Precision Health Engineering - Designing Health-Centered Mundane Technology to Increase Adherence
You. It, is my pleasure today, to welcome back Pablo Perez who, is an intern in my group in 2011. Ich well ok 2013. Was when we published the papers, but, he, was a, wonderful. Colleague, for us and many of us enjoyed working with him he, now. Is a faculty. Member at Stanford, University. In the medical school and he actually started a lab called the pervasive, well-being technology, lab which is near and dear to our hearts, and he's gonna be talking about precision, health engineering, today so I can't wait to hear what he's up to so thank you well. Thank you so much Mary for having me here and thanks for attending. This talk. I'm very excited. Yeah. So I'm here to talk about precision, health engineering. And um you, know I hope I can convey the message of let you know what's the the, new wave of understanding. Of what health means in. A second so when, a daily basis you will see like people who are like overworked. Or. A tired. Angry. Overweight. Unmotivated. Distracted. And. Happy. Feeling. Isolated. It. Will be the the. Dark side of of you know what mental health like, there's. Like 1 in 4 people who. Have some problems with mental health we're talking about a 2.5, trillion, dollars per year problem. 1.7. Of those are due to productivity, and absenteeism issues. So this is a serious, thing. And we, really need to understand, how to attack this so when, I talk about precision health I'm gonna focus a bit more on the mental health aspects, today it's definitely, a broader, concept but I will be focusing, slightly. On mental health and stress so, I'm. Gonna be talking giving. A definition I want, to basically say that precision holds not precision, medicine and I'm gonna spray that explain that in a second and, I'm gonna be also giving an overview of, more. Applied se, perspective, or precision health which will be what I call mundane stress, management, basically, how do we use very mundane objects, like PC. Mouths or or, display. A car. Or things. Like, messaging. Apps or just any other apps that you use on a daily basis in the office the car in the home and how to create, how. To do research across these different environments, which. Is what I've been doing so far so. Precision, health is in, one definition of precision holds the practice of personalized, health so you want to personalize, things and you want to basically be. Understanding. Health but what is health like let's understand what health means if you look at the definition of health by the world. Hair organization, it's not about sickness, we're like overwhelmed. By the perspective that Held's, about sickness and it's not about sickness it's precent, sickness, but. So far it's been like too much emphasis on, sickness, and infirmity and, that's great I mean we have to basically to support, this type of research but not enough on like, how do we keep keep your well-being, your physical, mental social abilities. Up to, par how to monitor, the baseline, of this, so when, I say that precision holds not precision, medicine what. I basically mean is that you know a, good analogy will be like you know if you look at humans they basically, have a sampling, frequency of, one visit per year and sometimes. Even less than that to try to monitor, and try to do detection, and try to do you. Know prediction, and prognosis and, most, often. Than that where we get slate detection, of a bunch of problems now. If you look at the jet engine, you, know in the aerospace. Industry. Well. It's very different you have continuous monitoring, hundreds, of sensors hundred, of super. High sampling, frequency, and that's, what you know keeps, this like you know high you. Know availability. Industry, going so, the hope is that we can transform. That and have the human, being monitored, constantly and, being, able being, able to really have, early failure detection and prevention, of individuals. So that's what precision, helps more about the baseline, than. Any type of anomaly and I, will say sadly, the med school and med schools in general don't really know how to deal with this problem like, because they have been studying sick people they, have not been studying, healthy, people right. By definition, so. At the end of the we still need to know what, it means to like do this monitoring, on on, a daily basis how to measure up an and for a long-term and this is not easy because it's not just about like putting sensors on people's like how do we get people to stick with those sensors how do we get people to like basically.
Understand. The value of like, prevention, and prediction. And proactive, medicine so. If. You look at you know that the state of the art of sensing, you know if we focus on the sensing part first you, know we have like the the wearables, that are, everywhere but there's a huge, drop out rates and this is what I'm trying to talk say, about like you know adoption right it's, it yeah we have great sensors, and some things that are happening is this is actually the Matt's cold is implantable, but, the dropout rates are still pretty large right, and many of them are not because they learn to do something or they are better at using where I was because they forget to charge it right or things like that or it just don't like it. You. Can also say like well how's their sensing, in other areas like mental health well sadly. The state of there continues, to be. Subjective. Monitoring, you know asking question to people more often less often whatever, about asking question it's, not a you, know at the most the, best way to to collect, data, and it's also very tiring, for the individual, and, finally. You know if you say like well you can do electrodermal, activity higher. Variability or like new modes, of sampling, they, continue to be obtrusive because, you have to learn to wear these things and you have to basically have to charge them and do these other types of things so so. Our hope is that you know we can move beyond. Just. Sensing. With wearables, and sensing with questions, and start looking at their environment as well and this is one, of the corners, of the projection, of what it could be the, future of you. Know precision, how, do we use houses. And very mundane things, around. Us to, generate, these let's call it health center Internet of Things if you must right, yeah. How, often if you go back a slide how often do you have to charge those shirts. You're. Not really and often. It depends on you know the type of sensing, that you do you, know many people advertise, heart rate variability, and what they do is they sense once that the once a day, or only, when you're sleeping or when you're sitting so, the reality is that if you really want to have this continuous monitoring, there's no good, battery solutions, yet for this continuous, monitoring you still we haven't solved that problem right, I'm. Wearing this ring for example which, is actually useful for sleep, tracking and it's, not bad I have insomnia, I need it but, it works only at night and it's good it's a use case right but the rest of the day is that, doesn't, really know what's happening with me but, we're getting close but still, you know form-factor mother I do wear jewelry. So it's better than wearing some guys don't wear other wearables except for these one because I have a torn ligament. So I to wear something but. That's a different that's a sickness that's not a help thing, so. Anyway that the end of the day yet is a good question battery's not there yet so. Can. We move. Not, only from sensing, can we also use, the. The. Infrastructure, to do they even behavioral, not just how do we maximize adoption. And adherence, can we get to the point that we can really like push, that and and I hope is that you know we can understand, why is this important, you know if you look if we if, we look at this graph that you know I've created between, like what I consider suppressive. And whether considers, engagement, engaging, I would say like you know micro, interventions, and wearables, should, be like you know not so high not so low obtrusively, speaking, and you, know they are actually short in the engagement party HSR shortly they died rather quickly, you know the wearable the intervention, people use it and then they discard, it and they, don't keep engaged for the long term. Sorry. We would say like you know not just and, and. IOT might. Actually give, us a bit less, obtrusiveness. Because, you know we're like not asking, you to wear anything and maybe, we'll increase engagement and, of course if you get to the point that we actually do, like implant saying configural, environments. We, will get to the point that you know it might actually have, higher engagement because.
