Using Real-World Data to Understand Treatment Resistant Depression in Mental Health
Speaker: Carl Marci, MD
On today’s episode of the MAPS “Elevate” podcast, we’re discussing the use of Real-World Evidence to understand and possibly intercede in treatment-resistant depression and why is it especially difficult to develop treatments against resistant depression using the traditional structure of R&D-led clinical trials.
Following is an automated transcription provided by otter.ai. Please excuse inaccuracies.
Garth Sundem 00:00
Welcome to this episode of the Medical Affairs Professional Society podcast series: “Elevate”. I’m your host Garth Sundem, Communications Director at MAPS. And today we’re speaking about the use of Real World Evidence to understand and possibly intercede in treatment resistant depression. Joining us is Carl Marci MD, Chief Psychiatrist and Managing Director of the Mental Health and Neuroscience Specialty Area at OM1. This episode is sponsored by OM1, a leading health outcomes and registries company focused on the measurement comparison and prediction of treatment outcomes. So first of all, Dr. Marci, welcome.
Carl Marci 00:42
Thank you, happy to be here.
Garth Sundem 00:45
And let’s start with the need for real world data are our W E. in mental health. Why is our web so important specifically in mental health?
Carl Marci 00:57
Sure. I mean, I think real world evidence has a role to play across the entire healthcare spectrum. But as a psychiatrist, if you would have told me 20 years ago, as I was finishing my residency training that 20 years later, I wouldn’t be practicing pretty much the same way I did, then I would have said no way, right? We are learning too much about the brain, we have new medications and new treatments coming all the time. We’re investing heavily in research, it’s not possible. Well, I will tell you, through my small clinical practice, it feels pretty much the same as it did 20 years ago. And when you compare that to say, what’s happening in oncology, or cardiology, where we’re making huge strides in reducing morbidity and mortality, it feels like a step backwards. Because the opposite is happening in mental health. Right, even before the pandemic that accelerated through the pandemic. You know, we have high single digit, low double digit increases in anxiety, depression, PTSD, ADHD, substance abuse, and sadly, suicide. So you know, Houston, we have a problem. So I think there’s a real opportunity to change the game. And I’m happy to talk more about that if you’re interested.
Garth Sundem 02:07
Well, I, am, you know, I know there’s no capital A answer to anything we’re going to talk about with mental health or treatment resistant depression today. But you know, in your mind is RW E, not the capital A answer. But is this a strategy that can make inroads against mental health challenges?
Carl Marci 02:24
Yeah, absolutely. I mean, first search, we have to start about, you know, talking about what is what is the opportunity, so the opportunity is, with with the way in the way home one is approaching it, is we combine healthcare data from multitude of sources, right, electronic health records, importantly, because that has both structured and importantly, unstructured, you know, clinical notes from clinicians. And we combine that with health claims, and pharmaceutical claims, and then other types of data, sometimes social determinants, financial information, and we organize that in the cloud. And we mine it for insights, that’s really the goal. When I came on board, the opportunity was a new specialty network and mental health with 3 million patients. So 3 million patients across 9000 clinicians and 2500 clinics in all 50 states. It’s an enormous data set. And so that’s the opportunity how do we mind that for insights that can help patients with mental health get get diagnosed sooner, and get the what we’d like to say the right treatment to the right patient at the right time?
Garth Sundem 03:33
Okay, so let’s get back later into what we can find from all of this structured and unstructured data. But before we do that, you know, I’d like to learn more about treatment resistant depression specifically, it’s almost like we’re going to talk about this as a case study for RW E or, or how it can help. So give us give us the background. What do we mean specifically? And maybe what is your experience with treatment resistant? Depression?
Carl Marci 04:00
Yeah, the reason we’re talking about it is because it’s a very hot topic, and it’s a hot topic because there’s some some new medications and new treatments coming online, through you know, the psilocybin ones. Ketamine has gotten a lot of attention esketamine transmitting transcranial magnetic stimulation, right? All kinds of opportunities, right. So we think there’s a real change, see change happening in in depression, treatment resistant depression is kind of self explanatory, right? It’s patients with depression, who have failed therapeutic interventions more than once. And that’s sort of what it is. And depending on who you ask, or where you look, it’s anywhere from five to 50% of depressed patients fall into this category. Now, that’s a huge range. You may be asking, you know, Dr. Marci, why is it such a big range?
