Thank. You so much for having me I really appreciate this opportunity to, speak. To all of you my, name is Marsha Chan Koska and I'm, going to be talking to you about some. Of the work. That myself. And colleagues, have been doing on, applications. Of GPS, GIS and, other geospatial, technologies, to understand. Environment. As related to health, so. I am. A geographer by. Trade, and name and experience. And so. I'm going to start to talk a little bit just giving, you a broader overview of. Space-time. And why we even care about it when it comes to health, so. Basically. Everything, happens, in a safe and at a time and I think there's such ubiquitous, things. Around, us that we often. Do not take the time to actually think about what they mean, and, when they come together, with. Environmental. Features they. Work, together to create context. And context. At the very basic definition is, basically the circumstances. That form the setting for an event and. In terms of which it can be fully understood and assessed. So. Why do we care. Context. Is basically constantly. Influencing, health there's all kinds of different ways that contacts, context. Really matters for health and. Furthermore. The, reason I am very, interested in understanding context. And how we, can measure it how we define, it is because. Essentially. When we talk about intervening, and when we talk about trying, to improve, health we. Need to come up with a way to have. A controlled, change in context, and so, it's really important, to first better, understand, what that context, is and. The. Questions that I grapple, with as a health geographer, are. Listed. Here I'm not I'm not going to read them all out loud to you but they really all get at these different, concepts. Of how can we better. Measure context. And then how can we use context. To actually, intervene, and make. A change for health. So. An overview of prease overview, of what I'm going to be talking about today and. I'm going to give you some just. Super. Basic geography, terms, that. Are applicable to expose, ohm sensors, and exposure based research I'm going. To delve, into some detail about how we can combine. Sensor. Data environmental. Data and behavior data and then. What, what does it mean when we combine that data and what do we get out of it I'm. Going to go into some, details. About working with spatial technologies, so talking a little bit about error and, some of the ethical considerations when, we look with this type of data and then. I'm going to end the talk moving. Into a bit, of a new space for me but it's the area, that I've started researching in the past six months which is using machine learning to. Take, all this data and start. Trying to make actual predictions, about health, behaviors, and. That, to me is really exciting, because it's where I think this. Field is going to be going in the next couple years and really where we.
Can Start taking this data that we've been collecting for, so long and really making an impact. So. First was started, with some of the basics so, what, it's space and what is place, these. Are two terms that geographers. Really. Like to get into arguments about and, there's, no really reason, to argue about them because they're just two different things. And both of them can have really great and meaningful. Implications, for how we understand, health but. When we talk about space, what we're really talking about is the, physical, planes. Of where. Things are so, we have XY and Z and then we also have at the bottom their time, because, things will change over, time so we're really, talking about. Putting. Things into this plane. Of existence. So. When we talk about place, what. We're talking about it's actually bringing in. Something. That gives us give that place some kind of meaning, or that space meaning so, we can put things, into the space and then suddenly they become a home or a, tree, or a, school. And. So these, two terms will often use interchangeably actually. Have a lot of different, implications. For what we can do with health data so. Some examples, about space and what, we can do with space just purely talking, about the XY location or, the latitude and longitude of where someone or something is we. Can start looking at distance, we can look at spatial patterns we, can look into spatial autocorrelation. Looking. At, clustering. A different phenomenon, we can look at different spatial orders, scales. And resolutions, hot and cold spots so. These are all, pretty. Traditional and Geographic methodologies. To actually just look at the, relationship. Of a bunch of points and how they interact. With one another. So. Some of some further examples here, space-time, views or you, know a darling, of geographers. We really love looking at where, people are going through space and time and then, what they're doing during those different time periods, and so. We can. We. Can actually include behavioral, components, in the sensing as we look at as the one who's moving through space and time and. I'll get into that a little bit later but. Basically. The, point here is that we're, just using. Latitude/longitude. To try to identify patterns, or. Important. Emerging, properties about, what. Think what people are doing in certain places. So. Then on. The flip side what. What. Can we learn from plates and I think places really where a lot of us that deal with spatial technology. This. Is really our true interest, we want to know not. Just where, someone is because, just knowing where someone is doesn't really tell us much we. Want to know what's around, them we want to know where they are in, terms of, what they're seeing what they're smelling what. They're breathing, in we want, to know what. Kind of opportunities. They might have around them and so, this is really where place makes space. So. Much richer, and where, we can really start to understand, how health, is. Being impacted, by what, is around someone, so. This, is really looking, at environment, talking, about the built environment, physical, environment, air. Quality. Demographics. Socio. Cultural and. Economic. Context, and, the. People around that person, all, of these things would, be classified under place because they put, meaning into where. Someone is and. We, can the. Two kind of main, areas. That people really look at when we talk about place, is access, and, exposure. So. We can look. Where someone we can look at where someone's oftentimes. Or someone lives or maybe where they work and we, can defer. We can try to identify what. They have access, to either by the time, it takes to get there how, far away it is is it, financially, available to them is it, in their ethnic, and cultural preferences, there's.
