Chatbots & health apps - AI in healthcare

I wear my watch every single day. I love it. Look at how bad my turn is. This is Lauren. She's a final year nursing student. As I was going through different placements for nursing, I would see the difference between my heart rate and then the heart rate that I was comparing with the standards that we use for patients. And I would notice that sometimes my heart rate would go a little bit higher, sometimes not, didn't really think much of it. Lauren had been wearing her smartwatch for three years, but hadn't really taken note of how it was monitoring her health. I had all of the settings on mute just because of all the notifications with everything it was one listing that I needed to be notified of. But then some unusual things started going on. I had started to get a few bizarre symptoms, like really sensitive to heat, having some odd other things going on as well. And then my watch ended up notifying me that I had all these missed notifications with my heart rate. So I'd opened up the app and I'd seen that my cardiac output and heart rate had been kind of all over the place. And then there was a massive dip that was in the space of a couple of days that had shown that the trend had gone from being normal for the last three years to extremely low and abnormal. So I went to the doctor, showed them what my watch had shown me. So she did some blood tests and she did a scan and at that point they found half of my thyroid had disintegrated and it had all aligned up with when all of those symptoms had started and when my watch had initially detected that something was going on that was a bit bizarre. Embedded in the health apps in these smartwatches is an algorithm that can detect changes in the wearer's blood pressure and heart rates. It's certainly not a diagnosis, but these simple devices are allowing people to self-monitor things like diabetic or heart conditions, sleep apnea, medication reminders, and even signs of memory loss and dementia. All this thanks to some clever data science and AI. Welcome to Everyday AI. I'm John Whittle, an AI expert from CSIRO, Australia's National Science Agency. In this episode, we'll be speaking to a number of people working with artificial intelligence to bring about positive changes in the world of health. We'll come back to Lauren and her smartwatch, but first we'll hear how AI can be used in the early detection of breast cancer and how chatbots are helping both children and adults with things like Parkinson's disease, chronic pain and autism spectrum disorder. A quick content warning. At around the 24 minute mark of this episode is a mention of self-harm and suicide, so please listen at your own discretion. I think we're at now is just programming computers to do increasingly complex tasks, tasks that would normally have been done by a human and my project or the research project that I'm working on is a case in point of that because we're teaching a computer to do a task that a human has always done, which is to read mammograms, to read breast x-rays looking for the presence of cancer. This is Dr Helen Fraser. She's a radiologist, breast cancer clinician and AI researcher with more than 20 years clinical experience in breast screening, imaging and cancer diagnosis. I'm particularly excited to hear about Helen's breast cancer AI project where her team is trialling the use of artificial intelligence to screen mammograms. Radiologists in general, not just breast radiologists, we're patent doctors and alongside pathologists that also patent doctors, we're kind of like the two professions that are really going to be positively impacted by artificial intelligence. I think we've evolved rapidly in that space. How are mammograms currently read? So what would be the process that somebody trained to do that as their job would go through? What kind of things do they look for? Every mammogram, which is a breast x-ray, is read independently by two breast imaging subspecialty trained radiologists and if they differ, it actually, another read takes place, a third arbitration read, which is by a radiologist that's extremely experienced with very high detection statistics. So it's a really time-consuming quite a costly process. It's doctors that have done their medical degree and then they're doing specialty training and to get a fellowship from the College of Radiologists, which is a five-year program and many go on and do a breast imaging fellowship. So it's quite a lot of time learning the trade. Is it that you're looking for areas of the mammogram that's maybe brighter than others or how do I know what to look for? So when a radiologist looks at a mammogram, we're actually looking for very subtle changes, subtle signals that might be appointed towards a breast cancer that is going to either develop in that area or is already developed. So some have a very systematized approach where they scan certain areas with their eyes of the breasts and methodically go through it. But some of our best readers actually get the gist of something as soon as they see the image and they track their eye movements. They're a little bit more random and chaotic, but very quickly they seem to hone in on that area of abnormality. The opportunity here is that we're looking for patterns and we're doing repetitive reporting. And to me that all points towards some intelligence really being able to assist us. It strikes me as you said that it's a very repetitive task and dare I say a boring task as well. I mean I can remember when I was a university professor and around exam time we would have 300, 400 exam scripts to mark by hand. But by the time you get to number 200 or number 250, you just want the whole thing to be over and you're not really, it's probably no big secret, but you're not really looking