Latest / Technology Now / AI may soon use our voice to diagnose illness
Transcript
- Michael Bird (00:09): Hello, hello, hello, and a very warm welcome back to Technology Now, a weekly show from Hewlett Packard Enterprise, where we take what's happening in the world around us and explore how it's changing the way organizations are using technology. We're your hosts Michael Bird and Aubrey Lovell, although Aubrey Lovell's out today. She'll be back next week. But in this episode, we are looking at something pretty incredible. Now, there's a lot of hype about AI and most of it is fully justified. Frankly, it can do some pretty amazing things. And today, we're talking to somebody working on what Aubrey and I are pretty sure is one of the coolest uses of AI we've heard in a very long time. Yep, it's something which could revolutionize both healthcare at large, but also the way that organizations care for their staff and customers; that is using AI to determine whether you're ill just from the sound of your voice. So in this week's episode, we're going to be talking about how AI can detect signs of ill health. (01:04): We'll be discussing why the research is so important that it's attracted millions of dollars in research funding, and we'll be talking about the wider implications for our long-term health and the way that we treat illness in society. Sounds pretty incredible, doesn't it? So if you're the kind of person who needs to know why what's going on in the world matters to your organization, then this podcast is for you. And if you haven't already done so, do make sure you subscribe on your podcast app of choice so you don't miss out. Right, let's get on with the show. All right then, so usually on this show, we take a good look at wider issues and events affecting the tech industry and why they matter to you. But this week, we're going to be doing something ever so slightly different. This is a single research project, so cool that we couldn't let it pass by. (01:53): So let's start by adding some context. In the UK, over half a million people are out of work long-term with chronic disease, in the US, it's around the 6 million mark, and in the EU, 18% of work report having long-term mental or physical health conditions. And that's just the people with diagnosed conditions. It's hard to put numbers on it, but estimated figures are thrown about of $3 to $4 trillion a year of lost economic output in the US alone due to undiagnosed health conditions. Now, that being said, it's no surprise that any tools which can help the situation are getting a lot of attention. The National Institute of Health has launched a $100 million Bridge2AI program to fund research, and millions of that have gone to a study called Voice as a Biomarker of Health. So we are joined today by its lead co-investigator, Dr. Yaël Bensoussan, Director of the University of South Florida's Health Voice Center. Hi Yaël, very, very good to have you on the show. Yaël Bensoussan (02:59): Thank you so much for having me. Michael Bird (03:01): So first off, can you just tell us a little bit about the project? Yaël Bensoussan (03:05): Yes, of course. So the project is a four-year project where our main goal is to get 30,000 human voices associated to diseases and other health information. So we're collecting voices from people who have no diseases and people who have other diseases from five different categories of diseases. So we're talking about respiratory diseases, mood disorders, neurological diseases, pediatric disorders and voice disorders. So really across a really wide area of disorders. Michael Bird (03:40): And the idea is that you could create a tool where somebody could talk and just from hearing their voice, you'll be able to say, "Oh, you might have this condition." Yaël Bensoussan (03:52): So obviously, this is a very long-term goal. The goal of our group, so we're a very large group, we're 50 investigators from 12 different universities, our group is to create this large database because we know that in AI and in machine learning, without data, without accurate data and high-quality data, the models we develop are not scalable, not accurate, they can be generalized, they could be biased. So what's really missing in the voice AI role right now is large databases that we call open source. And that's really the main goal of our group, creating this database to make sure the data is accurate, that it's AI-ready for other researchers to use. So of course, some of the models and products will be developed from our group, but the goal is to create this resource for other researchers as well and maybe industry. Michael Bird (04:41): And so what's involved in the collection of those 30,000 voices? Yaël Bensoussan (04:44): So much work. Michael Bird (04:45): I can imagine. Yaël Bensoussan (04:48): So much work. So first of all, before even collecting the data, you have to understand that a lot of different people are needed from different backgrounds. So we have clinicians, we have acoustic researchers, we have ML engineers, but we also have ethicists. So there's a lot of ethics involved in what we do. So just how to develop these protocol, how to collect the voice, who does the voice belong to? Are the patients going to be too tired after 20 minutes of voice data collection? Should we make our protocols shorter? What's accurate? So all of that was done this year and the preparatory work for the year. So that's just the preparatory work involved before starting the data collection. And then we collect data or we start collecting data