Latest / The Biological Psychiatry Podcast / Enriching Antipsychotic Clinical Trials with Speech Biomarkers
Transcript
- Elvisha Dhamala: That's all for today's episode. Thank you to Dr. Alex Cohen and Dr. Mark Obler for joining us to share their work on using speech biomarkers to improve clinical trials and schizophrenia. For our listeners, you can find the full paper in Biological Psychiatry and we'll include the link in our episode description. To learn more about our journals, please visit biologicalpsychiatry.com. You can find The Biological Psychiatry Podcast wherever you get your podcasts. now so you don't miss our next episode and follow us on our social media channels for highlights and updates. Thanks listening. ⁓ This is for informational purposes only and does not constitute medical or professional advice. ⁓ The expressed are those of the participants and do not necessarily reflect those of the Society of Biological Psychiatry, ⁓ joining me, we Dr. Alex Cohen from Louisiana State University, first author and Dr. Mark Oppler from Clareo, ⁓ co-author of a recent in biological psychiatry ⁓ that demonstrates how analyzing patterns can improve clinical trial outcomes. Dr. Cohen and Dr. Oppler, welcome to the Biological Psychiatry Podcast. Clinical trials are research studies involving human volunteers designed to evaluate the safety and effectiveness of new medical treatments, drugs, devices, or diagnostic tools. These trials have multiple phases. Phase one focuses on safety and dosage with a small group of people, typically healthy volunteers. Phase two evaluates efficacy and side effects in a larger patient group. Phase three confirms efficacy and monitor side effects in large scale, often randomized international trials. From the Biological Psychiatry family of journals, is the Biological Psychiatry podcast, ⁓ we discuss the latest research on psychiatric disorders and their treatments. ⁓ you for joining us. I'm your host and Social Media Editor, Elvisha Dhamala. ⁓ Alex Cohen: Thanks for having us. Mark: Thank you. Elvisha Dhamala: Psychiatric trials fail at much higher rates than trials for other diseases. What makes them so challenging compared to trials for other diseases? its family of journals, its editors ⁓ rights reserved, including those for text and data mining or use in AI systems. ⁓ trials. Phase IV occurs after FDA approval to monitor long-term effectiveness and safety in the general population. Clinical trials are very expensive and prone to failure. In fact, psychiatric clinical trials face some of the lowest success rates among major disease areas. Today, we're exploring a novel approach to improve these odds, one that uses something we all do in every conversation, the way we speak. Mark: ⁓ there's a lot reasons that trials fail. think the principal in clinical trials in psychiatry ⁓ is ⁓ inability to differentiate drug effects from placebo. ⁓ And this partially ⁓ I think, to the inherent noise in experiment, in any clinical trial. in psychiatry, we have the ⁓ added problems ⁓ of ⁓ First, highly subjective endpoints. We're often asking people about their subjective sense of improvement. We're evaluating sometimes relatively subtle changes, and again, subjective changes in behavior. But we're also trying to study diseases whose cause and underlying biology we simply don't fully understand. diagnosing on the basis ⁓ of syndromes. let's say, a study where we're looking at blood pressure, where we strap a cuff on somebody and numerically measure how much higher their blood pressure is today than yesterday, we don't have a cuff that we can strap on you to measure how depressed you feel or much better or worse your hallucinations are today than yesterday. So I think the subjectivity combined with other sources of noise, both in the diagnosis and the evaluation of these diseases, creates a tremendous challenge. when we're trying to ⁓ ⁓ artistic aspects, subjective aspects of this often make it really challenging, This is why I believe that the from subjective more objective endpoints that are closer to the underlying biology. of whatever these diseases actually are, is absolutely vital. Alex Cohen: yeah, this is an essential move towards precision psychiatry ⁓ is the use of some objective biomarkers that are closely to the psychopathology, ⁓ to functioning, and to the underlying biology. Elvisha Dhamala: So in this paper, you mention enrichment strategies. Can you explain what these strategies are in the context of clinical trials ⁓ and makes them so important? Alex Cohen: Enrichment ⁓ are using patient characteristics to participants that somehow enhance the signal detection of drug ⁓ These are done prospectively, so before randomization. there's three general strategies. But they involve things like trying to ⁓ placebo effects, ⁓ identify ⁓ people that more likely to placebo. or very quick responses that are unlikely due to drug effects. Strategies for identifying high risk patients that ⁓ show the disease state, ⁓ identifying patients that are more likely be responsive to treatment. Elvisha Dhamala: So that's a great segue into the next question. You mentioned earlier for something like blood pressure, we have a blood pressure cuff. So again, you needed something more objective in psychiatry and you decided to focus on speech. Why did you focus on speech specifically and is there something unique about how we speak that makes it important in psychiatry? Alex Cohen: Speech is a critical human ability, right? ⁓ It's important for ⁓ people they navigate their daily routine, relationships, work, recreation. during a interview, it's the primary media that's being evaluated. And it's something... easy to objectify. Speech technologies are well established. They've existed for at least 70 years. They've increasingly become sophisticated and more scalable. And I think their promise for capturing speech, ⁓ just for a few languages, ⁓ for people all across the world is improving very quickly. ⁓ the clinical interview, ⁓ speech biomarkers can be extracted from that are already ⁓ conducted. So there's no additional burden on the participant or the study team, ⁓ additional blood draws or tests or interviews. ⁓ Elvisha Dhamala: So you looked at speech latency in particular. Can you tell us what that is and why it's important to measure how someone is pausing responding? Alex Cohen: Yeah, of course, speech latency refers to pauses. And here we're interested in one particular type of speech latency, which is a turn latency. So that's the time in milliseconds between when one speaker stops talking and the other speaker starts talking. So the interview asks a question, how are you doing today? And the patient processes the question, formulates a response, and then begins articulating the response. In turn, latency is the time that it takes for those mental processes to occur. psychiatric interview or the interview is ideal extracting speech latencies because it's structured, ⁓ it's typically delivered by trained personnel, the questions are so it puts a bit of a processing load ⁓ on the patient. There are follow-ups and there are also lots of turns in an interview, so you get a reliable estimate. Elvisha Dhamala: you ultimately ended up using speech latency ratios ⁓ raw latency itself. Can you what that ratio captures and why it works better than just response Alex Cohen: We did some work depression and we saw that speech latencies were a good measure of psychomotor slowing. And we think with depression what happens is people ⁓ may be slowed. It takes extra time to process and articulate a meaningful response, but they're able to articulate a meaningful response. I think when patients have more profound social cognitive deficits, the response is also affected. And the speech latency ratio accounts for this. So we're including the latency, which is a measure of psychomotor slowing, but also the duration, which is a measure of response. together, we think that captures a person's ability to engage with the interviewer more than just their psychomotor ⁓ response time. no doubt that ⁓ is a kind of crude measure of somebody's response, ⁓ It doesn't include the sophistication or the number of concepts that people are ⁓ or ⁓ how linear the response is or how good a response it is to the person's question. And are many, many creative and sophisticated ways try and ⁓ somebody's response, using LLMs. ⁓ Our use case here required our measure to generalize across languages and across very different people. And there are big challenges in using LLMs that way. So duration seemed like a simple and interpretable and transparent way to capture people's response. Elvisha Dhamala: One thing that is really impressive that you just mentioned is this was a global trial with eight different languages. So how did you ensure that any biomarker you find would actually be generalizable and would work across such diverse populations? Alex Cohen: So ⁓ we first consulted experts the when they were available. We put in place a ⁓ assurance ⁓ pipeline to ensure that the was working across these different languages because they ⁓ use different modeling ⁓ and different training techniques. And then we the reliability and validity ⁓ of our. ratios across the languages. It was a bit of a challenge in that the sample size varied across these languages, ⁓ but generally we saw that the ratio performed. The mean values sometimes were different across languages, as would be expected, but it seemed to perform ⁓ psychometrically well across all languages. Elvisha