Machine learning to identify neural signature in people likely to respond to antidepressant medication

Machine learning to identify neural signature in people likely to respond to antidepressant medication

Researchers have discovered a neural signature that predicts whether individuals with depression are likely to benefit from sertraline, a commonly prescribed antidepressant medication. The findings, published in Nature Biotechnology, suggest that new machine learning techniques can identify complex patterns in a person's brain activity that correlate with meaningful clinical outcomes.

"There is a great need in psychiatry today for objective tests that can inform treatment and go beyond some of the limitations of our diagnostic system. Our findings are exciting because they reflect progress made toward this clinical goal, and they also show the potential of bringing sophisticated data analytic methods to psychiatry," explained senior author.

Major depression is one of the most common mental disorders, affecting about 7% of adults in the U.S. in 2017, but the symptoms experienced can vary from person to person. While some may experience many of the characteristic features--including persistent sad mood, feelings of hopelessness, loss of pleasure, and decreased energy--others may experience only a few. There are several evidence-based options available for treating depression, but determining which treatment is likely to work best for a specific person can be a matter of trial and error.

Previous research has suggested that specific components of brain activity, as measured by resting-state electroencephalography (EEG), could yield insight into how people will respond to certain treatments. However, researchers have yet to develop predictive models that can differentiate between response to antidepressant medication and response to placebo and that can also predict outcomes for individual patients. Both features are essential for the neural signature to have clinical relevance.
The researchers developed a new machine learning algorithm specialized for analyzing EEG data called SELSER (Sparse EEG Latent SpacE Regression). They hypothesized that this algorithm might be able to identify robust and reliable neural signatures of antidepressant treatment response.

The researchers used SELSER to analyze data from a large randomized clinical trial of the antidepressant medication sertraline, a widely available selective serotonin reuptake inhibitor (SSRI). As part of the study, participants with depression were randomly assigned to receive either sertraline or placebo for eight weeks. The researchers applied SELSER to participants' pre-treatment EEG data, examining whether the machine learning technique could produce a model that predicted participants' depressive symptoms after treatment.

SELSER was able to reliably predict individual patient response to sertraline based on a specific type of brain signal, known as alpha waves, recorded when participants had their eyes open. This EEG-based model outperformed conventional models that used either EEG data or other types of individual-level data, such as symptom severity and demographic characteristics. Analyses of independent data sets, using several complementary methods, suggested that the predictions made by SELSER may extend to broader clinical outcomes beyond sertraline response.

In one independent data set, the researchers found that the EEG-based SELSER model predicted greater improvement for participants who had shown partial response to at least one antidepressant medication compared with those who had not responded to two or more medications, in line with the patients' clinical outcomes. Another independent data set showed that participants who were predicted by SELSER to show little improvement with sertraline were more likely to respond to treatment involving a specific type of non-invasive brain stimulation called transcranial magnetic stimulation (in combination with psychotherapy).

Work is now underway to further replicate these findings in large, independent samples to determine the value of SELSER as a diagnostic tool. According to the authors, the present research highlights the potential of machine learning for advancing a personalized approach to treatment in depression.

"While work remains before the findings in our study are ready for routine clinical use, the fact that EEG is a low-cost and accessible tool makes the translation from research to clinical practice more possible in the near term. I hope our findings are part of a tipping point in the field with respect to the impact of machine learning and objective testing," the senior author concluded.