A human-first approach to A.I. could transform digital therapy

Depression affects an estimated 300 million people around the world, according to the World Health Organization. While seeking care for mental health disorders has become more common and less stigmatized in many parts of the world, actually receiving care can be difficult. It can be expensive or inaccessible — especially during a pandemic.

Video conferencing, telephone calls, and even text-messaging with a therapist have become more common in the past year but SilverCloud Health, an internet-based cognitive-behavioral therapy (iCBT) platform, offers a different solution. Their DIY health platforms enable patients seeking mental health tools to progress through care modules at their own pace with remote clinicians or coaches monitoring patient progress.

Now, thanks to a research partnership with Microsoft Research Cambridge, SilverCloud is exploring how machine learning can help clinicians understand the needs of their patients better than ever. The patterns found in the de-identified data of over 50,000 users could help create more effective and personalized mental health treatments.

What is SilverCloud? — While apps like Headspace or Moodfit guide you through mindfulness or mental health exercises, SilverCloud is more similar to services like wayForward. It offers cognitive behavioral therapy (CBT) in the form of internet-delivered programs.

Derek Richards, SilverCloud's chief science officer, tells Inverse that this platform is designed to offer affordable and accessible tools to people navigating mental health care.

"[This] is largely self-administered and then a human in the loop, the coach or supporter, is supporting asynchronously," says Richards. Like a teacher reviewing online assignments, through SilverCloud, coaches or clinicians can log on and see what progress a patient has made and where they might be getting stuck.

The future of A.I. and digital therapy

While Richards says that iCBT does have substantial clinical research to support its effectiveness, that doesn't mean that all users necessarily benefit in the same way.

To better figure out how people were interacting with their platform and what they were getting out of it, SilverCloud partnered with Microsoft Research to design a machine-learning algorithm to crunch over 3 million points of engagement data from SilverCloud users and uncover what makes them tick. The study's results were published this July in the journal JAMA.

Over the course of this two-year study, the team trained a probabilistic machine learning algorithm to evaluate which modules users interacted with most (e.g. building a "worry tree" or completing quizzes about introspection) and how long they showed consistent interaction. While the team report that all users experienced some degree of clinical improvement, the improvement varied across five different engagement subtypes:

• Low engagers
• Late engagers
• High engagers with rapid disengagement
• High engagers with a moderate decrease in engagement
• Highest engagers

Danielle Belgrave, a lead author on the study and principal researcher at Microsoft Research, tells Inverse that these results demonstrated an opportunity for better-personalized forms of care.

"Through understanding these distinctions in how people engage, we may be able to better personalize interventions in the future by recommending resources that lead to better engagement, and therefore improve symptoms more effectively," explains Belgrave.

Richards says that this information can be shared with clinicians or coaches supporting users on SilverCloud in order to help them better guide those who fall into these categories of engagement.

Beware of your bias — When it comes to distilling the complexity of mental health to variables and transcribing those to an algorithm, Richards admits that there are bound to be some limitations.

"Human behavior is extremely complex and the whole invention of using A.I. in healthcare is still very much debated," says Richards. "There has been a lot of noise around it [because] it's the new shiny thing, but really my view personally is we should use our A.I. intelligently. It should be I.A. — intelligent use of artificial means to enhance treatment."

And while Richards says they have yet to come across problems of algorithmic bias in their research, especially as demographic information like sex, race, or income was not included in this study, he tells Inverse that this is a dangerous precipice the team intends to approach with caution.

"... my view personally is we should use our A.I. intelligently. It should be I.A. — intelligent use of artificial means to enhance treatment."

Belgrave agrees and says they hope to address these kinds of problems early by designing machine learning systems that are human-first.

"[T]o achieve good A.I. that works for the person, we need to build A.I. tools that are based on a strong understanding of the person, and are carefully designed and developed for their benefit," Belgrave says. "A foregrounding of such a human-centric approach to A.I. development will be crucial to the success and future evolution of digital therapy."

Even with a more human-centered A.I. for iCBT, Richards says that SilverCloud is not attempting to replace teletherapy or therapy options. Instead, it should act as a complement to them: Both 1-on-1 support from a therapist and curated tools from a service like SilverCloud have a place in the overall improvement of a patients' mental health, Richards says.

When can you visit your A.I. therapist? — Going forward, Richards says they plan to begin incorporating more variables into their machine learning model, such as information about users' sex, age, and history of mental illness, to build a richer picture of individuals.

With more refined data, it may be possible for an A.I.-driven system like this to even make predictions about what a user's clinical outcome might be based on their current path of engagement. This information would be communicated to a supporting clinician or caregiver who could then help correct their course towards a more beneficial outcome, even if that path took them off the service altogether and into in-person therapy instead.

Belgrave tells Inverse that she hopes innovations like these can be used to bring important services like mental health care to communities that need it most.

"Improving access to health and medical services is crucial to reducing health inequity globally," says Belgrave. "Concepts like internet-based therapy are just one of a number of solutions that can improve access to care for people living in poverty or in underserved areas."