A Bunch. Of other ways to measure access, and, on. The flip side we, can also look at exposure. And so. Access. An exposure again. These are terms that often get interchanged. But there are two really different things, access. Really talks about what, the one could get, to or, what problems would. Have. If they tried well. Exposure really, is much more I. Would say. There's. Not much you can do about what you're being exposed to unless. You're specifically, putting yourself in certain situations, so exposure. Really is more measuring, at any point in time just, what the route of somebody so the people what kind of opportunity, they have and then, of, course toxic biological substances. So. You. Know really clear-cut example is, air pollution, we can map air. Pollution from various. Sensors on the ground and in the air and, then, we can use that information to better understand. What. Kind of exposures, someone. Might have, based. On where they live or where they move around. So. Um, I'm. Going to talk now about what. Happens once we start combining these. Different concepts, together and, then, what, happens when we layer in behavior. Into, that so. This is a graphic. From. A paper, that I wrote with jackal, encourage Jasper, hit burn and. We. Wanted. To explore, just, a different way that people are collecting, information about. Where. They are the, environments that they're in and then what they're doing this, paper looks very specific, to physical, activities, but. You can really take this and extrapolate, it to really anything so, things like. You. Know asthma, attacks or. Cancer. Exposures, or any. Really any behavior. Or health, related. Concept. You can take you can extrapolate these, concepts, out but. In, this figure what we're trying to show is that. There's. This kind of continuum. Of data, that we're collecting from people so we can, ask people surveys. And. We ask them to spell support, what kind of environment, they're in and we ask them to give us their. Location either, their home address, or their work address or where they tend to go we, can collect, things like travel diaries, and. We. Can then, move into actually. Pulling in things from, graphic information systems. Or GIS so. We, actually, get environmental, data and, we can tie that back to different self reports, so we can ask. The person where they live and then we take GIS, data to actually look at their environment from a data perspective and, then, we can then further ask them okay well what do you do in this environment, what. Are the behaviors that you tend to do in this area, if. We take it a step further we, can actually start putting behavioral. Sensors on people in this case again this was the. Use of an accelerometer and so we were looking at the collectivities, and sedentary behavior, and so, we can combine. Their. Physical. Activity with, these, GIS, base layers and, really get start to understand. What. Kind. Of environment, the person is being, active, in, now. Here here. Is where a lot of a. Lot. Of research in my. Opinion, we start getting into this question of are we measuring access, or are we measuring exposure, and. I, would argue that is the, third, picture we're really talking, about access, because we're not looking. At. What. Someone is actually going, and doing we're, looking at the types of environments, that they.
Could Go and do things in if they so chose to do so. When we get to that last slide when, we talk about looking at the environment and in this case the environment is walkability, and. We're. Actually taking GPS, data so Global Positioning System putting, it on the person, adding. In the accelerometer. And so we can actually see, what. They're doing where. They're doing it and what, the environment is like in those locations and so. Here we're getting to on what, we would argue is the most sensitive and accurate measure, of behavior, in environment. And we're, actually also here, for measuring exposure. Because. We're, we're. Seeing throughout. Their day and throughout their movement what they're doing and what types of environments are just in. So. When. We take this type of data. We. Our. Research, team has worked on putting this into a, larger framework of how we can take, sensor. And. Spatial, data and, use. It to understand, health. And health behaviors. So. There's a couple different areas, that, we. Need to think about the first is data and analytics, considerations. And we. Need to make sure that, we, have valid, and reliable measures. Of behavior, and spatial, context and. Compliance. Of when, we ask participants to wear these technologies, how are we coming up with the proper ways of. Making. Sure that the data is being collected the, right way and then. A huge. Area that is often, very much overlooked. And it's not until someone has all, the devices it, has started, their study and suddenly they realized that the quantity, of data they're collecting is, massive. And and. They're left with kind of trying to figure out on the fly how, to process, all this data but, processing environment is really important here so knowing. That you have the right software and, the tools to actually take. GPS, data match. It back with GIS data and then, if you're going to further integrate, some kind of health or behavioral sensor. You, know how do you get that into the data so. All of those things are things I would argue you to think about before you, start on a study or story, start on his path, the. Next area is really, then how do we integrate these things and, again I'm this, little, paper at a framework developed for accelerometer. But I think you can insert, any, other, health, outcome, or health behavior, here so. What happens is that you get in the, center there what, we call dynamic exposure. So, you can take, GPS, GIS. And, then health. Or sensor data and really. Start to understand. Where. Someone. Is doing what. It. Is that they're doing and then, the environment that they're doing it in. So. Once we have that data. The. Third event area. Is really how do we analyze that data. And this, is where a lot of my work has been in really, trying to better, understand, and come up with methodologies for, taking, that data and then, analyzing, it in some kind of useful way I think, a lot of times and. People that I talk to they collect. This data and, then they don't know what to do with it because it's overwhelming it's often overwhelming quantities of data and there's, a lot of noise in this data so it's it's.