It's Really in your body, now but, it's arguable, how you, know at Rousset this can be. It's. Not clear. To design interventions to, make right. You have to design interventions, to get people to use them and and deliver values over the long term and that's actually part of the challenge so. In, this, talk I'm gonna talk a bit more about this subset, of I would say like you know the ecosystem. Of potential. Interventions, I'm not gonna talk a lot about this one even though I I've, done some research but I want to discuss some ideas on this area they're, a concept that I also want to bring you know as a baseline. Of this conversation is this idea of like you know what. Is health for people like, why. Do healthy, people think about health and the, reality, is like healthy people were high, level of health don't. Really have a huge motivation, to stay or become healthy they, are motivated, to be beautiful. To be interactive. To be attractive, to be smart but, they are nothing like oh I'm gonna keep like my cardiovascular health. I'm gonna keep, my like mental, held up today that's not how people think like. They think about the products, of being healthy not about health and themselves and it's, only when the cheerier ation happens, that, people start, really getting like you know when you're hell double goes down that motivation, goes up and be like okay now I really care about health because, something, broke because something is now going well so health becomes, fundamentally. Important. But. You know precision, health works with these people and when I talk to my colleagues in the med school is like these are not, patients, right these are individuals. These, are patients, these, are individuals, they have very different levels of motivation therefore. Also very different psychology. So you cannot design the same way that you design for these people or these, other ones over here and. You know what we hope is that you know we can create engage in design to like, you know bring people, to keep them healthy and find. This sweet spot between engagement. And efficacy. You have to be efficacious about, sensing, and about the interventions, that you deliver but. You have to also be engaging, right. And and I call it a sweet spot because I think it's like, not. A trivial way of finding a very efficacious. Sensor. Intervention, and very engaging over, the long term these are challenging. Problems. So. One approach I've taken you know through my formation, and my, research. Is like combining, a few philosophies. To try to like do. Something about us one is design thinking you know for the creative, aspects, combine. It with the scientific, method in order to postulate. Hypotheses. But, once you have also explore, what are the needs of the of the user in, order to be creative but. Also to have validated, ways, of sensing, and intervening, right but, you know for those of you who have, experience. In both ends these are not necessarily. Like, the philosophies, are the, people who practice the philosophies don't necessarily, mix that well right these are like, sadly. Even though they shouldn't, somehow. They just found themselves, at odds many times right in the way that they formulate, questions and, problems, I, think, we should definitely train, people to think in the intersection, of these things to, create these engaging and applications. A sweet, spot I also, used tools like machine learning you know sometimes, it's very important, I consider, it more like an aide rather than an objective even, though it's the emphasis. Right now is a lot about machine. Learning but then the other component that's also very important, to me is implementation, right. There's a huge, field that's nascent, in the public health realm called implementation. Science right. There's an average of 17, years to get like a, good. Validated. Finding, in the med school in, the actual market like 17, years even, though it's been validated, for like, many many iterations so, we're, not really that great at doing you. Know inter, implementation. In the clinical world much less implementation. Outside have. In the health domain in. Other domains are very good like yeah Microsoft, Google all of you know how to implement products, that lead in the work for a long term and that's what we're what we need to like learn from each other in order to keep doing like you know all these like things and machine, learning helps, us with the predictability and, implementation.
Science, With its credibility. So. I'm going to tell you a couple of like projects. During this talk one, is the sensor, licensing which is the. Idea that I can like bring some ideas that we created with. Sensing, an AI bring. It from different domains like the office, into. The car and. And eventually, also, get. It to, work actually in a real car in a real environment. So, we keep evolving, towards, implementation and, then, this other domain that we call the desert project, broad project, in my lab is called subtle, interventions, which started, more in the Design Thinking area. Got. Validated. You know in a simulator, and now we're actually have. Revalidated. It on the, road see, revalidation. Is a very interesting word I'm gonna just make a small, parent this is if you go to the med school I remember the first thing that I got asked, when I joined the med school like almost two years ago is like how's. Your thing. Being revalidated. I'm. Like no because you, know Kai only, wants us to publish the monster innovative, things I'm like well, once you revalidated, three times then, we talk, that's. A med school right. They, don't just throw, things that have not been validated, and, revalidate, for them for them revalidated. Revalidation. Is very important, so. Anyways the procedure, helped design space, for, me also has another challenge. So. Combining, those those, things. I told you I think there's two types of like tasks. That you should go to do procedure health one is the exploration, part and then, the in debt validation that, we discussed but these are two different types of people, and two different types of projects so in my lab I have projects. That I will consider more like shallow and they are in the creativity, phase and I have people were very creative but, then I have projects. That are going depth and revalidate, and, like redo study. With very different type of people for example this central is sensing the type of person that I'm hiring now is like an expert, in bio bio, mechanical. Manipulation. Of, the body he. Doesn't know about design, but he knows exactly how, the body moves because. We're revalidating. The things that we have been finding about that on or, inside all interventions also we're hiring people who experts. In like you know machine learning and IOT to deliver, the interventions, well, I have designers, and other people doing that thing so the combination, of these two things I think it makes the lab very interesting makes, it also hard to manage rights are very complicated lab to management because you have people were like let's. Scratch the surface a lot. Like as. Long as you don't prove this I will never believe what you're telling me so, it's a very interesting creative, tension, let's call it that way and. I think you know Mike Rosario is one of those places where you also find a lot of these things super. Exciting, so, very briefly I've about me chest for the sake of you, know explaining, why am I doing this thing where I come from you know I I, gave.
