This is a presentation I gave at TransTech 2018 in the Bay Area - drawing from our research on mental health AI startups (https://emerj.com/ai-sector-overviews/diagnosing-and-treating-depression-with-ai-ml/).
The full video of this presentation is available online:
https://www.youtube.com/watch?v=CvrqoPpYF94
Integration and Automation in Practice: CI/CD in Mule Integration and Automat...
AI in Mental Health and Wellbeing - Current Applications and Trends
1. AI in Mental Health and
Wellbeing - Current
Applications and Trends
Daniel Faggella, CEO at Emerj
2. Presentation Outline
● Background in brief
● The unmet needs of mental health
● Category 1: Conversational Interfaces
○ Purported benefits / limitations
● Category 2: Behavioral Pattern Recognition
○ Purported benefits / limitations
● Believable benefits and predictions
● Risks to consider
● Concluding takeaways on this sector
3. Emerj (formerly
TechEmergence)
● Emerj is a market research firm
focused the implications of AI for
business / gvmnt leaders
● Organizations planning very
expensive technology initiatives
use us to support critical decisions
● We work with clients as large as the
World Bank, and startups who’ve
only raised $10MM
4. Emerj (formerly
TechEmergence)
● We examine AI across three
dimensions:
○ Applications (Possibilities)
○ Implications (Probabilities)
○ Strategy (Plans)
● Diagnostics, health, and mental
health have been a focus for us for
over two years, and this
presentation summarizes our
findings at the intersections of AI
and mental health
5. Note
● This presentation is remarkably short (20
minutes), so I will not be able to talk about the
full breadth of AI solutions
● My goal is to highlight the critical points
● If you’re interested in our deeper research into
individual mental health solutions, please see
the last slide in this presentation, which lists
our deeper reports and app comparisons
7. By the Numbers
● As of 2015, of the overall NIH funding for FY 2015: $30 billion, less
than 10% went to mental health and addiction-related programs
● More than 10% of lost years of healthy life and over 30% of all years
lived with disability; WHO, 2001
● Depression is estimated to cause 200 million lost workdays each
year at a cost to US employers of $17 to $44 billion
● Also mental illness it is hard to measure and quantify (unlike, say,
pneumonia, where we can be sure when you have it, and when you
don’t)
9. Vendors
● Vendors: Wysa, Ginger.io, Woebot, etc
● Purported benefits:
○ Leading indicators for therapists/Drs
○ Healthy tips / reminders via chat (with
machines and humans
○ Connection to a real human coach/therapist
● Limitations:
○ Chatbots still somewhat nascent
○ Chat is only one source of data
11. Vendors
● Vendors: Ginger.io, Sunrise Health (now
“Marigold”), Mindstrong, HelloJoy
● Purported benefits:
○ Leading indicators for therapists/Drs
○ On-demand connection to professionals
○ Extracting data from multiple sources
(text context in and out of app, location,
etc)
● Limitations:
○ Measuring wellbeing = hard
12. Vendors
● Data sources purportedly tracked
○ Exercise / Sleep
○ Location / movement
○ Self-report assessments (Daily, weekly)
○ Contents of text messages
○ Phone use / activity
● There are questions around both the
reliability of self report data (exercise), and
the conclusiveness of phone use data
● There are privacy concerns with this
sensitive personal data
14. ● We suspect that the core value proposition of the mental health
space lies in finding patterns across new data streams
○ Sources: Phone activity, location, app use, content of texts,
estimated hours of sleep, etc
■ Correlating new streams of real-time data (not
self-reported, but extracted from the world) to proxies for
user wellbeing (psychiatrist visits, self-reported wellbeing,
suicidal thought frequency, etc) seems promising, and AI
can be part of this mix
● We imagine a future where this proxy data can give a much more
accurate estimate of the wellbeing of a patient
15. ● We suspect that it may be years before AI can realistically suggest
actions or behavior change for the human
○ These suggestions require much more context than a chatbot
or iPhone can tell us (from childhood trauma, to fears, to other
medical issues, etc), and probably should stay in the purview
of doctors
16. ● Replicating human interaction is challenging, and seems
potentially dangerous
○ Text chat is a limited channel, and conversational AI is
exceedingly limited
■ There are a number of examples online of very disjointed or
nonsensical conversations with the bots
● Text is one potential source of feedback and data from the user,
and it could be a critical one for coaxing out high-level risks, like:
○ Asking the patient if they’re okay (if other data signals indicate
that they might be at risk)
● We feel less optimistic about chatbots encouraging behavior
change, or humans being accountable to chatbot suggestions
17. ● Conversational interfaces might be able to reach:
○ People unable to afford therapy
○ People who might be too shy or ashamed to try therapy
○ People in rural or remote areas with no therapist access
● While the opportunity to help those who cannot or choose not to
use traditional therapy, it is unclear what the best mental health
intervention is for this group of people
● It is blatantly evident that more research is needed in order to
discern the impact of different kinds of digital treatment (or no
treatment)
18. ● Good news: Startups in this sector seem to understand that the
purpose of their applications are to keep patients safe, and hand
them off to doctors
● The value proposition of Mindstrong (below) is representative of
many other firms - with “preemptive” insight being emphasized:
20. Business Model
● Business models create incentives:
○ Monthly fee:
■ Encourage forever use, whether or not the user needs it
○ Pay per referral to therapist:
■ Refer people to therapists, whether or not the user needs it
○ Pay per chat conversations with a therapist or “coach”:
■ Hire inexperienced / inexpensive “coaches” onto the
platform
■ Refer people to chat therapists, whether or not it’s actually
necessary or preferable
21. Business Model
● Startups have to find a way to be profitable, but it’s unclear
whether the for-profit incentives might bend the products in a
direction that doesn’t serve users
● This isn’t necessarily the fault of startups, but it is a problem that
will have to be worked out, and something that users should be
aware of
22. Long-Term Effects
● More scientific testing and trials are required here
○ We need long-term double-blind studies to assess the actual
value (and not research conducted by founders themselves)
● Testing would need to be done for any of the target groups or
target conditions
○ People in rural / remote places
○ Men / women
○ Age of patient
○ Depression / bipolar / anxiety / other
24. ● AI in mental health and wellness is likely to unearth solid
predictors of wellbeing and mental health risk - but we have no
robust proof at present
● It will take much longer to achieve a level whereby AI can suggest
actions and behaviors in order to tangibly improve wellbeing or
reduce risk
○ Even in 2-3 years, we don’t expect that these “suggestions”
will get beyond very generic and basic advice (i.e. “get 8 hours
of sleep” or “stay hydrated” or “do your breathing exercise”)
25. That’s All, Folks
The end.
I’ve included a list of some of the resources from this
presentation on my final “References” slide.
Feel free to send along an email with questions, or for a
copy of the slides:
dan@emerj.com
Twitter - @danfaggella
emerj.com