as carefully as those answers as you might have been in the first batch of 10. So yeah, it must be quite hard to maintain your concentration if you're doing that volume of mammograms. Yeah, well you said that, not me, but you're exactly right. Look at every single one of them very diligently and make a wonderful decision each time, you know. Of course, John, that's exactly right. About 95% of the mammograms that radiologists read are normal. The other 5% will show some indication for cancer and will be called into the assessment pathway. That means that the majority of the time that the specialists on Helen's team are spending reading scans will be looking through mostly normal ones. And this is where artificial intelligence can come in. The promise of AI and risk cancer screening is in fact, rather than being a radiologist that looks at 95% normal, you know, perhaps we can start to use our human creativity and our skill sets more to spend more time with the women in those situations, those complex cases or in biopsy procedures, for instance, where we really can add a benefit. And so shift that a bit of a paradigm shift from using the human skill sets and the human creativity into the more complex cases rather than doing normal, normal, normal, normal all the time. The AI technology that Helen's team is studying is now being put to the test in real world scenarios to see how well the algorithm is able to detect whether or not cancer is present in the image. With the help of AI, the project aims to improve on the accuracy of screenings and ultimately to reduce deaths from breast cancer. By training an algorithm with a data set of mammography images, the AI can learn to detect cancers with the knowledge and skill of a team of the best human readers combined, pooling all the expert skills into one place. So how does this AI work? In this case, Helen's team trained an algorithm on a cancer enriched data set. That is a heap of past mammogram images where the signs of breast cancer were either present or not. Once the algorithm had learned what it was looking for, it was put into an app and tested with images alongside cases where the specialists knew what the actual outcomes were. The first thing they tested was using the AI to read each scammed image and filter out all those 95% of mammograms that are completely normal. And we did get really good results, but one or two cancers slipped through. But the issue with that operating point, being autonomous, is that we don't have a human in the loop. And at the moment, and this I'm sure will change with increasing improvements and developments over time, it's just not with us now. We really are hearing loud and clear that we need to keep the human in the loop. Then they use the AI to simulate the second reading of the image after one of the human readers has had a scan. So the results of the simulation on that retrospective cohort have been very promising to sit the AI reader as a replacement. This certainly saves the doctor's time, but it also reduces the stress and costs for the people going through the screening process for breast cancer. I'll let Helen explain why. Coming into assessment is a really anxiety-provoking and costly process. Many of them really feel that they have a cancer, even though the majority of them don't. But along the journey of discovering they don't have cancer, they'll have extra imaging, ultrasound, digital breast tumour synthesis, and sometimes even a biopsy. And that's a really costly process as well. So we improved on the accuracy by about 20%. We also would improve, I think, on the experience that, because we've used an algorithm to read the mammogram, that timely process of bringing in two readers independently and a third if they differ, we've taken out one of the readers there and our service delivery for time to result from a mammogram is two weeks. So I think really we could probably reduce that down to a matter of days. Two weeks is a long time to wait for a normal result. And I think an algorithmic reader could help us make that a lot faster process. And then the other area that improved and addressed one of the known challenges is an expensive program to administer. And breast cancer is increasing. We've got an aging population. We need to increase our capacity. And when we looked at the variable costings of reading an assessment, replacing a reader actually took out 50% of reading costs and a considerable amount of the assessment cost. And we saw quite significant cost savings. The potential of Helen's project is really exciting. But AI is already in use in many areas of health care. For example, there's a computer that can diagnose human heart disease, computer vision tech that can identify skin cancers, and algorithms that can detect eye diseases as expertly as human physicians. But while AI can be trained to detect things like abnormalities in scans, things can be a little bit more complex when it comes to our mental health. And that's because that, you know, human language is incredibly complex. It's also very sloppy. It's vague. This is Dr. David Ireland. I'm a senior research scientist at the Australian eHealth Research Centre within CSIRO. My background is in computer science and electronic engineering. But for the last 10 years, I've been working with speech occupational therapists, physiotherapists, building them all the tech that they want and researching artificial general intelligence. David has had an interesting career, to say the least. Before coming to CSIRO, he developed algorithms for detecting brain strokes and breast cancer using microwave radiation. He's an active software developer for mobile and server apps, and he's a kind of legend in the world of chatbots for health. My work on chatbots has exploded in the last seven years, ranging from