Garth Sundem 04:53
Dr. Marci, why is this such a big range?
Carl Marci 04:55
Excellent question. So, what we would part of the reason is We’re not very good as a field in defining what it is. Right? So one of the most common definitions, for example, comes from the FDA. It’s known as the regulatory definition. And it’s essentially the failure of two antidepressant trials of adequate dose and duration. Now, that sounds simple, right? But if you think about is, what is it? What is adequate? Right? Is it six weeks? Is it eight weeks? What’s an adequate dose? Is it 20 milligrams of Prozac or 60 milligrams of Prozac. And then what happens in between when do I know when a patient has stopped one line of therapy, and started another and so what we know from our data and other data is if you just tweak those parameters a little bit, you get a very different population. Right. So that’s not really a very good definition. And so we have a problem with defining what it is, which makes it hard to actually address.
Garth Sundem 05:51
So in terms of real world evidence, you know, it sounds like we have existing drugs, we have our products of the world, we have a lot in that generation. But then we have a lot of new things coming online. Is our web most useful. What in now and comparing the results of the existing generation of treatments? Or are we starting to look at RW E to evaluate emerging treatments?
Carl Marci 06:17
Yeah, that’s a great question. So I think it’s a little bit of all of the above. But But first, we have to sort of get consensus on a definition. Yeah. So Right. So one of the things we’re working really hard with partners, is to compare, you know, a standard version of the regulatory definition with actual clinical documentation of the phrase treatment resistant depression is by clinicians in the chart, right. So we, they’ve been labeled, right, and comparing that to some existing data science models. And we’re building our own machine learning and data science model. And when you when you can compare all these three things, you know, there’s going to be a lot of learning. And I’m very excited over the next coming weeks and months to share some of those insights. But what I’ll tell you is that you do not get the same population, when you compare these things. And so that should inform the discussion about what treatment resistant depression is, step one, step two is then Okay, once you have a model, and you know, we’re going to make the case that, of course, our artificial intelligence machine learning model is the best. Because it’s drawing from so many patients, you can then apply that to do a number of things. First, find patients who aren’t labeled to treatment resistant depression, and identify them so that they get the treatment they need. Second, we can use that to predict who has at high risk for going on to develop treatment resistant depression. And then thirdly, we can begin to have a conversation about how to optimize treatments.
Garth Sundem 07:53
Boy, I was being really myopic in my understanding of our web in this space, I was thinking, Oh, treatments, treatments, treatments. And you’re saying that, you know, our web is a gonna help you define the condition, you know, be can be predictive for what population is at risk for this condition, and then only see our, you know, comes my narrow view of using our web to see how drugs are working. Fascinating use Cool.
Carl Marci 08:23
Let’s build on that a little bit. How are we able to do that, right? This is not only do we have this massive data set, but we have this this incredible tool, it’s called Patient finder. And it’s an artificial intelligence software tool that identifies unique patterns of information in the health records of known patients with the diagnosis. So we go in, we identify patients where an expert clinician has said you have treatment resistant depression, we then build a model based on that, and we look for patterns. And then we can take that and apply the model to any population. Patients who don’t have the diagnosis. And if they meet that pattern, either a little bit, a moderate amount or a lot that tells us something about who those patients are, and how they may go on to either respond or not respond to different treatments.
Garth Sundem 09:11
It’s interesting. My wife’s a psychologist and I see her do a lot of these screenings. And it’s almost like doing doing a doing a screening at a at a mass scale. Where where you can look out into a population and mass screen that population for risk factors and be predictive in the population that you’d be going to with what further Outreach?
Carl Marci 09:30
Well, first of all, you said a critical word there, which is scale, right? The ability to sort of scale these kinds of screening tools, because we’ve digitized so many health records in this country, which is really how this is possible. Right. You know, thanks to a lot of government incentives, intervention and funding. You know, we’ve essentially digitize our health lives that allows us to really scale and go look for things. So yeah, you can you can then begin to apply that at the individual level. Oh, so we can look forward to a day, hopefully in the not so distant future, where, you know, Mrs. Jones comes to see me and I put Mrs. Jones few parameters of Mrs. Jones into a computer, you know, her comorbidities, some of her past treatments, maybe age and some demographic information. And I can compare Mrs. Jones to 100,000 300,000 patients like her, and it’s going to spit out a report. And that report is going to tell me how likely she is to go on to have treatment resistant depression, or what types of medication she might respond to. Now, there’s some work to do to get from here to there, but I think we’re on the right path.
