Artificial intelligence is quickly making its way into behavioral health care.
Still, many in the field are cautious about adopting the technology in practice. Dr. Megan Jones Bell, the clinical director of consumer and mental health at Google and author of the recent JMIR on AI, noted that while there are many use cases for the technology, clinicians should take their time easing into using the tool.
Even so, she said AI has the potential to reduce clinical burnout, improve nonclinical training programs and boost measurement-based care.
Behavioral Health Business sat down with Bell to discuss ethically implementing AI into practice, leveraging the technology at large organizations and the future of chatbots. This interview has been edited for length and clarity.
BHB: How can behavioral health providers meaningfully integrate AI and large language models (LLMs) into their practice?
Jones Bell: Leveraging this technology has enormous potential to improve some of the intractable problems we’ve had in the mental health sector, such as provider shortages and quality access problems.
It’s a great time for providers or clinicians to be educating themselves about generative AI, specifically to understand what’s different about generative AI from AI that has existed and was leveraged previously, and to start playing around with this, not using patient data, not using it in a clinical capacity, unless their organization is using, a HIPAA-compliant solution.
For an individual clinician, like in private practice, I’d say start educating yourself about being a savvy consumer of this technology, understanding ethical considerations, and how to identify and understand signals around clinical reliability and validity.
I like to think about the analogy of choosing mental health apps and how, initially, when they were developed, it was very confusing for clinicians and consumers alike to understand which ones were the good ones. What are the bad ones? How do I tell the difference? What are things I should look for? Many of those same principles around how you understand quality, safety and privacy in any kind of technology product apply here, too.
The main thing for clinicians is to proceed with caution but leverage it in nonclinical capacities.
How are you working with larger organizations to leverage AI and LLMs?
We’re working with them beyond mental health in a broader capacity around billing, note summarization and predictive models. There are a number of things that we can do with health care organizations on a central operational level.
In terms of mental health, I can give a few examples. We’ve worked with the state of Illinois via our cloud public sector team to help them build a new portal for accessing mental health care for children and families in the state. AI is part of that because it is about helping route and match people, understanding needs and triaging.
We’ve also used this with Google.org and some of our nonprofit partners. The examples I’ll give you are related to training and education or using non-clinicians to deliver crisis support.
We’ve worked with the Trevor Project to help build a conversation simulator. This is essentially a training tool that lets them train more crisis hotline volunteers to meet their quality benchmark for training.
Volunteers can practice with this generative AI-based conversation simulator. We took a similar approach with Reflux AI on a veteran peer support training tool.
This approach is exciting to see scale and represents a step in the right direction as we consider the workforce shortage in mental health. Shifting to non-clinicians for certain low-risk needs, like peer support or crisis interventions, presents a great opportunity.
You can also see how this could help therapists in training or therapists learning, and practicing evidence-based treatments.
What role can LLMs and AI have in diagnostics? Are these currently being used or future-looking?
We know that large language models can do a fairly decent job at estimating a mental health diagnosis.
They’re getting better at differential diagnosis, but that is a bit contingent on where there is a large corpus of information about a condition. For more common conditions, like major depressive disorder, models tend to be better at that because there’s just more on the internet.
There is more training data available publicly, but we know less about specific, smaller populations or lower-frequency disorders.
But that’s an area that I expect to improve because it’s the kind of thing that a large language model is going to be good at taking language, extracting the key elements and really matching to a condition. It’s also something that you can validate and train to be reliable.
Getting that technology to the right clinical benchmarks in a safe and kind of privacy-preserving way is feasible in terms of diagnostics.
Another area that we think is going to be important is the need for more measurement-based care. We aren’t doing a very good job tracking whether our treatments that we’re delivering at scale are working for people. We also tend to get diagnoses wrong and then not give people the treatment they need.
A big part of LLMs is the data they have been trained on. What should the behavioral health industry be mindful of when incorporating data sets?
One thing to consider is informed consent and where the data comes from. As we collect data from patients, whether in research or clinical practice, it’s key to make sure the patients are really informed about what their data will be used for now and in the future.
Our first job is to protect the privacy and confidentiality of patients, which means that training data needs to be fully de-identified, but it also does not need to come from real patients. We can generate synthetic data, which we can use to generate training data and work with real clinicians to label that data, provide feedback to models, and improve alignment with evidence-based practice.
There are alternatives that are being developed to simply just grabbing a big corpus of very sensitive data and training a model on it.
What is the potential role of chatbots in care?
The first thing that comes to mind when we talk about chatbots and mental health tends to be, are chatbots replacing a therapist? And so I’d say, like there are, that is certainly a path that people are exploring, companies, researchers, except developers. Although the thing I’ll say there is for a chatbot to be making a diagnosis, delivering treatment for a disorder — that’s a regulated product — that is software’s medical device, it needs to meet specific safety and clinical effectiveness benchmarks and have a regulatory authorization from the appropriate authority, depending on the country.
That is a longer game and that’s something to be taken seriously and with caution.
Separate from that, there are a number of opportunities to use chatbots in more of a wellness space related to promoting mental well-being and healthy behaviors, as well as what we talked about with education and training. But on the wellness side, like with mental health apps, that is where we are going to see more people come to the mental health market, and we believe that they can be more effective at helping coach people around their behavior change, whether that be related to sleep, activity or stress management.
Companies featured in this article:
