Harnessing AI for Mental Health Breakthroughs with Dr Sarah Morgan

Episode 12 • 17th July 2024 • Women WithAI™ • Futurehand Media

00:03 - Joanna Shilton (Host)

It's my honour today to welcome Dr Sarah Morgan onto the podcast. Sarah has a background in theoretical physics and is currently looking at using AI to predict mental health conditions, so I'm really looking forward to learning a lot from speaking to her today. But before we jump into our conversation, let me tell you a little bit more about our guest.

Dr Sarah Morgan is a senior lecturer in healthcare engineering, based at the School of Biomedical Engineering and Imaging Sciences at King's College London. She leads a research group which focuses on developing AI and data science approaches to understand and predict mental health conditions such as schizophrenia. To that end, Dr Morgan uses a range of different data types, with a particular focus on speech data and brain magnetic resonance imaging. Different data types, with a particular focus on speech data and brain magnetic resonance imaging. Prior to joining King's, Dr Morgan led the AI for Brain Sciences group at the University of Cambridge. She holds a PhD in theoretical physics from the University of Cambridge physics department and co-founded the Cambridge Women in Physics group while she was there. She also holds a master's degree in physics from the University of Exeter. So, Dr Sarah Morgan, welcome to Women WithAI.

01:07 - Sarah Morgan (Guest)

Thank you for having me.

01:10 - Joanna Shilton (Host)

Oh, it's a pleasure. Well, all of that sounds incredibly impressive and really, really cool, and did you ever think you'd be doing this when you were a little girl?

01:20 - Sarah Morgan (Guest)

No, I think it was sort of when I was 15 or 16 that I got interested in physics and decided that's what I wanted to study at university. Yeah, I guess it sort of all kind of went from there.

01:34 - Joanna Shilton (Host)

So that's how you got here. So just maybe you can sort of explain all that in like really simple terms for everyone listening that might not know what theoretical physics is or how you've sort of gone from healthcare engineering to AI. So yeah, can you give us a sort of introduction to you please?

01:50 - Sarah Morgan (Guest)

Sure, so, as you said, I lead a research group that's really trying to develop AI and data science tools to try and predict and understand different mental health conditions, and we do that with a range of different data types, like brain images and short excerpts of patient speech and and yeah, since how I got there, I'm a physicist by background and I wasn't actually using AI during my PhD, but I was looking for patterns in sort of large, messy real world data sets and found that that was something that I really enjoyed doing, and then, I think, towards the end of my PhD, I wanted to do that, continue doing that in a way that was a bit more applied, and healthcare just seemed like such an interesting application area.

02:43

And I got interested also in studying the brain, particularly studying the brain as this sort of really complex network of different brain regions and how those regions work together, and I just sort of got interested in that and started looking at mental health as an application in Korea. The other thing that I really learned during my PhD was around working with people from a wide range of different backgrounds. So, though I was a theoretical physicist, I was working with chemists and biologists and experimental physicists, and those are very much skills that I use now as well. I think AI generally is this very interdisciplinary area, particularly for the applications of AI to different fields like healthcare, and so now my work involves working a lot with clinicians, also people with lived experience of mental health conditions, as well as, obviously, of AI researchers and researchers in other areas like linguists.

03:49 - Joanna Shilton (Host)

ably here as well until about: 1993

04:27 - Sarah Morgan (Guest)

I definitely it's a really sort of important point to raise. Um. I remember reading Invisible Women um by Caroline Credo Perez um some years ago now and just thinking, yeah, this is really really incredible. The, the, yeah, women just were so excluded in these different areas. Um, things like mobile phones as well.

04:49

I, I mean not just healthcare, but things like technology, mobile phones being quite large, and yeah, it's really hard as a woman, sometimes, like with a small hand, to be able to reach the buttons that you need to yeah, sort of things like that, isn't it? And obviously in healthcare it has real implications for women's health and for the risk of different conditions and so on.

05:18 - Joanna Shilton (Host)

So the data you're looking at, do you make sure it is equally split with men and women, or does it depend what you're looking into?

05:25 - Sarah Morgan (Guest)

So we try to make sure that we're really representative of the population that we're looking at. So we try to make sure that we're really representative of the population that we're looking at, and so I mean different mental health conditions can have different prevalences. So I work a lot on schizophrenia, which is actually more prevalent for men, particularly for the younger men, particularly men from men who are black or from other ethnic minority groups, and so we try to really make the research that we're doing representative of the population that's affected by the particular mental health condition. If you look at something like depression, most of my work tended to focus more on schizophrenia, but depression, for example and we have done a little bit of work on that in the past is particularly prevalent for women, particularly that sort of 14 to 20 or so years age range. That we're obviously having a sort of diverse representation of patients in the data that we're working with and particularly representing the populations affected.

06:41 - Joanna Shilton (Host)

And how has AI changed it? So were you using AI at the beginning of when you started looking into this, when you sort of transitioned from theoretical physics into AI and healthcare? How did you start to use it?

06:55 - Sarah Morgan (Guest)

Yeah. So I started to get interested in AI at that point and started to use initially kind of quite basic AI algorithms or different methods. AI methods that were kind of a good one to focus on at that point, I think sort of more generally in terms of how AI is changing the field. Historically, mental health conditions have generally just been diagnosed by looking at symptoms, so by clinicians talking to patients, asking them what symptoms they have and then making decisions about different treatments and so on on the basis of those conversations. And that's quite different to many other areas of medicine where we have blood tests or different images that are collected and so on, like scans and so on, and so that's something that we're really, I think, as a field, psychiatry is really keen to move towards more of these sort of biomarkers, biological markers that could perhaps help to inform different clinical decisions, and I think AI has a role to play in helping us to get to that point.

08:17

So a lot of my work, for example, is with brain images, and if clinicians look at brain images, say, from someone with schizophrenia and someone who doesn't have schizophrenia, you can't really tell by looking at those brain images as a human sort of, which person has schizophrenia and which person doesn't, but we're finding that computational models, so by training AI, these AI models are able to spot that difference, and so the way that we think about using AI in my work is really around trying to use AI to give clinicians extra information from, for example, brain scans that clinicians wouldn't otherwise have available to them. It's really about giving extra information to clinicians and not about replacing clinicians with AI. I think that's something that you do a lot of work with people with lived experience with mental health conditions, and that's something that they always really highlight as a concern If AI is actually really used to replace clinicians, because people with mental health conditions often really value the therapeutic interactions that they have with those clinicians, and that human-to-human interaction is really important as well.

09:44 - Joanna Shilton (Host)

Yeah, I can see that. Yeah, because you wouldn't want to just go and see a robot doctor, but you're coming to see the clinician and it's just that. As you say, it's a tool to help them. So you still need that human input.

09:56 - Sarah Morgan (Guest)

Yeah, yeah, it's absolutely something that I see as a tool rather than a sort of end point in itself.

10:06 - Joanna Shilton (Host)

And I'm imagining that it's quite a sort of wow. That all sounds really exciting right now. But how do you, what exciting developments do you think will sort of come from this? Like, how do you, how would you like to see it sort of I don't know change or get better? I mean, do you think it will get better, do you? You're not in the camp that thinks, well, it will take over. It's like we need to keep it as a tool. So how can it?

10:37 - Sarah Morgan (Guest)

I don't know what. What developments are you looking forward to? So I think for me, the real sort of challenge for me is can we use AI in a way that really benefits patients' lives? Um, and I see sort of two different routes to doing that. One One is if we can use AI to better predict different individual trajectories, so basically being able to use AI to maybe feed brain scans into AI and for the AI to be able to say actually this treatment might be better than this one because of this particular brain connectivity pattern that this person has.

11:11

So really sort of trying to help inform different treatment choices. And we're definitely not there at the moment and there's quite a lot of work to do to get there. So I think that's sort of one area that we're really trying to work towards. The other area where I think AI can potentially help us to have an impact is also in better understanding of these mental health conditions, because we're still, I think, still fairly early on in our understanding of what actually causes these mental health conditions, for example, what brain regions are involved, what combinations of different brain regions and connectivity between different brain regions is important, and if we understand those things better. And I think AI can help us to do that then that could potentially lead to new treatment options, because if we know, for example, what parts of the brain to target, then that would help with developing drugs so are there other people doing this as well?

12:21 - Joanna Shilton (Host)

because I guess it's. If it's just your data and the data that you've got the time to collect, how does it go into, like do you share that data or are you working with other universities or other departments to get more data? How does it go into, like, do you share that data or are you working with other universities or other departments to get more data? How does it? How does that all kind of fit together?

12:36 - Sarah Morgan (Guest)

So I don't collect data myself. I tend to work with hospitals and work with hospitals around Europe, worldwide, who are collecting data, and then I tend to be on the data analysis side of things.

12:52

so building computational models that are analysing that data, and what the field has found is that you need quite a lot of data.

13:02

It's generally true for these AI models that they're quite data hungry and where you see sort of real success stories for AI At the moment, those are generally areas where there's been really huge amounts of data available.

13:19

So, like with the large language models, which are very prominent and talked about at the moment, they've had access to huge amounts of data from the Internet, lots of kind of written data, that amounts of data from the internet, lots of kind of written data that people have shared on the internet, which is what's been able to make those models so successful the size of those data sets. And so for us, I think what the field's been learning kind of generally is that you need to sort of share data between different hospitals to get the sort of sample sizes that you need. Historically, people were maybe collecting, you know, data from maybe 20 or 30 people, and that was. You know it's hard to scan and that was a. You know it's hard to scan large numbers of patients and to recruit patients to scan them, and so, yeah, really, by combining different data sets and by working together, we can get a lot further.

14:23 - Joanna Shilton (Host)

And do you get to say what data you want or what's missing? Because, just thinking as you say about the kind of, or like um, the invisible women book and about women being missed out on research, do you get to say, well, actually we need to look at these ages and then, when you know we need to look at women at certain times of their menstrual cycle as well, I mean, do you go that deep into it? Because I'm guessing that you know, with things like depression and things that maybe are more prevalent in women at certain times or and men at other times, a lot of that could be all down to hormones and the, that our bodies fluctuate, you know, throughout your cycle. So every sort of you know four months, four weeks or so, I mean, do you get to ask for those or to try and encourage that data to be collected, or is it? Yeah, what's it dependent on?

15:12 - Sarah Morgan (Guest)

I think these are really sort of decisions that are made at the level of. The way it works is that different sorts of consortia of scientists come together to collect data and at that point they decide what's important, what are we going to prioritise. And it's often difficult because you don't want the burden on the patients to be too high, so you don't want to be giving of giving someone questionnaires that are gonna take them days and days to complete. Like there has to be some sort of choice often around sort of what to include and what not to include, um. But yeah, it's sort of done at that kind of consortium level where lots of different scientists with different research interests, um will come together and try to sort of read the consensus on what's most important.

16:04

The other area that I this is quite a bit about brain imaging. The other area where I work quite a bit is on speech data. I think that's interesting because historically, data that really hasn't been collected. Actually there's a field we've spent a lot of time collecting brain images from patients and actually not speech, whereas you'd think that speech in many ways is sort of easier to collect than brain images because you don't need a huge MRI scanner to do it, recording equipment, and you can do that.

16:37

But I think that the reason that speech data hasn't been collected so much is because historically, people didn't really have the methods. It was very time-consuming to transcribe the speech and then to analyse it, and you needed linguists and real expertise in the different sorts of linguistic models. And I guess that's another area where AI is really helpful, because we can now, because automated transcription is still far from perfect, but we're getting to a point where we can transcribe the data in an automated way and also analyse it in a much more automated way. So I think AI is also sort of helping or sort of changing priorities of what data is as well.

17:34 - Joanna Shilton (Host)

Because it can pick up on the tone of voice. Is that what you mean? Is that the data that you're collecting, so not just what people say, but how they say it or the speed they say it?

17:43 - Sarah Morgan (Guest)

Yeah, we're interested in both the sort of content of the speech and also the sort of acoustics. So how people are saying what they're saying and it looks like sort of altered language use is often a symptom of schizophrenia, for example. So people can sometimes talk in a way that's quite difficult to understand If they have schizophrenia and if they're in a particularly symptomatic state, and so that's sort of why we became interested in speech, for schizophrenia particularly. And yes, it looks like there's sort of signal both in what people are saying and also in how they're saying it.

18:34 - Joanna Shilton (Host)

That's amazing. And thinking back to when you said that the data that's being collected is decided upon by whoever's collecting it and what they think is important, I guess that's where we have to hope that it's not just sort of like completely male dominated field and sort of making sure that it is kind of a you know a spread of different well, different gender, you know male, female, different ethnicities, because I guess to each otherwise it's just what they think is important, not what maybe is best for everyone.

