Machine Learning and Antidepressant Response

A few weeks ago (Feb. 22, 2017), a new study came out in JAMA Psychiatry, one of the top psychiatric journals, entitled "Reevaluating the Efficacy and Predictability of Antidepressant Treatments: A Symptom Clustering Approach." In this study, Dr. Adam M. Chekroud and colleagues used a newly popular statistical technique, machine learning, to ask the question, “which patient with major depressive disorder (MDD) will respond to which medication?”

Until now, this choice has been basically trial and error.

This report uses data from 9 large previously conducted antidepressant treatment studies, including 2 large federally funded studies STAR*D and CoMED, and 7 drug company studies, comparing different medicines. It’s part of an ongoing process of using machine learning and related approaches to identify more effective treatments (see work by the primary author, AM Chekroud, and https://www.spring.care)

This inventive "machine learning" approach is being used to identify patterns related to better response to treatment across many branches of medicine including psychiatry. Machine learning finds patterns to predict treatment response in the data itself, rather than relying on preconceptions of researchers or clinicians about which symptoms are most important or how they are interrelated. The technique searches within the data set often testing the connections using an "N minus 1" method, subtracting one subject on each analysis, and other times taking part of the data set, say, one half, and comparing the pattern observed in that part versus the other. Then these patterns can be tested on data from other studies, to see if they are still predictive.

Most researchers believe that major depression is "heterogeneous"—that is, it includes several different conditions with similar symptoms, not a single disorder; so there’s a significant value if it’s possible to disentangle subtypes, both for medicine response and for understanding the differing biology of these presumably different conditions.

Researchers in this recent study found 3 major clusters of symptoms (what they call core emotional; sleep (insomnia); and ‘atypical symptoms’). In general, they found that antidepressants worked better for the emotional symptoms than for the other 2 clusters of symptoms. Some of the medicines were more effective than others in some areas (high dose of the serotonin-norepinephrine reuptake inhibitor duloxetine vs. the SSRI escitalopram).

One useful result of this study is an online tool, https://www.spring.care/spring-assessment, a questionnaire that can be used by patients and doctors to help making decisions.

One intriguing possibility: The study (and possibly the new tool) could help with "personalized" treatment, matching antidepressant to a specific patient based on their symptoms.

It could potentially help in development of new drugs which could focus on these clusters of symptoms. And it could possibly guide studies of the biology of depression since different symptom clusters may reflect different abnormalities in the brains of people with depression—different brain circuits may be involved in these different symptom clusters.

The study does have some limits: different study designs (two of the studies didn’t use placebo; the ‘blinding’ differed between studies), and 7 of the 9 studies were with one drug (duloxetine). Also, the symptom clusters weren’t the same in all of the studies, with somewhat different items in different studies, which suggests that they aren’t ‘written in stone’ and we may not have the definitive understanding of them. Also the study only looked at medication studies not psychotherapy studies; some forms of psychotherapy (CBT, behavioral activation) may be useful for symptoms that don’t respond well to medicines. The 3 clusters should be tested in data from other studies, to see if they hold up.

The biggest limitation to me as a practicing psychiatrist is that it’s not clear how well these findings can be used in the care of individual patients—for one thing, studies often exclude people with medical or other problems like substance abuse, etc.

Questions I’d have:

· As a doctor, how many patients would I need to treat according to these predictions to get 1 additional patient better than if I choose an antidepressant medicine at random?

Artificial Intelligence Essential Reads

AI Chatbots for Mental Health: Opportunities and Limitations

Is "Digital Thought" the Next Cognitive Domain?

· As a patient, how much do I increase my likelihood of responding to a particular medicine if my doctor follows recommendations based on this study or the online tool (spring.care), vs. if he/she chooses a favorite medicine based on "his/her experience"?

But most interesting to me as a researcher is that my group has a number of datasets that this paradigm could be tested in. Our group in the Depression Evaluation Service at Columbia Psychiatry has done dozens of studies over several decades, generally using the same rating scales that this report investigates. We could use the same method to look at our data set. One bonus: we also have done MRI and other imaging (such as EEG) on many of our studies, and we have begun to use this same method of machine learning on THOSE samples. We are looking for brain circuits that machine learning could identify as most related to medication (and for that matter, placebo) response. Many times the things that the machine learning technique comes up with are entirely unexpected, so while such investigations are in one way 'fishing expeditions' in another way they represent one of the best methods of scientific exploration—searching into the unknown. Of course any such exploration can discover random noise...so any such findings urgently require "replication" in a different data set...from an entirely different study.