Up You, know my business, career because, of family experience. In. Mental health and I've been studying mental, health ever since I joined the PhD, at. Berkeley and now, I'm doing what I like this well-being in precision health but, I'm a CS. PhD. Working, in a med school so, that's an interesting combination right, I, was. Hired to be part of this new center called the presidential health and integrated, diagnostic, centers abroad center, where my lab resides and it's, about trying to discuss these problems like how, do we design new things how, do we do the behavioral, economics, aspect how do we do the economic, about, economics, of adoption, of like proactive, health and it's, a very very new like way of thinking for the med school they are learning how to do even, with me like like we still don't understand you that well but we're starting to get something out of you. Just. To clarify. When I say well-being is a very broad topic there's a lot of things so, what I'm gonna do is over seem oversimplified. Tremendously, in the and in the focus of for this talk and talk, mostly about stress, management that's, naba all means what, well-being, means well-being is a very complicated large. Construct, that's been discussed, in, many ample. Terms in public health but I'm gonna focus on the stress management so so, let me tell you about stress management as an operationalization. Of life. Or the research that I do so remember. All these people that I showed you know they they went through all these things so, what, do these have in common right I don't know if you can help me find something. That's common, you. Know if you if you look at them whether, they are like you, know happy or angry or whatever. Most. Of these people also, suffer of stress or arousal arousal. Economic, arousal, though it's a driver for, like pretty much on any other emotion, and the, truth is that stress right now it's becoming a really complicated, problem, in America alone is 300, billion dollars. 80%, of adults have reported, to have issues with stress and not being able to deal with the stress and it, basically affects any part of your life like, if you read the fine, print something's, gonna break in your body if you talk to so, an internist, or internal. Medicine person. Will tell you like something's gonna break, we. Don't know yet what's gonna break and that's something that we have to actually start understanding better with more sensing. But.
There's Another commonality, also, across these people I don't know if you can spot what else is common, across, all these pictures here. Something. Specific, No. Well. These, people are indoors, so. We became an indoor species, we. Were not designed to be indoor species, look, at our arms and our legs we're, runners gatherers, hunters, were, like designed to move and, we. Spend ourselves sitting, right. So you wanna find a human find a chair right there easy, so. 93%, of the time indoors that means that this is not our new ecosystem. So. You know at the end of the day we have to attack stress in its ecosystem. So. Anyways. Stress you know is this very interesting thing that we talked about like challenge, and thread and you know how can, we cope with that thing can we do it that's how our body, produces. Resources, and he gets like tens because we're not you know, mammals. And we produce like muscle, tension to get out you. Know we have the year it's Dodson law that tells us that we need some, level of stress not too much not too little to, operate, and you know it's been proven mostly with animals and a little bit with with, humans and, Sapolsky, that tells us that you know stress. Is very useful, and zebras, don't get ulcers because, they use it very well once the lion, come they, are the zebra, runs and you know that, helps. Your metabolism to basically dissipate, stress and that's exactly what it was designed but. The truth is that you know these cycles, of like stress. And distress stress, and distress which are very normal in modern, times, because of the frequency of the stressors, they, are becoming actually, a problem. Because of these repetitive stress. To regenerate, our outer immune system and other types of problem so. At the end of the day we you know a stress is something that needs to be managed, not eliminated. Right it has to be managed part of life's part of.
Support. May. I have a question another question for you so we have the zebra that runs and you know that doesn't get ulcers but. What are the zebra need. To. Be able to actually do, and. Be efficient, with the use of stress. Pretty. Without. The Prairie the poor zebra might actually die, of, a. Heart attack if. The lion comes and there's nowhere to run oh, that, so. You know stress, management in the wild works because there's this pace because you, move because, that's what we do when we're stressed we get tense. Our muscles, get jacked up to defend with a fight-or-flight. But. Look at our modern, wild at 93% of our times we spend in these environments, right. So. Where do we run. How. Do we move, how. Do we manage our stress in, these things so. Then you will have a philosophical, conversation, oh let's. Get out and those do exercise, I'm like sure. Fantastic. Now. You spend 93 percent of your time here, so. You're basically designing, for the seven percent oh we. Want to convert. These people okay if you're super successful, say you convert 99, 9%. 10%. Of, those 93 percent still, you have 80 percent people, who basically do this every day so. Why don't we redesign, these spaces to. Do. A stress management, in. The wild this is the new world this is our wild. People. Would say no no that's not no we have to get out of offices no just. Accept it suck it up offices. Are the new wire for humans simple. Let's, the redesign, them to make them like you know viable to do a stress management. The. Problem is that you know right now stress management it's really like very lame like you know there's a lot of stress-related. Visits, to the primary care and only. Very few get, some support from primary, care like almost nothing they. Don't know where how to do it's true they send you with an Advil for your, headache right. If. You say well, then the mental health wear has to be taken care of so mental health we say one quarter of population, has issue the truth is that about 10 percent of them them get referred, to a some, kind of like supporting, institution. 20, percent of those get diagnosed, with something one-third, of those complete treatment and one-third of those do not relapse so if you compute. This like beautiful. Like cascade you, had a tremendously. An, efficient. System. For mental health doesn't work either we. Say okay not on a wait a minute it stress the zebra doesn't go to the doctor is your problem, so why don't you deal with that that's put, it back on the user so. What do the user says so you know at the end of the day what, do you think are the two major reasons for lack of good stress management. In. The new wild. Time. Lack. Of willpower and lack of time yeah. And these are like you know a large. Cohort, of a thousand, people that APA did, yeah. So we're. Also very bad at managing there's no time so we need support we need support so for me this mundane stress management is like I'm gonna take the the, bull by the horns, and like I'm gonna design for the 93%, of the. Time that you spend in these places offices. Cars and homes and. I'm gonna try to see a way if I can do a make, a chair into something that works for a stress or a table. Or a. Lamp. Or something, that's already there and then. Learn from each other. So. So, let's start with with, with these two environments first, the office in the home to tell you some of the things that we've been thinking and doing and then I'll tell you more about the car so. This was the first study that you know inspired some of the work we actually published this paper a few years ago and like can, we sense stress. Just by the way you move the mouse this. Is a five dollar mouse from Walmart, no sensors, at it just look at the way you move x and y and then. Try to create a biomechanical, model. Of the, arm. Which gets affected, by muscle, tension, because that's the. The normal reaction, and proved in the damping frequency, of a mass, spring damper model, of that thing will actually increase. Because. It's correlated with the K coefficient, of your muscle tension and we proved that and we show that you know it is possible to do that, we. Actually what, we did is we created a bunch of like punt but, bunch of tasks. Like point and click or drag and drop or steering and we, replicated, essentially the Fiedler experiment, and then we show that for every combination of distance and width. For, different targets we were able to actually find, a, difference. In a stress and no stress this is the mean without, necessarily, there are bars and you're, actually able with about ten random samples, and a simple, machine, learning.
