I think studies of motion artifacts however is a great example of how science should work. It was awesome – we had that paper in March of 2012, but in January of 2012, Koene Van Dijk and Randy Buchner published almost the same finding in an independent dataset. In February 2012 you got Jonathan Power and Steve Petersen with the same finding and also a method for handling it. So, three different labs, working independently with different data sets – all coming to the same conclusion -- producing three papers in NeuroImage at almost the same time. I think that's a great example of how science can work to provide convergent evidence.
EG: What do you think psychiatry should go for in the future? Where do you think is going? You mentioned looking for an actual ground truth, do you think we're going to get it at some point?
TS: People ask that all the time, “when is this stuff going to be useful?” And I think the first answer is: it's not useful now, to be honest. But that doesn't mean it's not a super important problem. If we do make progress, this sort of work could be incredibly impactful because psychiatric disorders are among the most common afflictions that humans get. But in the end, to really show something, we're going to need clinical trials and outcomes that matter. Some people who have been starting towards this, like Nicholas Koutsouleris and the PRONIA consortium – they are doing some really cool work. In the end we will not convince practicing physicians that this brain imaging in psychiatry matters until we show real results in clinical trials. And that's a challenge—perhaps a 10- or 20-year challenge. But I think we'll get there.
EG: Thank you Ted for taking the time to sit down for this interview. Looking forward to your keynote at OHBM 2021.