in our centers because it's very important. The way we modeled what we do is that we start by quality data that is supervised in our clinics, and then we're aiming to scale it for eventually people being able to donate their voice from home. Michael Bird (05:48): Wow. So what sort of diseases are you hoping that you might be able to diagnose or might be able to be diagnosed with this project eventually? Yaël Bensoussan (05:58): So we always talk about the lowest hanging fruits, the most intuitive ones, and then obviously maybe some that even humans can't hear. So when we think, as I'm a clinician, so I hear diseases all the time when people come into my office. If I close my eyes, I can tell with certain diseases. So Parkinson's, for example, is a very low-hanging fruit, a very intuitive one. Patients with Parkinson's have a very specific voice, monotone, slow-paced, they don't really change fundamental frequency when they talk and they really talk at a lower volume. So that's very specific. There's a lot of studies that have demonstrated that people with Parkinson's have a different voice than people without Parkinson's. So this is one example, we're working on Alzheimer's. And the mood disorders, we're working on depression and anxiety. (06:47): And you can tell that, you can tell your friend, "Oh, you sound sad today," right? I'm sure you've told somebody that one day. In voice disorders, we're looking at laryngeal cancer specifically and vocal cord paralysis, things, again, that as clinicians we're able to hear, but then there are some that are maybe not that intuitive and that are more difficult probably. So when we think about kids like speech delays, that's probably going to be a little bit more challenging for the data to talk. Michael Bird (07:13): And are you capturing frequencies beyond what the human ear can hear? Yaël Bensoussan (07:18): That's an interesting question. So what's really interesting about machine learning is before, when we did acoustic research, acoustic research is not new, it's been around for many, many years. And how people analyzed voice before is we extracted features like you say, frequencies or amplitude, so that we talk about in acoustic science and then we analyze them. What we can do with AI is with spectrograms, we can really analyze the sound as a whole without necessarily extracting features that the human decide that it's an important feature. And analyzing sound as a whole might bring us different information than analyzing sound with these specific features that human decided on. Michael Bird (08:01): What sort of technological innovations had to happen for this project to be able to happen? Yaël Bensoussan (08:06): Yeah, I mean, that's a great question. Again, machine learning changed everything. The voice research was around, I've heard a lot about the voice research projects. I meet a lot with pharma companies that said, "Well, we had voice labs in the 2000s and we were working, but all the analysis took so much time, we weren't getting results fast enough." I always give the example, I was a speech pathologist, I did my master's degrees and they would make us count syllables manually. We would count syllables manually. I mean, there was no way to make any progress fast. So of course, what changed everything in voice AI and why there's such an interest now, and while we knew that voice could diagnose diseases for many, many years, what really changed is the machine learning and the capacity of analyzing lots of data, really data science, lots of data really fast, with technology. Michael Bird (08:59): All right, final question from me. Can we try and put a date on it? When do you think a tool using this technology could be rolled out to users? It's a very mean question, isn't it? Yaël Bensoussan (09:09): Yeah. So it's a great question. I think, first of all, the database needs to be created. Then, in terms of machine learning models, tons are being created every day, with or without our team. If you're talking more about a tool, probably in the next three, four years, what's going to happen is tools that can measure health of an individual across time because that's easier. If I take your voice and you talk to your watch every day, and on Saturday you get sick, the technology can say really fast that there's a difference. And it can say, well, the difference in your voice has features of being tired or being sick or being short of breath. That's a technology, I think, a product that's going to come earlier. When we talk about diagnosis, we're probably about four or five years from really diagnosis because there's a lot of difficulty into diagnosing specific diseases. (10:04): So it's easy to diagnose a disease from normal, but what about the difference between Parkinson's and ALS, for example? And can we really make that distinction or can the app that we develop say, "Well, you sound like you have a neurological condition"? So specific diagnosis is going to be harder to obtain, but group of diagnosis or say, "Well, these features sound like you have a neurological condition." And I think that's what doctors do. When I have people come into my office, sometimes I can't say this is exactly what you have, but I say, "Well, what you have sounds neurological. I'll send you to a neurologist," "What you have sounds like from this category, I'll send you to that person." And that's really what's going to make an impact in public health, to be able to triage people, to screen people, screening and triaging. Michael Bird (10:49): Amazing. That's all