Classifier. Together, already seventy percent accuracy, with a very simple, tool. Which is damping, frequency, and the, way that we did that is using, a. Tool. That we borrow from speech processing, the linear predictive coding in. Order to do the generation. Of an, old poll model. Of your arm and try to establish, what, is the the. Major component. Of oscillation. Of your arm that's how we created this thing. At. The end of the day you know this is very simple you know it's just software it's really not heavy in any way and it can be implemented, in about. Four, billion mouths, that are already deployed in the world like, PC Mouse five dollar from Walmart right. We. Also recently, moved, and did. Studies. Replicated. Same study for touchpad, now touchpad, is a very interesting, problem it's way. More complicated the mouse it, has the extremely, complication, in the way that people use with, two fingers with two hands in different, ways also the arm and the fingers are way more complicated there I'm sorry the the hand and the wrist and the fingers are way more complicated, in an arm and it, requires, like more analysis, but we were happy to find that for, all the conditions all the combinations, of distances, and widths we, found actually, a difference in one of the features which was the area under the finger so. The amount of like let's, say as. A proxy for pressure, the, area the finger seems, to be clearly correlated. With. The amount of stress that you have for. Touchpads and where this papers in submission, right now oh these. Are the values, actually so we got people to stress with a stressor, like doing math and then, we measure the level of stress and then we basically show that, is area under the finger worked. And. Right now we're actually working we're, very proud we, got approval, to implement. Our technology. Both the mouse and. The. Trackpad in a clinic in the Hoover, Hoover, pavilion, is one of the clinic family medicine clinics, so right. Now burnout is one that, huge problem, in. The Med medical, domain where now there's a prevalence. Of higher than 50%, I'm. Not going to spend a lot of time but essentially we're now there's a combination of stress and dehumanization, and the humanization, happens, because of the, cycle, of production for, new patients, to be treated, so. Doctors, don't have the time to develop. An empathetic. Interaction. And that dehumanization. Let's call it that way generates, a lot of like emotional traits, and stress, so we're, going to be using the mouse and. Also, the trackpad, and all the computers, that practitioners are doing and where this. Is going to be longitudinal. In. This high risk population, now, one of the interventions that we're actually testing right now to try to modulate, stress, is, how. Do we enable movement, right so, movement is in or what that's whether the the. Zebras. One, way of looking, at movement is the. The, past 50. Years of. Studies incident, tourism, have, been focused in what's called moderate. To like active, movement. But. Now more recently there's, a new trend to try to look at what's called the low intensity, physical. Activity the, leaper. That. Means like even like changing, posture or moving, a little bit it's, predictive, of your mortality. So. What we implemented, was this desk, that, we call the haunted s that basically, that, thing that moves on its own and the, reason we created these robotic death which I know it's not look silly it's, because you no basically.
Sorry. We. Investigated, that 70%. Of users. Who. Start, doing these days. Remain, sitting, after three months so after, three months nobody uses this instant desk anymore just sitting there and then. You ask them why and it turns out that you know about, like 70% say like. Forgot. Apathy, I'm like, okay so if you don't care then I'm gonna start moving it for, you and see. What happens so, we. Basically did our early. Analysis. Of you know what, is the reaction that you have to Emanuel versus in an, autonomic, desk. Within a small cohort. Of 20 people and we. Probe to see how they felt about their stress and thing and we found no differences, between the, two they. Do become, very interesting, very, interestingly they, become very opinionated at the end when you say like, give. Me a reason, between the two they, will tell you very strong reasons about like I don't want to be remove the control how does a desk, there to tell me what to do and you, know I want to have control but. When they are doing the thing you would see people like literally, sitting and then going up and. I, didn't pay attention to the desk so. We're going to test these like concept, of what we call non volitional, behavior, change we're going to force you to move no. Persuasion anymore we're going to push in the limit like number - no -, see we can get you moving and we want to do it in longitudinal we're, going to start a longitudinal, study now in January and we'll, report, back you know we, hope that you know we can train a neural network that's. Very adaptable, very good that's actually in our brain did your network, to, basically get you used to go up and down without like you know complaining anymore right as opposed to training a super advanced machine. Learning thing, to, find the perfect time to move the desk there's some papers I have already investigated, this it's a very complicated problem, you want to make it perfect it just doesn't work so we're going, to go and in your face just move it and see if humans get used to that, so. That's one of the movement, things that we're trying, the. Other one that we're trying is like can we actually regulate breathing in a very like peripheral, way it should not be called subliminally, peripheral, peripheral. Way we're developing. This basically. Is a plug-in. That activates. A, breathing. Edge around, your, browser. That's. Linked to your breathing rate and as you start, like you know getting stressed it tries to bring it down to the same level. And it's very powerful you keep working you. Barely, notice the thing and we already have collected, like some preliminary, data that actually people do. Keep a lower breathing rate chest by the exposure to this very soothing. Frequency. Of the. Peripheral, vision we're. Actually right now investigating, what's the right application, because if, you study the the the. Visual, cortex the, peripheral, vision actually has a stronger, connection to, your parasympathetic nervous, system, as opposed to the the foveal vision the, part of the peripheral vision is has a stronger, connection so there's a huge potentially. We can activate, the, periphery. Stimulus. You might actually have a great chance so we think that this, will be even stronger you have bigger monitors, you know where you would really activate, your peripheral vision as opposed to like smaller ones or maybe in the car as well you, know maybe in the thing you need to we need to work with your rods the, rods not, with the cones right the cones are the ones that basically gets you alerted and like kind of aroused the, rods can help you calm down. The. Other thing that we have been doing you know is like this intervention sweets this, is actually the work that we work with Mary here, the the the the. Puff therapy, right how do we do this micro interventions, that was the early exploration, of this thing can, we get on a stressor, and use some kind of contextual, information to, provide a recommendation, of, very.