19:11 - Sarah Morgan (Guest)

Yes, it's also really about involving both people with lived experience of these conditions and clinicians in those discussions, because people with lived experience often have really great ideas about what they think is relevant and they have that experience to really know. And also clinicians can help sort of say what might be able to be collected in clinic. Because it's one thing, sort of having a tool that works great in a sort of research environment, um, but if you can't actually ever collect that data in clinic, then it's not, yeah, it's not really going to have that real world impact, um. So there are lots of different considerations that you have to yeah, you have to take into account there.

19:58 - Joanna Shilton (Host)

So I imagine that physics is quite a male-dominated field. Anyway, how have you navigated being a woman?

20:07 - Sarah Morgan (Guest)

in this field.

20:09

Yeah, so I guess there's so many different things to say about that. I think often the challenges that women face in male-dominated fields like physics now are often these sort of unconscious biases rather than sort of overt discrimination. I think often the sort of unconscious biases that I've experienced are things like someone assuming that a male colleague wrote a piece of code that I wrote just because he's male and therefore they think that's more likely, um, or yeah, kind of giving more credit to male colleagues. Um, there's quite a lot of literature showing that research papers by men tend to be rated more highly than research papers by women, even if it's the same research paper and you just change the name at the top of it. Like and similarly with cvs and so on. Um, professor Danny Bassett in the US has done quite a lot of work around sort of citation practices and how they're sort of different between different genders, um, which is really interesting. Um, I'm not sure about in terms of how to navigate that.

21:22

I think often what people say is around kind of not being afraid to voice your achievements and to remind people of your credentials, because they do have these sorts of unconscious biases and we all have these unconscious biases. Yeah, it's not just men but women as well have these unconscious biases sort of built into us, don't we? So I think that's one thing I think also, but it's difficult, isn't it? And lots of us don't really want to kind of shout about our achievements or to just get on with doing the job. But I think the other thing that people can do is sort of amplify other women's voices in meetings. So, yeah, sort of repeating something that someone said, or reminding people that a female colleague had that idea 10 minutes ago, in a kind of very friendly and nice way, but yeah, sort of trying to amplify those voices, I think is another thing I think the other thing that's really important to say is that these things aren't just true for women, but also for black people, people from the LGBT community, et cetera.

22:37

So there are lots of groups that are affected by these unconscious biases, and intersectionality is also really important. Black women face these unconscious biases in a way that's more than just the sum of being black and being a woman, but it's sort of an extra kind of cost to that.

23:00 - Joanna Shilton (Host)

Yeah, I think for me me, that is my biggest worry about ai that it is because it is only learning on all that historical sort of unconscious bias or the that we don't really realise that we're doing until it's pointed out. And then you're like, well, actually, yeah, you know why. Why are you thinking that just because he wrote it is better than if she wrote it? And that's the sort of bit that I'm hoping we can sort of overcome. I mean, yeah, that seems to me like the biggest challenge. I mean, but what's the biggest challenge that you're fighting with? Sort of applying AI to the mental health research? Does gender or bias come into it? Or are you with sort of applying AI to the mental health research? Does gender or bias come into it? Or are you just sort of because you're just looking at the brain or the speech? You're not thinking about that. But, as you say, if it's, if it's more prevalent in men or in women, yeah, I don't know how do you navigate that?

23:54 - Sarah Morgan (Guest)

I think these are definitely things that you have to keep in mind and things that we can. I don't know whether the internet does it this bias or that bias Like I say, working with people with lived experience can help you to sort of identify which biases might be particularly important and things yeah, really I mean obviously things like gender and age and so on that you do just check for, but sometimes people have ideas for other things to include there. I think the for me, the thing that I think we have to be really careful about with AI is that it's not able to replace humans, particularly when making high stakes decisions. There are sort of some areas where AI can and is already being used very widely and generally areas where they're not sort of very high stakes decisions. It's kind of quite routine, repetitive things, and their AI actually is already really quite prevalent.

25:07

I mean, for us, the sort of more technical example is that we often want to segment the brain into areas of different tissue types.