Parkinson's disease, autism, dementia, chronic pain, genetic counselling, and people wanting to quit smoking. Would you like to give a voice sample now? Yeah, I can do that. Please say "heard" five times. Heard, heard, heard, heard, heard. Thanks for that. What you've just heard is a sample of David's personal favourite chatbot he developed, called Harley. It was originally developed for people with Parkinson's disease. I was working with a group of speech and occupational therapists. We were interested in the language difficulties people have with Parkinson's disease. It was just for conversing about what books they've read. Then we took it out to community groups and we said, "This is a chatbot that's going to talk to you and ask you questions. What are you doing? What are you up to?" It's going to look at your voice and it's actually doing a lot in the background, looking at how people are articulating vowels and the pauses in the sentences. People with Parkinson's will have issues with vowels, because that's when your articulators are fixed, but your vocal cords are vibrating. We're looking at how constant those vocal cords are oscillating. We're also looking at the delays, the natural flow in the conversation. David developed this Harley app on a $15,000 grant, but it had a big impact and generated a lot of interest in developing chatbot technology for other health areas. But David had his own personal reasons to be invested in its success. But then something interesting happened. While David was developing this chatbot to encourage his son to speak, he noticed his son interacting with another chatbot, one that's a little more everyday. I've got videos of him using other apps and speaking and interacting with Siri. I think because he was at school, he was learning things like spelling, won't say a word, and then come home and I could hear him, "Siri, spell hexagon." Later I found out there's quite a considerable amount of evidence that things like Amazon Alexa and Siri are really good for children that are having speech and language issues, because they can go off by themselves and try and copy what the device says, and that can improve their speech. This anecdotal evidence that David found was from reviews on Amazon products. David's team looked at every app they could find in Amazon store, then read through every single review. So we found all these reviews of parents saying, "Yeah, this, you know, Siri's got my son talking batteries, he always pronouncing words better." So it got me thinking about sort of long-term research goals was in the idea of virtual companions, and that's where the child or the user can speak to the device about any topic and then it would have a meaningful response. Have you been up to anything interesting today? We watched a good movie called The Zookeeper. Can you tell me more? It's about a zookeeper who works at a zoo and the animals help him. "Human life just sounds so fascinating to me." You may be familiar with chatbots that pop up and help you when you're browsing an online store or maybe looking to upgrade your insurance. The ones that ask how they can help you. You will also have heard them on the phone, those automated message directories that pick up signals in the words you say and convert them to directions. We talked about this a bit in the first episode of this podcast series. So what is a chatbot and how does it work? David describes a chatbot as a computer program that imitates human conversation. The person can type, speak, draw or even press a picture on the screen and then it goes into what I call the brain of the chatbot. The brain determines what the response is based on the context. So based on the current topic, what they previously said. So it's not one algorithm, it's a combination of multiple algorithms that fall under what we call natural language processing. Here's a sample of a chatbot learning the rules of language. "There's a big black cat outside my window." "Okay." "What colour is the cat?" "Black." "What size is the cat?" "Big." "Thank you." "You're welcome." However you interact with the chatbot, whether through speaking or typing or pressing a button, there are AI algorithms at work that interpret what you're communicating and then decide how best to respond. "The speech recognition has improved greatly and that's when the computer can hear someone speak and convert that to text." So on the one hand, David's team are training these algorithms with language so that they can respond in a way that makes sense. But they've also got to work out how they respond. It's like their personality, the tone of voice or the particular way they will speak back. "We'll think of the personality beforehand, we'll think of what it's going to say, how it's going to say it. I like to model it around graduate speech therapists that are young and quite bubbly in my smoking project. I followed a doctor who's, you know, a treating doctor for people with lung conditions and I recorded how he would explain things and then I would use his language to put it together." Getting a chatbot to learn a few responses may sound kind of straightforward, right? But what if you're wanting that bot to pick up on a person's feelings and emotions? Having meaningful responses is where all the work happens. So we'll also do a lot of, you know, training on sentiment. So sentiment is usually falls into sort of binary like a positive and negative and it can have, it's sort of the emotion conveyed in the utterance. And it can vary for different contexts. We'll see how that performs and then we might retrain the algorithm. So like I said, the chatbot is a combination of algorithms. Often have to fine tune these in health so that we have a nice sort of meaningful and safe response to the user. So for example, sentiment