Dhamala: So the study that you did was building off of a phase three trial for a schizophrenia drug. Can you describe the study design of the clinical trial itself and how you incorporated your speech analysis into it? Mark: ⁓ design of the study was very similar to many pivotal trials schizophrenia ⁓ for new ⁓ treatments where you have ⁓ randomized into one of two arms patients are either going to put onto ⁓ placebo ⁓ matched for all the characteristics that... we believe matter, right? Appearance, you the pill looks exactly the same, sort of thing. Or onto, you know, one of two doses of active treatment, a low dose and a high dose. ⁓ again, this is fairly standard, I think, across almost every large trial that we see. ⁓ the details will differ. ⁓ the idea here was that, you know, we were not changing anything about the underlying trial. In fact, the underlying trial had already been conducted. So there was no way for ⁓ approach to affect the results. For this study, we had several advantages. ⁓ One is was already employed a set of technologies that enabled us to very easily and effortlessly capture the audio. ⁓ So study, instead of using paper forms where clinicians had their interviews and and all of their questionnaires, there was actually a tablet that was provided to every site by Clario. So that every assessment, everything from ⁓ clinical global impression scale to the positive and negative syndrome scale of PANS, ⁓ to assess motor function, all of these things are being collected electronically. And the PANS assessment, which was the primary outcome, PANS being the gold standard for assessment of efficacy treatments in schizophrenia, this was captured electronically along with video recordings of each of those assessments in every language for every patient everywhere in the world. ⁓ So that was the essential design. one of the beauties of this approach is that it's invisible. You don't really see it. Whether you're recording audio on a tablet or through some other means, it's become such a standard part of psychiatric evaluation, nobody really sees it anymore. ⁓ Elvisha Dhamala: So is it safe to say that speech is perhaps a far more accessible biomarker than potentially some of the other technologies that are being considered for biomarker development? Mark: I don't think there's any question about that. Speech is one of the things that defines our species, isn't it? It's what differentiates us, I think, from other primates. I would add to that speech probably one of the most complex behaviors, one of the last to emerge neurodevelopmentally. Because of this, I believe it utilizes ⁓ so many that we have engage as human beings. Social resources when we're interacting with another person, cognitive resources to dredge up process and the information. Neuromotor resources to deliver it. All of this comes together in speech and in the act of a clinical interview, which again is a very standard part of any clinical assessment, especially in psychiatry. It's the most ubiquitous and probably the most accessible form of what I would regard as biological information we have available to us. what makes speech, speech latency and speech latency ratio even more beautiful in my view. is the fact that it generalizes across every language, every culture, every country, every person we've evaluated using this technique. This is valid, this is replicable, and it's meaningful. Elvisha Dhamala: So in the study you analyzed over 2,300 recordings from psychiatric interviews. Can you talk a little bit about your main findings? So when you enriched the clinical trial sample, what did you see in terms of the treatment effects? Alex Cohen: Yeah, the treatment effects improved typically sometimes doubled in nearly every and secondary total and subscale score. We saw statistically significant improvements ⁓ baseline to end of treatment nine out of the 12 outcome measures. ⁓ And this is impressive because we had the participants. So a big power reduction there. I think this was particularly for negative symptoms where the treatment placebo separation jumped from about 0.39 to 0.71 with the enriched sample. Elvisha Dhamala: let's talk a little bit more about that. Can you tell us why negative symptoms in particular are so important in schizophrenia trials and why perhaps they're so challenging to address? Mark: negative symptoms have been a target for treatment for a long time. they're one of the things that we've been struggling to both understand better and also to treat more effectively. Mostly because there are good reasons to believe that this is the primary barrier between patients and their recoveries. the resumption of the lives that they lost they became ill ⁓ is really blocked by these deficits in motivation, language, engagement, interest. ⁓ There are lots of good reasons why ⁓ world wants to find a better treatment for negative symptoms. ⁓ study that we're talking about today was actually not a negative symptoms trial. It was a general trial for overall antipsychotic efficacy. yet negative symptoms proved to be incredibly important in defining a group of patients who really had meaningful responses to drug. I believe this is because even when our focus is on more obvious, the more classical of ⁓ psychosis, ⁓ negative symptoms still fundamental part of that disease and disease process. And I think what speech latency ratio does, it identifies patients who have these fundamental deficits and who probably have a very specific form or phenotype of the disease that is particularly sensitive and responsive to drug treatments. Alex Cohen: interpreting ratios very much still in its infancy. And so ⁓ with this study, we were primarily replicating and extending a study of bipolar depression that we'd done with Seth Hopkins, with Steve Sabo and Josh Siegel, who were then at Sumitomo Pharma. And the process requires some symptoms to anchor our speech biomarker in. So we can decide what a cutoff value. should be for enriching the sample. So we decided to use negative symptoms for this study, and that was primarily for ⁓ replicating ⁓ methods. Since then, ⁓ we've on, ⁓ and we're looking directly at placebo effects and treatment response ⁓ across wide variety of symptoms, including negative symptoms, ⁓ but ⁓ general mood and anxiety, and things like that. Elvisha Dhamala: the effect sizes that you mentioned are very impressive. So you could double the drug placebo effect with half the amount of participants. What does that actually mean for clinical trial design? Mark: To me, ⁓ the ⁓ most important extension ⁓ ⁓ effect, I think, is the reduction ⁓ in size, complexity, ⁓ length, and the cost of trials. ⁓ I also think that ⁓ ethically, we have to think about implications too. ⁓ A lot of people ask, know, why are clinical trials now so difficult to do? 20 years ago, 30 years ago, we could do studies that were much smaller, much more effective, ⁓ and they didn't take as long. Now it feels as though there's an entire industry that is ballooning and the costs are just spiraling out of control. you know, what I see within the scientific elements of the work that's been done, I see a promise of ⁓ smaller, efficient, costly more ethical trials where we identify a specific group of patients with a well-defined who actually stand to benefit. from the treatment that we are investigating. Costs aside, timelines aside, that to me is worth all of the effort that we've expended. And I look forward to seeing a day when we can do faster, smaller, more precise trials. Elvisha Dhamala: getting back to speech latency what do abnormally long speech tell and what cognitive or social processes might these be tapping into? Alex Cohen: a really good question, and I think that's one that we're still disentangling. ⁓ I think it's clear that during an interview, at least, this ⁓ of ratio ⁓ an indicator of somebody's engagement in the process. Engagement can reflect many factors, some cultural, ⁓ some but certainly cognitive, social, and motivational factors. ⁓ and ones that are important for psychopathology, especially serious illness like ⁓ schizophrenia psychosis. We ⁓ sort engineered this measure of ratio so it would be relatively stable over time and it would be used during the baseline and screening sessions of a clinical trial. So I think what we were really interested in is identifying people ⁓ who were disengaged, people who were really struggling to engage during the session. And this abnormally long ratio, which is caused by long latencies and short response durations, ⁓ very clearly reflects somebody who's not engaged in the session. Elvisha Dhamala: In your prior work, you also found that there was a speech latency biomarker that's useful for enrichment in a phase two antidepressant trial. What's similar about that work versus this work, how might the work that you're doing overall generalize across psychiatric illnesses and clinical trials more broadly? Alex Cohen: There's a lot of similarities. ⁓ The methods we used are almost identical, particularly for making decisions around enrichment. ⁓ The general results, very similar, you know, with about the subjects and in some cases a doubling of the effect. We use the same automated speech analysis pipeline. And the feature seemed to work across ⁓ diverse people. And particular, you know, ⁓ in study, seven languages. think the are simply the feature in that we were looking for latencies and ⁓ speech latency ratios ⁓ in this ⁓ ⁓ the and population. The reason why it's working may be slightly different in the work, ⁓ what