Anthony Barrows has had a long-time interest in using a multidisciplinary lens to explore complex systems. It seemed only natural that when presented with an opportunity to use computational tools he was developing to study one of the most complex human systems – the brain – he was eager to jump in.
“A colleague was working with a computational model of behavior, and it seemed like an interesting, novel approach for studying cognitive control, which I wasn’t especially familiar with, that would be easily combined with predictive modeling, which I was more familiar with, to ultimately give some insight into human behavior,” says Barrows of the University of Vermont. “I think this multidisciplinary approach is a great way to do science.”
Barrows and colleagues’ new study, published in the Journal of Cognitive Neuroscience, specifically looks at brain-behavior relationships in the Adolescent Brain Cognitive Development (ABCD) Study study, a large, longitudinal study of brain development in adolescence, looking at more than 11,000 young people from across the United States. “Improving our understanding of such relationships is critical for identifying opportunities for intervention to move youth towards favorable outcomes and away from unfavorable ones,” Barrows says. The study gives cognitive neuroscientists a new way to use computational models to test, understand, and predict brain activity patterns in the ABCD dataset and potentially beyond.
I spoke with Barrows about this new work and its implications (which was part of his recent PhD defense), as well as the future of this area of research.
CNS: What new insights were you seeking with this study?
Barrows: A colleague had recently developed a mathematical model of the processes that underlie behavior on the ABCD Study’s stop-signal task, which requires individuals to make simple perceptual decisions and to inhibit their responses when they are signaled to withhold from making a decision. They showed that this model could be used to measure a set of parameters that represent the brain mechanisms thought to generate those behaviors, and that the model produced data that was highly similar to children’s actual behavior in the study. A natural next step was to explore how closely these mechanistic parameters related to brain activity, as measured by neuroimaging, at the same time. Characterizing relationships between brain imaging and behavioral processes may help us understand the neurobiological underpinnings of cognition, psychopathology, and the possible impacts of substances that may affect cortical and subcortical development.
CNS: What is your favorite way to describe how the models work that you used in the new paper?
Barrows: There are two kinds of models in this paper. The first is a formal mathematical model that outlines a theory about the underlying brain mechanisms that allow people to make decisions and inhibit their responses when needed. It makes specific predictions about how long it takes someone to respond to a stimulus on a screen, the choices they make, and, in this case, how often they can withhold their response when asked to do so. The model is fit to behavioral data using Bayesian estimation. This form of model estimation is ideal for fitting complex models to humans’ behavioral data, which are often noisy and have limited numbers of observations. These models are interesting because they offer estimates of the underlying mechanisms of each person’s performance on the stop-signal task, while also taking everyone else’s performance into account (in this case, through the use of informative prior distributions) to get an accurate estimate.
The second type of model uses machine learning-based techniques to figure out which brain regions are most strongly related to the mechanistic parameters generated from the Bayesian models. This technique is broadly called predictive modeling, because you train a model to predict behavior using brain function from 80% of participants, then you test that model using brain imaging data from the remaining 20% of participants. If the behavioral predictions are accurate, your model is probably doing a decent job at characterizing the relationships between brain function and behavior in your sample.
CNS: What were you most excited to find? Were any findings surprising or contrary to popular opinion?
Barrows: A fair amount of research seeking to link brain function with observed behavior focuses on “localization,” searching for specific brain regions responsible for particular behaviors. A growing body of evidence suggests that patterns of brain activity that are widely distributed across many brain regions produce behaviors, rather than specific locations.
What we found supports this notion: although some region-specific associations emerged — like better general information processing being associated with activity in the anterior cingulate and bilateral insula, regions implicated in cognitive control and performance monitoring– the predominant pattern was widespread across the brain.
Further, we found the strongest associations between brain activity and a model parameter called efficiency of evidence accumulation (EEA). Estimated from the ABCD Study’s stop-signal task which, again, is designed to assess inhibition, EEA is best thought of as a general index of an individual’s ability to efficiently gather relevant information when making decisions. EEA appears to underlie performance across many different cognitive tasks and therefore can be estimated from performance on tasks beyond the stop-signal task. Remarkably, we found that EEA estimated from another task in the ABCD study (a working memory task) is highly related to the same quantity on the stop-signal task and that you can predict EEA on each of these two tasks well using brain imaging data from the other task.
The agreement between EEA across two very different cognitive tasks suggests we are measuring something with trait-like stability that may be a general underpinning of cognitive ability across many tasks.
CNS: What’s next for this line of work?
Barrows: The longitudinal nature of the ABCD Study affords an extraordinary opportunity to quantify the development of cognitive control through adolescence and young adulthood. This work only considered the baseline data collection wave, collected when participants were 9-10 years old. The years following this age range represent a critical period for the development of cognition and an excellent opportunity to model longitudinal development of cognitive processes.
CNS: Is there anything else I didn’t ask you about that you’d like to add?
Barrows: Associations between brain function and behavior are notoriously difficult to characterize and prone to small effect sizes. We note that although our study includes one of the strongest reported association strengths in the ABCD Study stop-signal task to date, we still only explain about 25% of variance in behavior using neural data. This limitation is meaningful, because it underscores the complexity of brain-behavior relationships, and how characterizing the neurobiological correlates of cognitive control is very much an ongoing field of work.
-Lisa M.P. Munoz