analysis, the chatbot might know, okay, the topic is about chronic pain. I've asked the question, how are they feeling? And they've come back with this big long utterance saying they're feeling terrible. The chatbot goes, okay, this is negative sentiment. This doesn't sound good. And then it can sort of decide what it's going to do. And it might be like, Hey, I might need to speak to a human. Or I might need to call someone or it doesn't action. And then it converts all this, its decisions. And then it speaks back to the user. One of the first AI chatbot systems was called Eliza. It was built in the 1960s at the MIT artificial intelligence laboratory by Joseph Weisenbaum. Eliza would give canned responses to users that would make them feel like they were engaging with someone who understood what they were saying. A bunch of variations were made on the original scripts by amateur coders. One of the most well-known was called doctor. In this instance, Eliza was trained to respond like a Rogerian psychotherapist. Doctor Eliza was made to sound like it was reflecting on user questions by turning the questions back at the patient. Here's a sample of the script that has been dramatized for an old documentary. A human, the patient speaks first. Men are all alike. In what way? They're always bugging us about something or other. Can you think of a specific example? Well, my boyfriend made me come here. Your boyfriend made you come here. He says, "I'm depressed much the time." I'm sorry to hear that you're depressed. It's true. I am unhappy. Do you think coming here will help you not to be unhappy? Weisenbaum originally designed it as a parody of psychotherapy, but what he didn't expect was that users would have real emotional connections to it. Even his own secretary reportedly asked Weisenbaum to leave the room so that she and Eliza could have a real conversation. After two or three interchanges with the machine, she turned to me and she said, "Would you mind leaving the room, please?" We've obviously come a long way from Eliza. Let's come back to Dr. David Ireland. One of David's biggest projects is working with speech pathologists, looking into bullying for children and teenagers on the autism spectrum. In this example, the chatbot is trained with responses it can make when it hears that a child is being bullied or teased. It can change its voice and personality and then roleplay with the child what they might say in response. There are other exciting outcomes of this research too. I've been working with a occupational therapist who's a pain specialist at the Royal Brisbane in Brisbane. She treats people that have had a trauma. That trauma might have been from a car accident, it's healed, but there's nerve damage so that the person is living with chronic pain. That's pain lasting more than three months. Many of them are on opioids and that can't go on for too much longer. In this case, doctors have to try to find out what behaviours the person is doing and whether there is a causal link to the medication use and the pain increase. It asks a lot of clinical questions like where is your pain, how long have you had it, so it's asking what a clinician would ask. It then provides education. David's not only interested in spoken language, he's exploring other ways of communicating too, particularly for those with chronic pain or special needs. Particularly children don't know why they're feeling pain. They don't understand that it is sort of a defence mechanism, it does have a purpose. When it comes to children with autism spectrum disorder for example, or who might be non-verbal for other reasons, David and his team worked out another way to interact with the chatbots. So I believe it's estimated about 50% of children who are non-verbal who can't speak have chronic pain, which is usually related to dental and teeth aches and stomach issues because of their limited diet. So we let the user draw, not only just speak, but they can just draw how they feel. We found that was very popular with children. We got all sorts of pictures come back. This was the technology behind a particular chatbot that David's team created to communicate with people about their pain. Its name is Dolores. We basically went around to pain clinics and different hospitals and found people waiting in the room saying, "Hey, do you want to come and talk to Dolores at our chatbot and tell us what you think?" So we got 60 participants, we got overwhelmingly good feedback, we got some really nice conversations. Some of the drawings were really interesting. The one I liked was a volcano erupting, which is people with chronic pain that is called referred to flare-ups, where it just explodes and bursts. We found red was very common with people that have higher levels of pain, which I believe is consistent with art therapy literature. The first version of Dolores couldn't really see what they draw, but they could see the colors so it could say, "Hey, that's a lot of red." The next version will be able to see what they've drawn and say, "Hey, that looks like a tree." The potential of AI to be used as a tool by health professionals like this is exciting. But we can't talk about them without also considering the ethical challenges this technology faces. Because what happens when an AI that can detect human emotion spits out a dangerous answer, or gives bad advice? There are plenty of stories of avatars and chatbots gone wrong. In 2016, Microsoft released an artificial intelligence chatbot called Tay. But the bot started posting inflammatory and offensive tweets through its Twitter account and was shut down after only 16 hours. More recently, Google's chatbot was accused of being racist and criticized for the lack of diversity in the engineering team that built it. Metta, formerly Facebook, have released