Mundane Things, I remember, we went on and say like we're gonna pick the most mundane things but the ones that are popular though so. We went on and look at the very popular things, that use online that. You already know how to use so, I'm not going to teach you anything new I'm just going to get you going to, do exactly the same things that you do and how, do we like recommend, with a machine learning tool. What, is the best thing to do at certain point we. Created, and sorry for the busyness this is part of our paper we created a suite, of interventions. That, basically, were essentially. Two parts a prompt, and a, link and we say like go, do this in facebook like for example one of them was like go and look at your like time line. And just look at a good part of your. Your. Life and that's, actually, highly, correlated with, the three good things technique. That you seen in, in. Positive. Psychology so, we found actually we work with the clinical psychology, to find an intersection, between the popular, apps that popular, mundane apps with. Popular. Techniques, in psychology, and it's, amazing actually many of these things I would say people I would, say they are using it already, with a purpose of like stress management right without them necessarily calling them that way we, will use a bunch of things from the phone in order to train, our contextual, bandit and try, to basically deliver, these things over time and. We were we, ran out for weeks you know with a bunch of like some things that we did daily, weekly etc. And we did this construct, of 2x2 between, random and machine learning recommendation. And also the possibility, to select or not select and. We eventually saw that you know there was a lot of personalization, that happened across all these different interventions, with all the different participants, we had like I we. Started with like 80 or something winded up with at, 27. Right or I'm not remembering so that dropout was equivalent. To to what you see in other apps but you, know we saw like there was personalization. And if you see an aggregate of these we we basically the random. Recommendation. Was this, versus, this is what we ended up at the odd week for after. Training several times the context of and contextual, bandit and the, funny parts that when, you asked, them separately, to people, which ones they preferred they were highly correlated with the ones that the the algorithm, picked for themselves -, one. Thing that we observed was very important was this novelty effect this, graphs a bit busy but and. Also pay attention to, the scale turns out that you know, people. Started. Testing, for those who were in the selection. Section. They were testing, new apps I wanted to explore, instead, of like exploit after, a while they started, becoming a bit more like used to like you know what they were you, receive from, the algorithm, but, after two weeks they started, again exploring. Exploring, experience program like what happened we supposedly. Already, identified, the best interventions. For you but, now you're not happy, why not then we did interview, a couple and they say like I'm. Bored, yeah. Okay breathing is good for me moving like doing Yasha what, else is there out that's. A huge challenge I would say for in the wild design, the, boredom a novelty is really a complicated, problem, that needs to be attacked and I think it can be done systemically, with algorithms as well, the. Other thing that we found very interesting was this paradox, you know many people who dropped early, in the study dropped. Because, they said, your, app stresses. Me up I'm. Like no way there's a stress management I'm, supposed, to help you not to stress you up they.
Say Well the thing is that once I installed that thing in my in my phone I keep, being reminded that I had the stupid app and reminds me of my stress and how I stress I am and I'm like so. Awareness for stress management is not trivial, everybody. Thinks aware this is the beginning of all health interventions, I disagree. For. Stress you have to be very careful awareness can actually be an accessor Bayesian factor we, have to be very good how you design these things so. One of the extensions. Of the work that we did with Mary is this idea of puppets this is what we got funding from, the human center AI, movement. In, Stanford. So we're actually doing a, similar, idea but, with, micro, chatbots and what we're trying to do is basically create a family, of chocolates, because, those could not understand, their will they the, word of like childhood. Development, for, mental. Health it's being very hard the lies that at first you, know mental health Chadwell was created in the 50's, what. Happened 70 years and we haven't cracked the nut it's. Not a trivial problem to do because you know it's very hard to keep track of all the complexity. Of life and mental health basically, about life so, what we did is we proposed to work together with dan jurafsky who, is my my, my. Researcher. A professor, in linguistics, it's like let's create a bunch of ups a bunch. Of micro childhoods and then then let's create a, recommendation. Layer, that will recommend. The. Childhood that works for you for the different like you know context, and see we can find a posse. Of chatbots that, will work for you and discard those who don't work for you so. In that case you can actually like you know do some kind of like trimming we. Basically started, with a wizard or a bus to validate if this idea of having multiple chapel works and we basically had variable several. Chad Watts versus control which was one single chat board and people, found that you know sorry this should have been the other way around people, found that you know having multiple childhoods actually a bit more. Useful. Than having a single childhood, we. That's the very beginning when we started with a wizard of us simulation. But then ever since we actually deploy the, app that we created in telegram, and we. Basically. Did it where we applied most of the money in creating this app that basically developed. The, delivers. This this. BOTS and we have been testing. It with students, with non students, and now we're going to do a larger cohort, we've. Observed that you know most, of the stressors that people report when they talk to the bots are about. To work in school and productivity, and some, with relationships. Etc so, we need to learn to like you know identify. The things and the other thing that we observed is that by, you. Know asking them was it helpful or neutral, we, are right now like in a complicated. Points, because we there's people who find the bots useful, but, there's also people who didn't find the work the, bots useful in this but we start investigating, we realize that you know there, is very important, problems, here that need to be attended especially, from the NLP perspective, one of them is parsing. Of the stress or when the you ask when the bot asks like how are you doing what you're stressful, and somebody says like I'm stressed because I'm running late and you know I might get fired or maybe get like you. Know punished, or something. What. Is the stress order is it that you're late or that you're might be get laid or a compounded, of those two and if, you understand, how to parse and basically, talk back at the, user exactly. What the stressor is. The. Empathy, that the rapport builds up immediately like oh that the, but understands, me if. You just say oh sorry sorry, you're stressed it's. Very rough the interaction, so parsing, a stressor. From an open text, question, it's.
