With $100 million in investment backing, Brightside Health, a San Francisco-based telemental health provider, serves people with clinical depression, anxiety, and other mild to severe mood disorders, including those at high risk for suicide. Mimi Winsberg, MD, the company’s chief medical officer, recently spoke with Healthcare innovation about the company’s concept of “precision prescribing” and leveraging data to optimize treatment plans, as well as using AI to help predict mental health crises.
Healthcare innovation: I want to ask you about some research recently published in JMIR Mental Health which analyzes the performance of large language models in predicting mental health crisis episodes. Before we do that, could you help us set the stage by talking a little about your background and Brightside Health’s approach?
Winsberg: I am a Stanford-trained psychiatrist and my experience during my fellowship was in the management of bipolar disorder. I’ve been in the digital health space for about 10 years. What I observed, certainly in treating patients with bipolar disorder over the years, along with other psychiatric conditions, is that it was very helpful for patients to track their symptoms, and we could be much more successful in predicting their episodes. If we had a good record. of their symptoms. 25 years ago, patients did this with pencil and paper, and then with the advent of the digital health movement, it was very important to me that we be able to use some of the technological tools that we have at our disposal. willingness to do things like remote symptom monitoring and even treatment prediction based on symptom cluster analysis.
Not all antidepressants are created equal, but many times in the mental health arena, selecting an antidepressant is actually a bit of a guess and check process for many providers. What I hoped to do with some of the technological tools we had at our disposal was to create a database and take a more informed approach to treatment selection that takes into account everything from the patient’s current presentation of symptoms to things like Previous drug trials, history and so on. That’s what we built at Brightside and it’s built into the backbone of our digital health platform that Brad Kittredge, our CEO, and Jeremy Barth, our CTO, created seven years ago.
HCI: Does that involve looking at not just how this individual patient has responded to, say, different medications, but also looking at the entire database and seeing how people respond and clusters of symptoms and things like that?
Winsberg: That’s how it is. It is not just based on the individual. It is based largely on the published literature that exists and also on a very robust database that is probably unparalleled in the sense that we have treated over 200,000 patients. We can observe the patient’s attributes, the presentation of symptoms, and the treatments and results. We can say, ‘Who else do we have who looks a lot like you and how did they do with this treatment?’ And we can make some predictions accordingly. This is one way to approach treatment selection. We have published extensively in peer-reviewed journals about the success of this model. This is all exciting, because it really helps move the needle in a field that, I would say, has been less data-rigorous than other fields of medicine.
HCI: Especially when the pandemic hit, there was a huge growth in the number of telemental health providers. How do you stand out in that field, among patients, payers and provider groups?
Winsberg: Telemedicine 1.0 puts a doctor and a patient in a video interface. That can solve a lot of access problems, because you’re no longer dependent on those two people being geographically located. Allows you to leverage providers in an area to provide services in an area that may have a shortage of providers. But that’s just the beginning of what telemedicine can do. As you said, a number of companies emerged from the pandemic that intended to solve the access problem. At Brightside we see that as something that is at stake. We existed before the pandemic and telemedicine was just one of our goals. What we really tried to do was take a more precise, quality approach to care.
So in terms of differentiators, one is the notion of precise prescribing, which is our proprietary language, so to speak, around the data systems that we use to make treatment selection recommendations. It is clinical decision support, so a machine does not decide which treatment is best. You tell your psychiatrist, who then uses that information to better inform your choice. But that precision prescription engine is proprietary to Brightside and is definitely a differentiator, as are many of the other AI tools we’re actively deploying and publishing. In terms of health systems that partner with us, we believe it is important to showcase our work and publish it in peer-reviewed journals where the data can be objectively examined and evaluated by anyone interested.
HCI: What does the payments landscape look like? Does Brightside have partnerships with health plans or health system organizations?
Winsberg: We have national contracts with many payment systems and we obtain those contracts by demonstrating the quality of our work. They have access to the data so they can look at our results with a very informed lens and have obviously determined that our results meet or exceed the quality that they would expect to be able to afford them.
HCI: Do you have any contracts with Medicaid managed care organizations?
Winsberg: We started with commercial payers and then launched with Medicare, and now we’re also rolling out Medicaid nationally.
HCI: Let me ask you about this research recently published in JMIR Mental Health. Could you talk to us about how it was carried out and what it demonstrated about large language models and their implications?
Winsberg: Large language models can digest a large amount of textual information quite quickly and synthesize it. So when a patient comes to our website and starts registering for services, we have a question for everyone that says, tell us why you’re here. Tell us what you are feeling and experiencing. And people write anywhere from one sentence to many paragraphs about why they seek care. The provider typically reviews that response, along with other structured data.
In this experiment we took the information entered by patients and completely stripped it of any identifying information, and showed it to a set of experts who reviewed the text data, along with information about whether the patient had previously had an attempt to suicide. . Then, apart from that, we fed that information into a large language model, ChatGPT 4, and asked both parties (the experts and ChatGPT 4) to predict whether they thought the patient was likely to have a suicidal crisis during the course of their attention.
What we found was that the language model was close to the same accuracy and predictive ability as trained psychologists and psychiatrists. Now, the caveat in all of this is that providers are far from perfect in their predictions, so just because I’m a psychiatrist doesn’t mean I’m going to predict this, but that’s the best we have right now. It raises a broader philosophical question: When AI is implemented, do you expect it to be as good as humans? Do you expect it to surpass humans? For example, with self-driving cars, you have to be better than humans to want to implement it, right? That’s why we took the same approach in medicine when we started training these tools. To implement them widely, we would need them to be much better than humans, but what we are seeing, at least in this example, is that we can do it as well as humans. What we found is that for a human to do this task is very laborious and also very emotionally draining, so having an automatic alert that you might not have had otherwise can be very helpful.
HCI: In this particular use case, if you could get the tool to be really very accurate and that triggered an alert, how would that change the care plan?
Winsberg: We make a large selection of patients based on the information we obtain about them at the time of admission for treatment selection purposes. For example, we have a program called crisis care, which is for patients who are at high risk for suicide, and it is a particular therapy program that is based on collaborative assessment and management of suicide. When patients enroll in this program, they have more frequent and longer sessions with their therapists that specifically discuss suicide risk and manage reasons for wanting to live, reasons for wanting to die, etc. So if we found out that a patient was identified as high risk, they would be referred to a higher acuity program.
Similarly, there are certain pharmacological strategies you could employ with higher-risk patients. You can move them to a level two treatment selection, rather than starting with a level one.
HCI: So, in summary, are you saying that the research shows that these tools are promising, but they are not yet ready for implementation?
Winsberg: What I’m saying is that we still keep humans informed at every step. We view these tools largely as co-pilots. They are more like a GPS than a self-driving car.
Another example of an AI tool we are implementing is a scribe, a tool that can transcribe a session and then generate an interim note for a provider.
Yet another example of AI is that we also provide our providers with information about care. There are many elements of the chart that you should review before speaking with the patient or while speaking with the patient. Depending on how extensive a patient’s history is, it’s nice to have a tool that can summarize various aspects of care for you. And LLMs are pretty good at this. So we’re just scratching the surface in terms of the ways that AI can improve the quality of care delivery, as well as reduce the provider burnout that we’re seeing with great frequency across the country right now and everywhere. the specialties.
