Can Machine Learning Predict Antidepressant Response?

Pauline Anderson

January 20, 2020

An inexpensive portable electroencephalography (EEG) device may help determine antidepressant response in patients with major depression, new research shows.

Investigators found that a machine-learning tool that analyzes resting-state EEG signals was able to predict response to the antidepressant escitalopram (Lexapro, Allergan) with an accuracy rate of 79.2% in adult patients with major depressive disorder (MDD).

Dr Faranak Farzan

"We're not too far from the day when doctors will have this rather inexpensive, portable EEG gadget in their office and could use it to make more informed decisions about prescribing medications. This is something we need desperately," study investigator Faranak Farzan, PhD, scientific director, Center for Engineering-Led Brain Research, and chair in technology innovation for youth addiction recovery and mental health, Simon Fraser University, Burnaby, British Columbia, Canada, told Medscape Medical News.

The study was published online January 3 in JAMA Network Open.

Determining Utility

EEG records the oscillations of electric activity in the brain through by means of electrodes that are attached to the scalp. Recordings reflect the synchronized activity of thousands, even millions, of neuronal populations in the brain.

Previous research provided compelling evidence that resting-state EEG can predict response to antidepressant medications. However, for the most part, sample sizes in these studies were small (10 to 50 subjects at most), and participants were drawn from a single center, said Farzan.

"This is not enough to train a machine to learn a pattern. Machine learning depends on the amount of learning samples ― in this case, patients with depression. To get good results, you need a lot of patients," Farzan said.

In addition, Farzan noted that previous studies suggested that EEG only "somewhat" predicts patient response to a particular treatment.

"It was not clear whether this is good enough to start building a tool that will be useful for clinicians and patients," she said.

For the current research, Farzan and colleagues used data from the Canadian Biomarker Integration Network in Depression (CAN-BIND) study.

They collected resting-state recordings from 122 patients with MDD from four Canadian centers (mean age, 36.3 years; 62.3% women) before treatment started. For these patients, the mean baseline score on the Montgomery-Åsberg Depression Rating Scale (MADRS) was 30.1.

High Accuracy Rate

In addition, in a subset of 115 participants (mean age, 36.2 years; 62.6% women), the researchers analyzed EEG data recorded 2 weeks after the treatment started .

All participants completed 8 weeks of open-label treatment with escitalopram, a selective serotonin reuptake inhibitor. The researchers chose escitalopram because of its favorable efficacy and side effect profile, said Farzan.

Patients were deemed responders if they demonstrated a 50% or greater reduction in MADRS score from baseline in the first 8 weeks of treatment. For responders, the mean MADRS score was 29.5 at baseline and 7.9 at week 8.

The main outcome was the ability of EEG data to predict treatment outcome, measured as accuracy, specificity, and sensitivity of the computer model at baseline and after the first 2 weeks of treatment.

The researchers used a well-established machine-learning algorithm known as a support vector machine. This "classifier" analyzed the large amount of data to detect patterns that predicted treatment response.

The authors noted that the model identified a number of features that had been previously shown to predict antidepressant response in MDD. These included parietal alpha measures and anterior cingulate cortex activity. The new tool identified additional features, such as the asymmetry in complexity of neural activity between brain hemispheres.

"A lot of previous studies focused on a particular feature, but here, we tried to combine all these diverse features and allow the machine learning to tell us which one is more predictive," said Farzan.

Results showed that the classifier could identify responders with an estimated accuracy of 79.2% (sensitivity, 67.3%; specificity, 91.0%) using only the baseline EEG data.

A Decade Away

There was no significant difference between study centers in terms of accuracy, which suggests a high degree of generalizability, said Farzan.

For the subset of participants for whom additional EEG data were recorded after 2 weeks of treatment, the accuracy increased to 82.4% (sensitivity, 79.2%; specificity, 85.5%).

The accuracy improved after 2 weeks because the machine learning tool takes into account how the brain is responding to the treatment, explained Farzan.

The findings suggest that this machine learning approach can speed identification of effective treatments for patients with MDD. Currently, physicians can't predict whether a particular medication will work, and they must use a trial-and-error approach.

Currently, remission rates are about 30% to 40% after a single medication trial and 50% to 55% after a second trial of a different antidepressant. Remission rates decline progressively with subsequent medication trials.

The authors note that the model predicted treatment outcome at 8 weeks, not individual patient response to escitalopram, which would require a randomized clinical trial.

The variability in patients' responses to escitalopram may partially explain the variability in treatment outcomes, said Farzan.

At this point, it is unclear whether this machine learning approach would also predict outcomes for other antidepressants and for nonpharmaceutical treatments, such as cognitive-behavioral therapy. "This is a question we hope to address with further research," said Farzan.

Future studies will also investigate whether predictive accuracy improves with the addition of other variables, such as depression severity, speech patterns, and sleep hygiene, she said.

It will take about a decade, said Farzan, before this tool will be ready for use in clinical practice.

Important First Step

Commenting for Medscape Medical News, Mark S. George, MD, professor of psychiatry, radiology, and neuroscience, founding director of the Center for Advanced Imaging Research, and director of the Brain Stimulation Laboratory, Medical University of South Carolina, Charleston, said the study is "an incremental and potentially quite an important step" toward enabling clinicians to choose appropriate drugs to treat depression.

Currently, there is no blood test or imaging test to diagnose depression, only clinical interviews, rating scales, and symptom assessment, said George. In addition, there's no way of predicting which treatment will be effective for any given patient, so physicians rely on a trial-and-error approach.

"There has always been this dream that we would have a scan or an EEG or something that would help us diagnose or parse out who should get what kind of treatment. We have talk therapy, different medications, and brain stimulation, so there's a big medicine cabinet, but the question is, who should get what when?" he said.

The beauty of using machine learning is that "it leaves no stone unturned" and that it can find something in a dataset that predicts or correlates with a response, said George.

"The question is whether what has been found is transferable to another dataset. Is it valid for all people with depression, or is it valid only for that dataset that you tried it on?" he said.

He noted that the study doesn't show whether the machine learning predictions are unique to escitalopram or would apply to any antidepressant.

"It would be helpful to clinicians if we had something that would tell us magically which drug works in which patients," he said.

He also noted that results may simply be an indicator of which patients had less severe baseline depression. However, he believes this tool goes further than current methods of determining depression severity in that it's a brain-based approach. "It gets us closer to biomarkers of actual brain activity," he said.

Researchers now have to determine what's unique about patients who respond to escitalopram, said George. "Where's the signal coming from in the brain? What about those people who don't respond? Do they have a different form of the illness?" he asked.

EEG is limited in that it only accesses signals from the surface of the brain, he added.

Some researchers are combining EEG data with brain imaging, which might improve prediction response, said George.

George noted other study limitations, including the fact that it was an open-label study and that there was no control group. However, the noted that the researchers "are not trying to predict one medication vs another."

CAN-BIND receives financial support from the Ontario Brain Institute, which is funded in part by the Ontario government. Additional funding was provided by the Canadian Institutes of Health Research, Lundbeck, Bristol-Myers Squibb, Pfizer, and Servier. Farzan has received funding from the Michael Smith Foundation for Health Research, the Natural Sciences and Engineering Research Council of Canada Discovery, and the Canadian Institutes of Health Research. George has disclosed no relevant financial relationships.

JAMA Netw Open. Published January 3, 2020. Full text

For more Medscape Psychiatry news, join us on Facebook and Twitter.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.