actually really, really rather quite cool and incredible. Thank you so much, Yaël. And we'll be back with audience questions for Yaël in a bit so don't go anywhere. Okay, next, it is down to you, our audience, as we open the floor for you to give your recommendations on books which have changed the way that you look at the world, life and business in the last 12 months. Now, they could be technology-based, they could have changed the way that you work or they could have just made you look at the world in a slightly different way. If you want to share your recommendations, there's a link in the podcast description. Just record a voice note on your phone and send it over. Erin (11:35): My name is Erin, and something I read recently that really changed my perspective is The Great Forgetting: Earth is Losing its Memory by Summer Praetorius. And this appears in issue 47 of the Nautilus Magazine. And Summer is a paleoclimatologist, and in the article, she talks about how climate change and the history that's stored within ice and deep layers within the Earth is being impacted by climate change. And she also compares that to her personal history and personal memories, and it's just a really amazing illustration of the effects of climate change. And I'd highly recommend it. Michael Bird (12:24): Fantastic, thank you for that. So it is time for questions from the audience, you've been sending in your questions for Yaël on AI and voice, and we've pulled out a couple. The first question is from Leah, in your neck of the woods, in Tampa, Florida, who wants to know if the AI could be used as a therapeutic tool to aid in stroke and injury recovery as well as a diagnostic tool? Yaël Bensoussan (12:47): As a therapeutic tool? That's a very good question. So I mean, for voice AI specifically, where there's really a lot of potential is in treatment outcome measure, how do people respond to certain treatment? So maybe not to treat, but really to help guiding the management of what treatment we choose and how we know that a treatment is working. So for stroke, obviously with certain types of stroke, speech is very much affected. So let's say, in stroke, we have to give an agent that releases the blood clot really fast, seeing how the voice change can tell us if it worked or it didn't work. So I think voice AI will really change in how we measure outcomes with different drugs, with different treatment, and that that's really the biggest potential. Michael Bird (13:36): Cleo in Vancouver is a transcriptionist and wants to know if and how the AI overcomes things like regional accents and different languages. Yaël Bensoussan (13:45): That's a great question. So there's a lot of people working on that. The key is training your models on large enough data that includes these accents and languages. And it's really a question of numbers. If you train your models on five white men from Kansas, then your models are not going to be generalizable. And that's why it's so important to have large projects like our initiative and consortium where we're really collecting large amount of diverse data. So it will, but we just have to train our models on proper data that's diverse enough. Michael Bird (14:21): Brilliant. Thank you so much, Yaël. And again, we'll drop a couple of links in the podcast description for more on these topics. Right then, we're getting towards the end of the show, which means it is time for, apologies for my singing, This Week in History, which is a look at monumental events in the world of business and technology which has changed our lives. Now, the clue last week was it's 1984 and you'll want to block some time out for this game. Did you piece it together? Of course you did, it is the launch of Tetris, the all-time favorite puzzle game that was created by Soviet programmer Alexey Pajitnovv, who released it for use on Soviet-built computer architectures. Now, from there it spread to illicit western computers inside the USSR and then was, of course, quickly smuggled out of the East to Western Europe. There, a load of companies tried to claim they owned the code and format, spawning decades of legal battles. (15:19): Now, ironically, as an employee of the state, Alexey didn't see a penny from the game until the fall of the USSR, when he was finally able to claw back some well-earned cash for his creation. Next week, we're heading to 1983 and a pioneering journey. You know what it is? Shh. Right, that brings us to the end of Technology Now for this week. Next week, we're going to be looking ahead to the tech event of the year, at least that's what we think. Yes, it is HPE Discover. Certainly the biggest event on mine and Aubrey's calendar. We'll be discussing some of the exciting updates, innovations, and announcements from the event. And if you have even a passing interest in enterprise tech, then this is going to be one that you will not want to miss. In the meantime, do keep those suggestions for life-changing books coming in using the link in a podcast description. (16:08): Until then, thank you so much to our guest, Dr. Yaël Bensoussan, Director of the University of South Florida's Health Voice Center, and to our listeners, thank you all so much for joining us too. Technology Now is hosted by myself, Michael Bird, and Aubrey Lovell, and this episode was produced by Sam Datta-Paulin, and Zoe Anderson, with production support from Harry Morton, Alicia Kempson, Alison Paisley, Alex Podmore, and Ed Eveston. Technology Now is a Lower Street Production for Hewlett Packard Enterprise, and we will see you next week.