Garth Sundem 10:41
Okay, so to put a peg in at one one use this approach would be screening at scale. Is that just really screening at scale?
Carl Marci 10:48
For sure. Okay, cool. And then predicting at scale? And then predicting at scale? And then informing at scale? Right, which is what what are the best treatments? And we can talk about that?,
Garth Sundem 10:59
Well, yes, let’s, uh, okay, so informing it’s getting, we’re getting finally into the area that I thought was the only use of our web, but maybe let’s go into treatment. And you know, in treatment, resistant depression seems like a good place to stay. But what are we talking with our web in treatment? For?
Carl Marci 11:17
Sure. So So now, we assume a world where we have a model that is robust and accurate identifying people with treatment resistant depression, the next thing you need to do is have outcomes, okay, because we need to know who’s responding to what. And one of the challenges in real world evidence and real world data is gaps in the data set, right? Because the real world is messy, right. And sometimes we don’t have a PHQ nine, which is a standardized scale for depression, commonly used for screening on every patient every month, like we would like. So this is the second use of real world data and real world evidence, and artificial intelligence where we can apply a similar approach that we did the TRD. But now, instead of looking at just the structured health record and the history of that patient, we’re going to look at a single encounter. And we’re going to use a computer to read the notes from the clinician, and we’re going to generate a PHQ nine score, which we call an E PHQ. Nine, because it’s an computerized estimate. And we can fill in a huge number of gaps with that kind of technology.
Garth Sundem 12:30
Okay, take me back just as well. So first of all, we’re going now into the unstructured data. And we’re trying to, and we’re trying to say how are patients doing? Can you remind me so that the PHQ nine is a standard depression score? So does it say I mean, improvement on the on the PHQ, nine implies that your depression is becoming more managed, and you can like, correct. So you’re looking for shifts in score over time.
Carl Marci 12:57
So the PHQ nine is a noted item scale, that’s where the pain comes from. And it’s modeled after the Diagnostic and Statistical, Statistical Manual. And it’s well validated. And it’s actually fairly simple, it takes five or six minutes to complete. That’s why it’s so widely used, and you get a score of essentially zero to 27. And then the higher that score, the more depressed you are. So you can create different cut offs. And imagine if you if you’re doing this sort of, say, every month or every couple of months, you can see disease progression. And then you can look at our large data set and say, Okay, these are the groups of patients who are responding to medication a, these are the groups are responding to medication B, these are let’s, let’s do that same technique, and look and see, well, what do they have in common? What are the what are the signals that we can see in their health record and their data, that’s going to show us that they’re able to respond to that treatment. And then we can do that for medication B, and medication C and medications D. And then I can take Mrs. Jones and say, okay, Mrs. Jones, you have a higher probability of responding to medication, B, then C, or a, I’m going to start with medication B, I can’t do that today. As a psychiatrist, I have to just sort of use my you know, my own history and a little bit of guidance from the literature to make a recommendation, I think it’d be a tremendous benefit. If we could look at a report and say, you know, Mrs. Jones, you’ve got a 50% or 2x Probability of responding to this medication. And that one, would you like to try it?
Garth Sundem 14:27
At the beginning of this conversation, you brought up oncology and it almost seems like Prozac is chemotherapy, and you’re talking about going to, like a genomic profiling approach. And saying, here are the here are the factors of your experience that can predict almost like a fingerprint, which treatment even modality or drug or whatever is going to best, you know, have the highest likelihood of affecting you so we’re becoming more personalized through this RW e approach?
Carl Marci 14:57
So Garth, is you’re spot on with the thing fingerprint, except where oncology is has tissue because there’s an actual tumor, or blood cells line that you can actually extract and put under a microscope and put all kinds of tests. That’s the genomic part. We’re using fino omics or phenotypes. So we’re looking at expressions of behavior that patients have that clinicians are capturing, either in structured or unstructured data that really reflect who you are, clinically as a depressed patient. And that’s where the fingerprint comes in. So we’re essentially doing the same thing that we did in oncology. You’re absolutely right. Except instead of with tissue and biomarkers, we’re using phenotypic markers and behavioral markers, and clinical rating scales, tickets at the same place.