Not Trivial because the knowledge domain of stressors is huge, so, we're working on trying to construct, a very large database to. Try to start parsing these stressors, in a. Much, more efficient, way from. An NLP perspective. The, other one is we have to answer this question can, I provide, support sometimes. People want to use these chat BOTS even, though they are designed for daily stress management for. Things like you know like, my. Son. You, know it's suicidal, and wants to like very, complicated, problems, now, those level. Of attention. Of treatment. Cannot be dealt with like a simple, childhood that only gives you simple, therapeutic, advice those, should be escalated so we need to determine in, terms of severity and complexity, can we support and basically, escalate, those where we can and the, other one is that's very important, we cannot just launch this system in a random way like we launched, the pub therapy, improv therapy we launched literally. With our random, you know assignment. Like traditional, contextual, bandit and people kept trying here. They create a very strong, you, know connection to the very first conversation so, the very first conversation, with the user the, very first bot that a test is not great. That's. Pretty much the end of the of the, engagement. So you cannot, just start with a random. Based. On one so we're creating a free, train run. Recommendation. Based for this problem. The. Other thing that we did with the puppet that I think it's just work quickly, sharing is we, did some qualitative, analysis, and what we did is this card sorting technique. Where we asked them to like take some cards, that look like you know with the stressors, in them and try, to ask them to like put. Them in different boxes that had a different, but. So I said like for example what kind of blood do you want for financial, problems oh I want the problem solving one what, do you want from relationship. One eye one the humor one, and we asked them to distribute them and to talk about that and that was all very interesting but then the part that's more interesting about the whole thing is that you, know we then added humans, and we say okay let's say you have humans, and BOTS and you. Have humans always available so forget about availability. Redistribute. The cards after they have done the bot and they. Don't didn't, put everyone, in the human side they left some in the bot side I mean like why are you leaving but humans, aren't aren't, humans supposed to be better to, provide support and all that like. Yeah but no sometimes, they are not better sometimes. They take too long to give you advice sometimes. They are like I have to like engage and go through all these like conversations chess to tell them what's going on sometimes. I don't trust them they. Want BOTS in their lives people. Want these BOTS as part. Of an ecosystem of support. So that's why we're trying to work with these ecosystem, the. Delays part is fascinating, we. Also tested, we learned when, we were doing this like qualitative, assessment, they told us we want the bot to have a delay I said. Why do you need a delay I'm like well you know the promise that I send my my stress during like immediately, responds back to me and I. There's other people who have seen this before too like but it's a pot it's already, processed, do you know two milliseconds, I know it doesn't matter I know but still I need the, body to. Feel, human, so, that need to have delays and feel human is a very interesting one some, of them wanted to use multiple, BOTS they say I don't want you one but I want to start first calming myself down and then, I go into problem-solving and then maybe I do some breathing regulation. The. Human body was a very interesting one this. Was love, and hate relationship some. People love the humor one it was like a simple joke type of thing and so, people hated, it with their whole hearts it's like how do you dare to try to make, me laugh when I'm stressed like, don't do that ever in.
Your Life and. Now we're testing also this in formal language we're testing what if we actually add lol's, and little faces and like sup, and instead, of how are you doing is like what's up man type of thing again. Very devices younger people find. It kind of cute and interesting old. People, older, people oh my god there was one person work, got very upset with you lol like. How is this boss, but like, saying. Lol at me I'm stressed. It stupid but you, know so these little things we're investigating, we linguistical and tried to design an interesting, platform that combines the, HCI, design, with NLP problems, and try to deliver this microbots. We're, also investigating, the voice part, but in the voice part. The. Voice version, of these BOTS we found a very interesting problem, like and, maybe there's a solution like I haven't investigated deeper, enough so. We did it actually in a Whizzer a bust format, in the car. These. Pauses that people make when they are closed so how are you feeling and like well um, you. Know it's been hard you, know I have all this from all that like you know searching. For the reasons, on how to spread your emotions, that. Are very natural, for humans to make pulses you. Break, this voice interfaces. Immediately, like. Try to express, your feelings to Alexa, or to, any other things, you know and tell, me your theory or anybody it's tell me if they can actually understand that you are not done talking. Very. Complicated, is and we try to do phonetics, and proxemics, and other things English. Is a very complicated one seems like Japanese is more doable Japanese, seems to have like some process. Of the inflection, that will tell you that you're not done or done but English is all over the place so where that's, that's one of the things that we will be working in the future can. We actually also find a solution well with a lot of data we might be able to train our deep neural networks on but we don't have a lot of data for stress, management in the way. And. The last thing that we're doing right now just says that mostly as you know the. Future of what we're doing is we're, actually combining. Stress. And productivity as well so one of the challenges with healthy people as I said like stress is not enough work watched why do I have to manage my stress for what and. One of the key challenges is productivity, so we're really bringing together some. Of the knowledge that we have created with sensors, and some, of the knowledge that. Michael. Bernstein and his team has created with habit lab this is a new tool that they created to do procrastination, management, only exclusively, procrastination. But, procrastination.