25:20

So you get a brain image and there are different types of tissues in the brain and we want to say which parts of the image are which tissue type, and AI is already really helpful there, because historically people had to go and kind of draw onto these images kind of lines of where this tissue ends and that one doesn't, and obviously that's really laborious and AI can help with that very easily. So there's some things like that where they're kind of really easy wins for AI. But then the things that we're starting to think about now so there's some things like that where they're kind of really easy wins for AI, but then the things that we're starting to think about now are much more complicated. And if you're thinking about making a decision about what treatment someone should go on to for a condition like schizophrenia, then that decision really has to be made by a clinician and it's, yeah, really just about using AI to help that decision.

26:18 - Joanna Shilton (Host)

Um, yeah, so I guess that's the thing that I'm generally most, most concerned about in in the work that I'm doing yeah, I guess, and from the sort of patient side of it as well, you probably feel a bit happier if you know that, yeah, a human or a clinician has actually had a say in it and you're not just getting the information straight from AI. It's definitely been looked at.

26:45 - Sarah Morgan (Guest)

Yeah, it's important for a whole host of reasons that AI can't take responsibility in the way that a human can take responsibility for making decisions and the sort of value of these therapeutic interactions that we were talking about earlier.

27:01 - Joanna Shilton (Host)

Yeah, yeah. So going back to the kind of like male-dominated field of physics, because you, as mentioned in the well, when I introduced you, you set up the Cambridge Women in Physics Group. How?

27:16 - Sarah Morgan (Guest)

did that come about and is it still going?

27:18

Yes, so there's a group that I set up about 10 years ago now with Dr Hannah Stern, who's now a lecturer at Manchester University and a really fantastic experimental physicist there there, and I think our motivation for setting up the group originally was just that Cambridge University physics department is quite large and back sort of 10 years ago there were a number of women who were in sort of different groups of the department, but they were quite isolated in those groups and didn't really have so many opportunities to meet each other, and what we wanted to do was to sort of build a space where people could meet each other and could then support each other, which I think is really important for women who are in this sort of male-dominated environment important for women who are in this sort of male-dominated environment.

28:17

Um, and yeah, so that was our sort of motivation for creating the group and it is still going 10 years later. Um, so I think there was definitely a yeah sort of a need for that. Um, I have gone through successive kinds of different leadership and different things involved with it over the intervening 10 years, but, yeah, it's really nice to see it still going strong.

28:42 - Joanna Shilton (Host)

Fantastic. Well, that's kind of one of the reasons that we started doing this podcast as well. It's just, you know, it's to amplify the voices of women and it's to support each other and just sort of get that message out there that, you know, this shouldn't be a male dominated field, you know it is, you know, open to women and we need to kind of jump in. And, as you say, I know that quite often, you know, as women, we do just think, well, I'm just going to get on with it and I'm not going to shout about it. I've, obviously, I've done this. People should just recognize that and realise that. And it's that kind of know, maybe shouting a little bit louder about everything that we've achieved. And especially, you know, in all the work that you're doing and looking at all this, you know, all the ramifications of that are, just, you know, incredible. So it's, yeah, I mean, ai must be sort of part of everything that you do. Now.

29:28 - Sarah Morgan (Guest)

I've quite a lot of what I'm doing. Yes, yeah, yes, I think, and I think it's going to continue to be, and yeah, it's really exciting at the moment actually, because we've got so many sort of new AI techniques and the field is moving so quickly, which is a challenge, also definitely a big challenge, but I think it is an exciting time, particularly with the advent of these more sort of generative ai techniques, where ai is actually generating data um rather than just being used to analyse data um and yeah, so that's that's quite exciting for us really exciting.

30:12 - Joanna Shilton (Host)

So if people want to learn more about everything that you've told us or um sort of get involved, um we're doing, I think, is there any, any, any recommendations that you have any things, uh, any sort of websites to visit or how they can get in touch with you?

30:24 - Sarah Morgan (Guest)

um, I think linkedin's probably the best place um so on, linkedin I can send you yeah, share a link, and so on for that.

30:33 - Joanna Shilton (Host)

We'll share that. Well, thank you, sarah. It's been really interesting to speak to you. I just wish you every success with everything you're doing, and we'll probably have to get you back on the show in a year or so or something and just find out about all the developments that have taken place and how it's changed. Dr Sarah Morgan, thank you for coming on. Women with AI.

30:51 - Sarah Morgan (Guest)

Thank you for having.

Share Episode

Shownotes

Transcripts

Follow

Links

Chapters

Video

More from YouTube