Can AI predict depression?

S

cientific progress is built on failure. From penicillin to the pacemaker, many of the greatest discoveries came about by accident. And to test any hypothesis, you must first take a leap in the dark. Alphabet’s X (formally Google X) spent three years developing prototype technologies to better understand depression.

The ultimate aim – to find a biomarker for the condition using machine learning technology – was unsuccessful. But other researchers may be able to build upon the work to find better diagnostic tools for mental health conditions in the future.

X is Alphabet’s so-called ‘Moonshot factory’ that has been likened to a real-life futuristic version of Willy Wonka’s chocolate factory. The top-secret organisation aims to solve humanity’s greatest problems by inventing radical new technologies. Past and present X projects include driverless cars (Waymo), delivery drones (Wing) and contact lenses that measure glucose in the tears of people with diabetes (Verily).

It is not surprising that X set its sights on solving mental illness. Poor mental health is a huge and growing problem. Approximately 322 million people suffer from depression globally according to the World Health Organization.

But the way doctors diagnose these conditions, using screening questionnaires, have limitations. Depression and anxiety can manifest differently in different people and patients can interpret the PHQ-9 test for depression or the GAD-7 test for anxiety in distinct ways.

There’s no objective measurement for diagnosing either condition as we do in other areas of health, such as diabetes and hypertension. X researchers wanted to change that; Project Amber was born. “Our journey started by asking the question: what if we could make brain waves as easy to measure and interpret as blood glucose and use them as an objective measurement of depression?” wrote Felten for a blog on X’s website in November 2020, just after Project Amber wrapped up.

The team knew from neuroscience studies that certain patterns of electrical activity in the brain are linked to symptoms of depression. For instance, anhedonia (not experiencing pleasure from the things you usually enjoy) is a hallmark sign of the condition.

Neuroscientists have found that people with depression consistently demonstrate lower electrical activity in the brain’s reward system in response to winning game-like tasks. This can be seen on an electroencephalogram (EEG) – a standard test used to find problems related to electrical activity of the brain. To the X team, this type of research offered a path to a potential biomarker for depression.

So, the team replicated the anhedonia studies in its own research with Greg Hajcak at Florida State University, US. But what they really wanted to develop was a similar test that could be used in the clinic, rather than a laboratory.

“For EEG to come out of the lab and into the real world as a mental health assessment tool in a primary care doctor’s office, counselling centre or psychiatric clinic, it needs to become more accessible and usable at scale,” Felten summed up in the piece.

Project Amber focused on making EEG data easier to collect and interpret, and how the technology could be better applied in the real world. The researchers built low-cost, portable prototypes of EEG headset devices.

The final prototype looks like a swim cap and is relatively easy to set up. The researchers say the device can be ready for measuring brain waves in less than three minutes. Like the advanced laboratory device, it has sensors to measure reward and cognitive function.

The most recent prototype from Project Amber. Credit: X

As well as making the device more accessible, the team then looked into tools that would make the data easier to interpret. Machine learning was an obvious choice. Project Amber collaborated with Google’s DeepMind AI experts. Obi Felten, X’s head of getting Moonshots ready for contact with the real world, explained that the researchers wanted to denoise the signals at scale and work out which aspects of the signal are most relevant in determining whether someone is depressed or not.

In a paper, that has not currently been peer-reviewed, the researchers suggest it is possible to extract EEG features that might suggest mental health conditions such as depression and anxiety using machine learning. The team said they were able to do this for an individual participant rather than a group, which implies clinically useful information from EEGs can be derived with much fewer data samples than is currently possible in the laboratory.

Other research groups have applied machine learning algorithms to EEG readings. An IBM team claimed to have developed an algorithm to classify seizures with 98.4% accuracy. And research published in Nature Biotechnology in February 2020 described using EEG and machine-learning algorithms to predict which patients would benefit from antidepressants.

Alongside device development and machine learning projects, the Amber team also conducted more than 250 interviews with potential users of the technology such as people with mental health problems and healthcare professionals. The researchers wanted to know what impact finding a more objective measure of depression and anxiety could have on patients’ lives.

The team did not manage to identify a specific biomarker for depression after three years of Project Amber research. But Felten summed up three main learning from the project. Firstly, the team concluded that mental health measurement remains an unsolved problem.

Even the surveys and scales that are available are underused by GPs and counsellors. The researchers suggest that any new measurement tool that is developed would need to create value for both the person struggling with their mental health and their healthcare professions.

Secondly, X believes that while an objective biomarker test for mental health would be undeniably useful (if one even exists), it should not necessarily replace the subjective assessment tools that ask people about their experiences and how they’re feeling. Combining subjective and objective metrics could be powerful. And this is a topic that demands further research.

Lastly, the team found patients and healthcare professionals had mixed opinions about using EEG as a diagnostic aid for depression. Some doctors felt a clinical interview was a better way of assessing someone’s symptoms. While some patients felt nervous about a machine being able to ‘label’ them as depressed. However, both groups of people felt using EEG for ongoing monitoring – to capture changes in mental health over time or in response to treatment – could have value.

Although the Amber team was not able to find a single biomarker for depression and anxiety, the research will live on. X has released its hardware and software designs in open source on GitHub. The Amber team has also donated 50 unused EEG headsets to Sapiens Lab, which runs the Human Brain Diversity Project, an initiative that supports mental health research in low-income countries.

“Addressing today’s challenges will require new partnerships between scientists, clinicians, technologists, policymakers and individuals with lived experience,” wrote Felten. “Now more than ever, more diverse voices, more multi-disciplinary collaboration, and more open sharing of knowledge are needed to unlock better mental health for everyone.”