Blendobot, which has been known to spit out fake news. Social media platforms came under heavy criticism for their AI algorithms when a British teenager died from an act of self-harm. A court ruling found that the negative effects of online content contributed to her death after the app fed her with content about depression and suicide. It's unavoidable that when chatbots are trained with human scripts, and particularly with data from the depths of the Internet, that they can never be free of human biases and the darker sides of human behavior. David Island's team are continually working on ways to avoid this. He says that one of the criticisms he often receives is that he's attempting to replace professionals with chatbots. The technology I'm trying to make is really just to augment and to connect people that might be isolated. I'm not trying to replace humans. A lot of the mental health modules that we build in all my chatbots, if it detects any issues, it can say, "How about we call Lifeline right now?" It's not just about replacing humans either. If an AI served a befalse diagnosis, it could potentially cause harm. I want to bring us back to nursing student Lauren and her heart-monitoring smartwatch. AI-enabled smartwatches or health trackers are definitely helping people like Lauren put the missing pieces of a health puzzle together, or alerting us when something may be potentially worth exploring. They do not take the place of a human doctor and certainly cannot provide a diagnosis. Here's Lauren again. My GP thought it was really cool, and then when I went to see the specialist, the specialist was like, "Oh, you can't trust the watch," and I was like, "Of course you can't trust the watch," but it was what prompted me to go to the doctor in the first place because here I was thinking I was run down and stressed, and then my watch was telling me, "Well, actually, according to its data and its AI, it was something more sinister that was going on, and whether that was right or wrong for a bunch of other patients could be completely different, but for me, that was what I needed to prompt me to go to the doctor and to take time off away from work," which I think nurses are very negligent in doing, so yeah, like it helped me and it was that that kind of really led me to the cascade of findings after that. I put the question to Dr Helen Fraser too. She said, "There's another reason that this technology is not likely to replace humans anytime soon." We thought that women might not be that excited by some perceptions of AI or that it's malevolent robots that are going to destroy humankind and that sort of thing, but interestingly, they don't necessarily want to understand the math of it, but providing it's evidence-based and evaluated satisfactorily and improves the program, they're really excited about it. The thing that did come out of the focus group is that they really expect for the human to be in the loop. They really don't want this autonomous co-pilot arrangement and most of them came back or articulated that they really would like to have a radiologist directly involved. AI's use in many industries has come under criticism for perpetuating human biases and health can run the same risks. I asked Helen for her thoughts on how the demographics collected for things like training algorithms can potentially impact the data. These are of considerable concern to us and to groups globally and we are in an opportunity where in terms of biases that might be encoded in our data sets, we're very cognisant of that and we do capture information in our demographics, country of birth, language other than English spoken at home, Aboriginal, Torres Strait Islander women for instance, so we haven't solved the problem, but I think we've started that journey of being cognisant of it. Again, I always come back to the strengths of our data set. I don't know any other medical imaging data sets that would have that demographic information that enables us to see does our algorithm perform equally as well in those cohorts of minority groups of women and the generalisability, does it generalise to those cohorts of women, which is so important to us and is definitely works that will be undertaking. If implemented in a responsible way, it's exciting to imagine just how much this technology will advance into the future to help take the load off for health professionals and give us more power in monitoring our own health. As we develop legislation, public education and regulation around AI, it will continue to become safer and more ethical. But ultimately, artificial intelligence will always be better when used as a tool in collaboration with doctors and professionals. I'm John Whittle. Thanks for joining me for Everyday AI. Up next, we'll be hearing how artificial intelligence is being used to restore ecosystems and how anybody with a smartphone can contribute to science projects that help scientists better understand the populations of animals across the planet. You can detect and identify whales that are moving through and through automated recognition of whale vocalisations and real-time detection. There's actually the opportunity there to prevent ship strikes. We can also be monitoring fishes and amphibians and a whole suite of species. I'm really hopeful that in the years to come, we're going to see this technology be applied to many conservation questions across the globe. I'm John Whittle. Everyday AI is a CSIRO series created by me and Eliza Keck. Alexandra Persley is our supervising producer and Jess Hamilton is senior producer from AudioCraft. The AudioCraft production team is Jasmine Mee-Lee, Cassandra Steeth and Laura Brierly-Newton. We'd love to know what you think, so please subscribe to Everyday AI and leave us a review wherever you get your podcasts.