we saw was fast speech latencies identified people that were primarily showing a placebo In this study, we do see some placebo response, but we see also ⁓ enhanced treatment effect in people ⁓ long ratios. ⁓ We still have lots of work to do in terms of figuring out exactly going on ⁓ and neurobiologically. ⁓ Elvisha Dhamala: So what would you need to see before this becomes standard practice in all clinical trials? Mark: feel that for certain applications, certain use cases of this approach, ⁓ the data exists today to consider validity ⁓ and the reliability or the reproducibility ⁓ of the effect. I think it's also worth pointing out that right now there are several large studies ongoing that are using this prospectively, meaning they're actually evaluating speech latency and speech latency ratio, as patients are screened into the trial. that data is being used in a variety of ways. So, you know, the, the case, for the approach, I think, both in the recent publication and in the prior publication are very well made. I do believe that we need to complete analyses of additional data sets and also to really dig into the underlying issues, right? So why, for example, are certain patients classified one way and others another? What's different about these two groups? In depression, we're identifying people who ⁓ may have depression, but maybe not biologically mediated depression that really requires drug treatment. That may the vocal biomarker negative versus positive, the slow versus the fast speech latencies. ⁓ In schizophrenia, ⁓ or the schizophrenia is, depending on your point view, ⁓ will probably wind up being more complicated than that. But again, I suspect the overall shape ⁓ that it takes as this method progresses will feel very much the same. I don't even necessarily know that ⁓ this ⁓ all specific to ⁓ depression to psychosis. ⁓ Alex, has also done some fascinating work in neurodegeneration that suggests a role for speech latency. And I think we have good reason to believe that other psychiatric conditions and even other neurological and non-CNS conditions may show evidence of changes in speech latency and its derivatives. So validation is imminent for those of you who really enjoy things that are validated. Alex Cohen: I think there's a lot we can do in terms of improving the precision of these vocal biomarkers so that we can better interpret any individual, given their ⁓ age, culture, their language, ⁓ given the context, which interview it is, ⁓ who interviewer is, much more. And I think, you know, ⁓ we're getting involved in that process so we can sort of optimize these vocal biomarkers in a much better way, knowing, for example, when in the session to look at, when people may be particularly ⁓ with their engagement, and ⁓ being to interpret their data for. ⁓ decision making, I think, ⁓ be much improved. ⁓ know, what I envision ⁓ soon, hopefully, is ⁓ than a one size fits all cutoff, which we used in this study, more nuanced interpretation ⁓ of a ⁓ speech based on, ⁓ their kind of demographic factors one that can be used ⁓ to aid clinical ⁓ decision making. Elvisha Dhamala: So if there's one key message you'd want clinicians and researchers to take away from your work, what would it be? Alex Cohen: For me, would be that very simple speech measures ⁓ have been used for decades ⁓ easily be measured and used to assist clinical decision making. Mark: You I think ⁓ the takeaways here are that ⁓ these are no longer shiny toys or ⁓ curiosities. real. And I think ⁓ there's opportunity to use them for something that really matters. think one of the things that has held back the field ⁓ of biomarkers ⁓ is In addition to their dizzying complexity, very complicated measures in some cases, but also a lack ⁓ of It's easy ⁓ throw in cool sounding or cool seeming tools without really understanding what they're for, how they work, what their purpose is. ⁓ We're not to see this advance. ⁓ know, complexity is unavoidable sometimes, right? There are simply very important, complex measures that we need to employ. That said, more simplicity can finally introduce into our daily work, think the better off we're all gonna be. I'm excited for lots of reasons about this, but most of all, I'm excited to see what comes next. other thing I would say would like to thank ⁓ all of folks who ⁓ have collaborated us, ⁓ who have contributed to all of the work that's been done. ⁓ Certainly all of my colleagues at Clario who worked very hard to ⁓ do sometimes very things like ⁓ organizing audio files ⁓ cleaning data. All of that contributes mightily to these advances. And I thank very sincerely all of the folks at Reviva and other companies who put their faith in us. I hope that we have something really brilliant to show for it in the near term.