Is Very emotional problem. They haven't touched the, affected component, of procrastination, they only basically give you tools that basically, block Facebook, or like. Delay Facebook, and so, some, of these times actually. These behavioral interventions, work but, the thing is that procrastination, is emotional, I mean if you think about a procrastination. Is really almost, like none. Logical, like how why, will you procrastinate, if it's going to hurt you, hurt. Yourself, so. It's very emotional, that the psychology, behind is coming so we're combining the habit lab learning. With the pub therapy, stress management and papads in trying to create like, an interesting like you know probe, of a stress and productivity habit, luck has already like 12,000, users and we're like working with these people and they have, like 50 new users, every day being. Recorded for, this plug-in that they created. Alright, so that was the office let, me tell you a little bit about the car so. The. Car why the car though like so actually. It's not about the car. I'm. Gonna skip this one we. Already went this it's. Not the cards that commute. The. Commutes are very interesting problems so remember, there's no time to. Deal with stress well. That's because you're running and working that. Commutes one of those times when people actually don't really have a very strong opinion of like their, time years it's like will. You do a breathing, exercise in, the commune. Yeah. Probably yeah. Would you do something yeah I'm. Not working I'm not either at, home I'm transitioning. Well. The commutes are very interesting thing because in the US people. Are spending cars, 60, minutes a day every, single day on average. This. Is a hundred and twenty eight million people that jump in a car every single. Day 87. Percent of the labor force, commute. By car in this country, this. Is 128. Million hours, no. X 2 hours. Per day so. This is like a lot. Of opportunity. 114. Million of those commute alone. By. Themselves, in a car. Can. We do something to confer to transform, the commute which is for many a bug, in life, into. A feature again, you know I'm like taking, the bull by the horns on like a commute. Yeah sucks well let's. Do something about it right. How. Can we use a commute, to break the exacerbation. Chain between office, and home you know to do work-life balance and, basically. Create the opportunities. That we want to do that so again, doing the centrally sensing and subtle interventions, we, actually. Did. The same studies. That we did for the mouse were in the car and also, we've been dealing with like breathing, regulation, in the car so the, sensor, the sensor part is essentially. An extension of the what we said before like you know. We. Can use these. Beautiful sensors, for movement in the car there's this movement for, the upper limbs and for the lower limbs and they are already being encoded pretty much the drive wire by wire cards. Modern-day backward they all have encoders, already you, can actually go and buy an obd, and onboard. Diagnostics, USB, from like online, and get, like data out of your car in no time so. How do we repurpose, this car and and the, reason is that we're also doing the same type of approach that we had before in the in the car instead of looking at the autonomic nervous system analysis. We're, looking at a somatic nervous system analysis. By looking at how you move things, that's. How. We basically, approach, this problem. This. Is a traditional, psychophysiology. That. You can get in a car as well and this is what we think we can add value, like looking. At movement from your neck your shoulder, your legs your, arms. And. Converted, into an unobtrusive stress, sensor, so. What we did is we basically started with a simulator, study, because. You never know the effects of stress so, we got people driving, and we monitor essentially. Your turns right there, are anything that's positive, is a right turn anything that's negative is, a left, turn and we got them we, basically, wanted to do the mass spring damper model, again but this time with like angular, movement.
And We, put people in both, stress. And and, and and calm, conditions. The stress condition is a bit peculiar though because in the mouse you, can get like a hundred clicks, in like about a few, minutes you, can add a hundred you cannot get a hundred turns in a few minutes in a car so, you have to stress them up but then you have to strain the stress so what we did is we did applied math. Not. The same that we did in the other one and then we apply heavy-metal during. The drive and this is actually a former, yeah there is actually a very interesting journal, psychophysiology, paper, that shows that heavy. Metal is actually pretty stressful for ninety, percent of the population, and very, calming for another ten percent. All. Right so the experiment was like you know 25 people you. Do the baseline, with a video then you calm them down then. You do the driving thing and, you know counterbalance, everything and, as I said you know we have math but also heavy metals to keep the stress high while you're like driving, there, were a couple of like old, ladies, or. We. All skill them without having and they said is the worst experience of my life I have never had, something like that. Again. Damping free is the damper in frequency gonna change we. Use this advance. Simulator, for driving, that basically had high frequency, high sampling. Rate to. Try to get a more precise, reading. And we. On purpose, didn't put any type of traffic or any type of other stimulus, because otherwise you will be stimulating, other components, of the somatic, nervous system. So. Simply, a driving. Learning, curve and. What we did is we basically did, a simple analysis, we took the turns we got the absolute values then we found, the peaks and look at the monotonically, increasing, segments, of turn. And then, we basically picked, the, first part of each segment the. Was you, know in a turn because that's the first part when, the intention, occurs with your muscles, the first few. Degrees after, that there is other other, forces. That influence, the movement including, the the, the the movement of the wheel itself so we wanted to keep the very first part, of the of the movement which has a stronger, you, know a muscular, activation. And, then. We apply the same technique, for the order linear predictive, coding which, is a speech technique, speech. Processing, technique, to predict, the second-order, mass spring damper. We. Look at basically what's called the under damped poles which is the stable you know poles over the system, and we were able to predict, the, damping frequency, correctly. We. Validated, that stress, that people got very stressed, we measure also the, pleasure our elements, and as you can see people really hated, the heavy, metal like really, badly. And. Here this yeah we basically proved, that damping, frequency, also works and we published this paper in KY last year this is the fast, and furious paper. There's. A title of the paper, the. Beautiful, thing that we also found that we didn't need ten minutes of driving which was more or less what we did to get like a la with, about eight turns. We. Were able to also see already a signal with, the first eight turns so, that means that you know even when you're like just moving out of a parking lot and getting into a highway where it's more stable you might still get be able to detect something and, we also simulated, we decimated, the signal in order to get a lower sampling, frequency, because the OBD that you get the signal is that sample that a thousand, Hertz like the simulator, that we did is usually, about around twenty to fifty Hertz so, we wanted to see if it works with like decimation, and it worked well as well. We. Chose replicated. These in the wild and we are like should. Be published, in in would soon, and, you become we. Did it in the garage or. Under, the office where we work we work in the theranos, building with Elizabeth, Hong Jisoo. So we moved there we took over he had not to do with Tara knows it's just the left. Over. Of that building, so, we use this because since it was abandoned, world abandon its remodel, that this, was the first lab of the med school in that building and we're very proud to say where they we have the first lab which is a garage with, a car and, we proved that it actually works also very well in, the wild and we're like submitting, this my my my postdoc, did. It. We're. Gonna skip this one so that's essentially. With a sense of the sensing part now. On the subtle interventions, part. How much life. Okay. So in the cellar interventions, we started testing like will people actually do movement, in the car well as you can see I'm obsessed with movement, and we, saw. That you know can we guide them with like you know a chair that, has a bunch of like mothers to cogno, to, like do, things like guide you like it's, a swap swipe, up and down like to see if you can bend or do twists and things, by. Activating, these motors we wanted you to do like side stretches, or head turns or or or other, things and eventually the one that really worked was their breathing, they want they will do, guided.
