These researchers want to stop treating depression with trial and error

These researchers want to stop treating depression with trial and error

A promising startup is using deep learning to tailor treatments to patients

Depression is the leading cause of disability in the world, and the World Health Organization estimates that over 4 percent of the world’s population suffers from depression – a whopping 322 million people.

Yet the largest clinical trial on depression to date, the Sequenced Treatment Alternatives to Relieve Depression (STAR*D) trial, found that only one third of patients get better after their first course of treatment, and that patients are less and less likely to improve as more treatments, including things like SSRIs and cognitive behavioral therapy, are tried.

In part, this bleak pattern of treatment response may be because depression is an incredibly heterogeneous disorder. To meet the criteria for depression set out in the Diagnostic and Statistical Manual (DSM) a patient could be sleeping too much or too little; they could have no appetite or gain a significant amount of weight. While it’s easy to imagine that these patients may require different courses of treatment, there are not currently effective tools to predict treatment response. Treating depression is typically a process of trial and error, and unlike other medical disciplines, advancements in precision psychiatry have been slow to develop.

Deep learning and depression

“It takes time – that’s two weeks, three weeks, four weeks, and they come back and they aren’t doing better yet and you need to try something else,” explains David Benrimoh, a psychiatry resident at McGill University and the chief executive and medical officer of aifred health, a Montreal-based startup that hopes to change this.

The aifred health team, originally comprised mostly of undergraduate students at McGill university but already vying with established companies, hopes that by applying the power of deep learning to assess biomarkers for depression, they can determine which treatment will be most effective for each individual patient, eliminating the often long period of trial and error between diagnosis and relief.

“I had a patient once – he had been treated for over a decade for depression," Benrimoh continued. 

"It took years to find a treatment that worked … and he had been through almost every conceivable treatment, before he got there, some of which have significant side effects. So it was years of suffering and finally we found something.. … Clearly there is something that would have worked for him, so wouldn’t it have been nice if we had some way to know what that would have been earlier on?”

Aifred is designed to complement existing diagnostic tools and help physicians choose the treatment that, based on the patient’s profile and how that treatment has performed in thousands of other patients, will provide the best outcome. Sonia Israel, a co-founder and director of scientific partnerships (and a colleague of mine at McGill), tells me that data points could include everything from genetic, metabolic, or pharmacokinetic profiles, to neuroimaging, to any sort of peripheral marker from blood or urine.

“The power of deep learning is that it doesn’t need every feature to be filled," she says. "It works relatively well with missing data.” But the more data available, the more precise the result.

A 'missing piece'

Israel and aifred's other founders, who had begun working together as part of a NeuroTechX chapter at McGill, came up with the idea last winter. After observing the gap in clinical practice either firsthand or through their coursework, the students decided to apply the skills they learned in their club, which aims to increase collaborative neurotechnology development, to address the issue.

“I really just got the sense that in psychiatry there was a missing piece," says Kelly Perlman, aifred health co-founder and director of research. "There was personalized medicine in all these other departments, but not psychiatry. People would finish lectures by, ‘in the future we hope that there is going to be a better solution to personalized medicine in psychiatry,' or, 'we need this,’ but it seemed like the problem was too big to tackle. That’s where we were like, maybe we can use deep learning.”

In just a year, aifred health is now competing on a global level, against teams from established companies, for the IBM Watson AI XPRIZE, a worldwide competition aimed to tackle the world’s grand challenges using artificial intelligence. The competition, with a total of $5 million in prizes, opened this year with 147 teams, and it runs until 2020. Aifred health is one of the 59 teams selected this year to move forward in the competition, and they were further selected as one of the top 10 teams developing the most innovative and important applications for artificial intelligence, and for one of the two Milestone prizes in the IBM Watson AI XPRIZE competition.

“It feels really awesome – I’m really proud of my team," Israel says. "I’m learning so much and we’re really doing, I think, something meaningful.”

While a full version isn’t ready yet, the company is starting to train a prototype of their model on clinical data sets from Canada and the United States. Once training is complete, the group hopes to release a limited test version and put aifred through some of the first-ever clinical trials for a decision aid based on artificial intelligence.

Using deep learning with psychiatric data sets is still a new field. But Benrimoh tells me that, recently, the team got a good sign for the future of aifred: their deep learning model was able to predict from a psychiatric data set whether someone had experienced suicidal thoughts with 98 percent accuracy – a level comparable to applications of deep learning in other fields. And it's bringing sorely needed progress toward precision psychiatry.

Comment Peer Commentary

We ask other scientists from our Consortium to respond to articles with commentary from their expert perspective.

If you want to know more, you can ask the scientists directly via Twitter.

David Haggerty


Indiana University School of Medicine

Trial and error are the best two words to describe psychiatry as a whole, in my opinion. While all of science is basically just trial and error, other fields – like cancer research – have really figured out the trial part, and minimized the error. 

To truly create progress in psychiatry, using network models, like deep and machine learning, to understand the nodes of data that we look at is the way forward. This approach starts to address the gaps in our research, which are the same gaps that the clinical population falls through, which is often a painful and frustrating experience. 

The group profiled here gets it right: trying to find the one golden biomarker in a haystack that is infinitely large hasn’t provided many dividends, and this shift to combining data and finding correlations that translate into improved clinical outcomes is deeply helpful. 

I hope to see labs and companies take it a step further and apply some of the same logic to things beyond detection, like medication management and mood tracking. Projects like this are the things that keep me engaged in the field, and give me hope for the future.


The Massive Science Consortium is group of scientists dedicated to making science accessible for everyone.

Apply to join