Garth Sundem 15:46
Okay, so you said, you know, your PHQ, nine scores, your your inventory, or I’m sure I’m misusing that word, but your PHQ nine scores. Can can imply how depressed you are and change over time can imply getting better or getting worse. But not everyone has that. So that’s a gap, right? That’s a gap in the data set. And so we hear that term, so much gaps in Medical Affairs, what what do you mean by that? And how does our web fill in these gaps?
Carl Marci 16:21
Sure, so So the gaps are just what it sounds like, there, it’s missing data, right? It’s information about a particular patient or group of patients that we wish we had, but we don’t for some reason, it might be that the data was never collected, it might be that the data got corrupted, or might be that, you know, for whatever reason, it didn’t get into the database. So there’s any number of reasons why we might have missing data. And the problem with missing data is it creates a statistical challenge. So this is where using the machine learning approach that I was describing earlier, we take a group of patients who we have PHQ, nine scores for, we take the encounter date, and and their clinical record their clinicians notes. Now those notes have to be of a certain detail in length, or they have to be a certain order of complexity. And then the computer reads them, and looks for patterns in the note associated with that encounter, and that score. And we do that enough times where we get that fingerprint you were describing earlier. And then we can apply the same approach to look for patterns in notes that don’t have the PHQ nine, and assign an estimated score. And what that allows us to do is literally 10x, the number of PHQ nines we have. So if we have 500,000, PHQ, nine, on our five to 10 million patients, that 500,000 goes to 5 million. And that’s a lot of gaps to fill. And that’s what allows us to then look at disease progression and change over time and begin to predict who’s going to respond to what?
Garth Sundem 17:58
Well, so I’m tempted to ask how that works, and ask if this requires human coding at the front to train the AI that then goes in and works with this unstructured data. But the second I do that Dr. Marci, everyone is going to turn off this podcast. So instead of going smaller, let’s go bigger. You know, does this same approach or a similar approach? Or the same model? Does this generalize beyond treatment resistant? Depression? I mean, more outcomes, more data means better outcomes. Does it generalize this approach the model?
Carl Marci 18:33
Yeah, absolutely. And we’re getting better every time and we meet in the field, but also here at OM1, you know, when you think about our broader dataset, right, we have 300 million patients. So we have data on, you know, a large majority of the country. And we’re working in other areas, immunology, cardio metabolic areas, I haven’t really focused on mental health and, and neurology. We started with depression because it’s common in and you know, in the World Health Organization says, you know, it’s the number one worldwide cause of disability in the world, right. So okay, let’s start with something that’s very common. We know that depression makes other physical illnesses worse and harder to treat. And it’s very expensive, right? You know, billions and billions of dollars are spent on lost productivity, utilization of health care, etc. So, you know, we thought it was a good place to start, but you’re absolutely right. Next up, we’re looking at models for bipolar disorder. We’re looking at models for schizophrenia. We’re looking for models for generalized anxiety. We’re looking at models for autism, PTSD, Parkinson’s, migraine, some sky’s the limit. So what we were hoping is we can really get quite good at building these models, and they take time. They don’t happen overnight. But you know, I think the first model we built here probably took on the order of two years. The second one was a year you know, when Hour down to about three months, which isn’t bad. Hopefully we can get down to three weeks or less.
Garth Sundem 20:05
So is it is it define, predict? And then what treat or recommend treatments? Is that sort of the three step process to these things?
Carl Marci 20:16
Yeah, I think that’s I think that’s exactly right to define what the condition is, and model it so that you can find it in those who aren’t labeled. So you can expand the universe and make sure you’re not missing anybody. And then, you know, predict who may have it or who is at high risk so that you can kind of shine a bright light on them and direct them to the right care they need, and then help inform what the right treatment for that population is. I think there’s a fourth step, ultimately, when we get really good at this card. And that fourth step is prevent, right? So if we can get to the point where we’re these models are robust enough and done at scale. You can imagine a world where we can prevent dreamer resistance, because we’re actually identifying you as an adolescent. Right? Who as prone to depression, because we know adolescents depression go on to have harder to treat depression as adults. So why wouldn’t we do that?
Garth Sundem 21:08
Well, I think that we should make that the topic for our next podcast, Dr. Marci, the prevention of treatment resistant depression or mental health issues, but for today, let’s leave it there. So thank you for joining us. And to learn more about how your organization can partner with OM1, visit OM1.com. That’s OM1.com. MAPS members don’t forget to subscribe. And we hope you enjoyed this episode of the Medical Affairs Professional Society podcast series: “Elevate”