Breathing With this type of chair like that the chair basically, tells you when to breathe in and when to breathe out and. They basically said that's what I want to do deep breathing and maybe other types of breathing like sigh breathing like tell. Me when to sigh and things like that so so. We went on that that was an exploratory, study and we basically did it breathing I was. Going to ask you to like take a 10 second deep breathing so you can feel the power of the breathing so, we can do like 10 second deep breathing and see like how do you feel, let's. Do it together so I can also take apart. So, see deep breathing is pretty soothing right even ten seconds, the. Problem, is that it's really, soothing to the point I can't get you to sleep so, you're driving though, we. Don't want that so we actually wanted to test it can be done while driving so we basically got people driving, and breathing. Slowly and we saw that you know we can go from nineteen to, eleven breaths per se. Per minute in about forty seconds with the guidance which, is not bad you know it helps you like reduce, your breathing rate and, chest. With our simple guidance on your back I. Was, going to explain you know the reason of like deep breathing is because it modulates, basically, your. Oxygenation. And also your respiratory, sinus arrhythmia that's, how breathing. Regulates. Stress. And we, focused mostly on ongoing from, normal, to. About, thirty percent below, normal. Now. Really, deep, breathing is considered to be around six breaths per minute when you're really good and the benefits are fantastic but, you really have to be very good to do that thing you many. People actually like, cannot, do very, very deep breathing so we want we. Focus on only the early part which is the 30%, reduction and, we tested, several interactions. You. Know an accordion, like interaction. Or a counter, like interaction. Or spear, alone interaction. To see which one or just one, a counter, that basically tells you and eventually. Most, people prefer a very simple interest a swipe, that goes up and a swipe that goes down so. I'm like the other ones but this metaphor. Was the best one the one that we picked we, implemented, that and we basically got people like you know linked to your breathing rate and we, contrasted, it with a voice-activated.
Single. Like you know the car you're driving the car tells you breathe in breathe out breathe in breathe out talk to you we. Ran this study, in. The city and also in the highway, and we tested the haptic versus, it what is the reason we you do this and career research is because cities, and highways are very different. Places. To test any of these things and then, you have you know typical. Baseline, and then the different conditions and then. You basically cut. The counterbalance this thing I'm gonna skip. This design. Period, so this, is the the apparatus where the person was monitoring, the, person and this is the result so we had people, that reduced, their breathing rate while. They. Were exposed. To this the the, guidance but. Also after, they were left the guidance was turned off that was pretty cool so you turn off the guidance and they remain, breathing, slowly. Now. The voice was rather heavy also started a bit higher the baseline right but, it was a with. Better than ethic but, the sustained, effect of the haptic is a bit more like flat it was a bit more sustained, in the other one so both of them worked I wouldn't. Say that I have to beat the other one both of them work they, both work, and they both reduce also her, availability, or, increase. In this case right there are MSSD, I, think. The, other the other element, is also to test, well. Sure, your brilliance no but are you are you gonna get in that accident, so we measure, two. Element safety and performance I want there's a difference, between safety performance and car safety means that can you die performance. Like you suck at driving you. Can still suck a driving, and, drive, but. Not be, in a place of like so, we didn't find any safety issues and we found that people, perceive. The haptic to be less distracting, and we, probe that a little bit like you know what do you prefer they say I prefer happy this is their, preferences. Because. It's less cognitively, demanding means. That you know that I don't use my part for anything the, other one I have to be paying attention and, it's, also easy to reengage this was fascinating, they basically say like I'm driving and I don't want to bridge so I just pulled myself before, where I don't need to feel that thing do, mine my new Varuna and then I go back ok now I can keep continue, they. Prefered guidance, over. Over. Chess a prompt, a notch they, said no neither one a notch I want you to keep telling me how to breathe because otherwise I get confused and I keep driving and I don't do it and. They found better it many other like you know subjective, like easy to use efficacious, etc, and. Yeah. Some people say well I prefer that you put this in a highway not when I'm turning right or when, I'm like in a red light or something like that that would be good one. Side, effect of this breathing was we tested also autonomous, driving so we also had the autonomous driving and look at this guy 40, seconds, in and, he's, lippy, badly. Sleepy so if, you don't have level 5 automation, don't, put these things that I was telling the mercedes-benz people though we're gonna put it like do, you have level 5 out to mention in your cars no it's like they don't do it because you, might kill people with this thing you. Will put them to sleep very quickly. So. We actually also tested, well what happens if we breathe. Faster, than can. We wake people up and, we have a paper that actually shows that we can wake people up by getting you to breathe faster so we're working with a truck company now to, try, to like improve, performance. And, we. Tested it in the same garage this is a different, cohort, different. Even like circuit. And, we, tested it and we proved it that. It. Actually works. Yeah. I'm gonna skip, a little bit but basically we. Are tested, you know the control versus now we use a control, and the intervention, work really well and this is for, non conscious breathing right so its replication, of the other one but we also did it, stress, intervention. So we stress them up and then we verify, if we can reduce you're after a stressor, can we reduce your brain already it actually works even better when. When I stress will spread so I can bring your non conscious brain rate low I can really bring you down back, from a stressor in. The car so. Anyways this is the summary, of what you can do in a car they commutes a beautiful, thing so it's not really about the cars the commute is such a great, opportunity, to, do something in the mundane, world to, really like do not obtrusive sensing, and interventions.
You. Know I have a bunch of projects that you can see I'm not going to go through all of them but they are all in the realm of design science, and implementation, and machine learning we have, like six eight people working in my lab right now we're, very excited of like you know finding, other collaborations. With many of you it will be really nice and. You, know we want to keep repurposing. Things you know we did something happier. With. Mary as well can we actually add