Machine learning, fMRI offer insight into OCD patients’ response to therapy

A team of California researchers has developed a method for predicting responses of obsessive compulsive disorder (OCD) patients to cognitive behavioral therapy using machine learning and fMRI, according to work published in the journal PNAS.

Few researchers have been able to identify solid psychometric and demographic features that correlate to an OCD patient’s treatment response, PhD candidate Nicco Reggente and colleagues at UCLA’s David Geffen School of Medicine wrote. Most OCD patients, who often struggle with intrusive, recurring thoughts and repetitive behaviors, are often referred to cognitive behavioral therapy (CBT)—with a typically high success rate—but CBT is expensive, stressful and time-consuming.

Being able to predict a patient’s potential response to treatment would allow clinicians to appropriately allocate resources that support CBT for OCD and advance the current understanding of OCD as a neurophysiological disease, the authors said.

“I’ve always been interested in personalized medicine,” Reggente told Radiology Business. “Acknowledging that each individual is unique in their response to treatment, medication, et cetera allows for more precise and effective options for individuals to help. After meeting Jamie Feusner, the study’s senior author, and learning about the differences in response to CBT for his OCD patients, it felt like a natural extension to leverage my machine learning and neuroimaging experience to try and predict treatment.”

Reggente and the team applied a multivariate approach to their research; no studies to date have used the method, they wrote, and multivariate analyses of brain connectivity can offer the advantage of simultaneously capturing patterns involving multiple brain networks. Analysis was applied to whole-brain resting-state fMRIs acquired before and after patients underwent four weeks of intensive CBT.

Reggente said fMRI may be a pricier option for research, but it was the only viable one to examine the functional connectivity (FC) of particular brain regions with high enough spatial resolution.

“It would be much cheaper and quicker to use EEG data, but it’s unclear if we can resolve the source of the EEG signal with enough spatial specificity to narrow in on the same regions that were used for this successful classification,” he said.

The search was data-driven, the authors wrote, but restricted to networks where connectivity patterns had previously been found to change with treatment. Additional analyses of patients’ amygdalas were also included, since the amygdala is known to show value in predicting response to CBT for OCD individuals.

After combining machine learning and cross-validation to assess the power of FC patterns and predict individual post-treatment OCD severity, Reggente et al. found that pretreatment FC patterns within the default mode network and visual network “significantly” predicted post-treatment OCD severity, explaining up to 67 percent of variance. These networks, the authors wrote, were stronger predictors than pretreatment clinical scores.

“I believe that, once replicated and extended, this method could be used to screen and direct OCD patients,” Reggente said. “If the classifier were to predict you to be a non-responder to CBT, you could be directed down another route, like medication or neuromodulation. The same concept could theoretically be accomplished to predict non-responders to medication so that they could choose CBT instead.”