Caricaturing Brain Activity Yields Better Prediction of Behaviors

Caricature artists exaggerate distinctive features of an individual, deepening a cleft chin or multiplying freckles. Yale researchers have now applied a similar approach to maps of neural connections, emphasizing individual differences to see if they yield useful information.

Turns out they do, according to the researchers' findings published Nov. 3 in Nature Neuroscience.

Researchers have been constructing and studying these maps, known as connectomes, to see if they might be predictive of, for instance, behaviors or mental health conditions.

This research has so far found that connectome activity that is similar across individuals is important and can be predictive of behavior all by itself. So the remaining activity has largely been cast aside.

“But what’s going on in that activity? It has been left behind, so we really don’t know whether there’s value in it,” says lead author Raimundo Rodriguez, a PhD student in the Interdepartmental Neuroscience Program at Yale School of Medicine (YSM).

When Rodriguez caricatured connectome data, minimizing shared activity and thereby emphasizing individual differences, he found that connectomes were better predictors of several features, including age, IQ, and emotion processing.

“What we’re finding is that information carried in caricatured data is distinct from non-caricatured data,” he says.

For the study, the researchers used Human Connectome Project, UCLA Consortium for Neuropsychiatric Phenomics, and Yale-developed functional magnetic resonance imaging (fMRI) datasets.

First, Rodriguez identified key patterns in the datasets where, across individuals, different brain regions activated or deactivated together while individuals were completing some sort of task. And then he removed those shared patterns from the fMRI data that were collected when the individuals were at rest.

Doing so, Rodriguez confirmed, made individuals look less like each other in the data. It also made separate scans from the same individual easier to identify from the rest of the scans. “That told us that this method really was caricaturing the data,” says Rodriguez.

The question that remained was how this might affect efforts to predict behavior and characteristics.

“We looked at a range of features and showed that often times, this caricaturing method actually improves predictive capabilities,” says Rodriguez, who works in the lab of Dustin Scheinost, PhD, associate professor of radiology and biomedical imaging at YSM. “But what’s interesting is that it didn’t do so for everything.”

Caricatured connectomes better predicted individuals’ ages, IQs, sex, and BMI, as well as performance on tasks assessing emotional processing and the ability to identify similarities across objects. But caricaturing was less predicative of borderline personality disorder.

“That there’s this nuance in where prediction improves and where it doesn’t tells us that this method isn’t simply ‘cleaning’ the data. It’s not just removing noise,” says Scheinost, senior author of the study and associate director of biomedical imaging technology at the Yale Biomedical Imaging Institute. “Instead, these findings suggest that differing activity patterns may hold important information for predicting some characteristics and behaviors while shared patterns hold importance for others, meaning the two are offering different types of information.”

The researchers found support for that idea as well. When they combined caricatured data with non-caricatured data, they got even better predictions than with either one separately.

Going forward, the researchers want to identify what behaviors caricatured data does and doesn’t work for, and whether there are patterns underlying why.

“Ultimately, we’ve shown that there is this new source of information that we’ve so far been ignoring,” says Rodriguez. “But we can use it to improve prediction.”