Important, To come up with analytical, strategies, that can, actually tell you something of interest and, so I've broken it down through looking at the. Data in three ways you can look at it through time through. Space or, through behavior, but by applying any of these lenses and really focusing, on okay. When are they doing activities, or what. Areas, are they doing activities in or, what behaviors, are we really interested, in, that. Really helps already filter, out the data quite a bit and so you're not working with these massive, volumes but you can actually start to fine-tune, your questions to. Be more specific so I'm. Going to give an example of a paper, that we recently, did that we wanted, to look at how insulin, resistance, in breast cancer survivors. Was. Related, to physical. Activity promoting. Environments. So, our behavior, or our health outcome, in this application. Was not physical, activity it was actually insulin, resistance, but, we wanted to know if, people who. Breast. Cancer survivors, who had more exposure to walkable, and recreation, environments, if. They, had better outcomes, and insulin resistance the, theory there being that if, they have more exposure than they might actually be engaging more in those environments and then that would help. With the influent resistance measures. And. So. This. Was a methodological. Paper, and so what we were trying to do is actually compare, whether, these. Measures. Of dynamic. Environmental. Exposure and. How. They how, they compare to a, static. Home, buffer, measure so. When I say dynamic, versus static static. Really, gets at looking. Just at a person's, home so if you look at that figure this kind of blue. Buffer. Buffering. Approach where we're just looking, at where someone is up one particular time or just throughout. Their day we. Just assign it one location and we look at the environment around that one particular, location and then, we wanted to compare that with this more dynamic, approach where we actually have their GPS track and we're looking at their exposure. Throughout, the day and the. Methodology, we chose. To apply to this is a, kernel. Buffer we, did that because, the kernel buffer takes, into account time. Waiting, and so places, where they spend more time will get waited higher versus.
Places That they spend less time will get waited much, less. And. So, our hypothesis, was that dynamic, GPS measures of exposure would be more significant, than home static, measures. Basically. Based. On the reasoning, that, it's. Going to be a more, specific. And accurate measure of where someone is and what their exposure to walkable, environments, might be. So. A, brief. Overview, of the results but if you look at the figure at the rate these. Three figures are, showing, these different. Ways of conceptualizing and, measuring, space. And. Exposure, and environment, so, the background layer is a, walkability, as. Though it's a measure of walkability which, was calculated with various GIS variables, and, in. The first figure we, look, at a static. Exposure so we're looking at their home this, is where the person was living and then, we buffer, it and then we're basically taking, walkability, for, that home buffer, so. In the next figure, in the middle one we're. Looking at total dynamic exposure. And what this means is we took all their GPS data and we. Apply, to kernel density. To. That data and then, we extracted. The walkability, based, on that kernel, and so that's weighted for where they spend the most amount of time so you can immediately see, here that there's a quite a bit difference between where, they're actually going versus. Their home and so while they do spend quite a lot of time in their home they're also spending time. In other areas, around that neighborhood. So. In the third figure what, we were interested in knowing is, well. Lettuce. By. Looking. At all their GPS points, we're really introducing, a lot of noise because. Maybe. We're really only interested in, where, they're actually being, physically active and, the types of environments, that they're actually being physically active in so. That third figure is, excluding. Any GPS points, where they didn't have, physical. High, physical activity, and so, we're only looking at areas, where they have high physical activity, and. Then looking at the environment around those, so. When you put these into a regression model. So, I mean the main takeaway, for us was that we. We. Saw, that the static. Exposure the just home environment. We found no association. Between insulin. Resistance, and. Walkability. But. When we put, it we would put in these dynamic, measures we, actually did see an association. And, it's. In the expected, direction so, more. Walkability, leads, to less. To. Lower, insulin, resistance. We. Didn't see too big of a difference between this dynamic, or the, activity. Versus, just total dynamic but. I think more than anything it helped us kind of think through. Again. What were, we interested in measuring were, we interested in measuring all. The possible, places the person could go to in all the places that they're exposed to are we, only interested in measuring where. They're, actually doing the behavior of interest, and I, will probably argue that the middle figure would, be the most pertinent. Because, we. Get into this question of well, it's, someone, is walking, somewhere in, order to be active, well then they're likely walking, to a place that promotes, physical, activity, but. If we actually look at their. Exposure, in all, throughout. For whole entire day we. Can maybe get a better understanding, of well, they, were in, an environment that because promote walking but they didn't walk or, vice versa and then what that means for, their health outcomes. So. I'm going to get. Into some of the finer. Details of. Using GPS, and spatial technologies. We'll. Get to, too crazy here but I wanted to give people a little bit more, information. In, case they were interested, in doing this type of research. So. I, kind. Of wanted to just give a brief, overview of, where. Spatial. Science is currently are and exposing research and, health sensing I would. Say it's the vast majority that is right now in measurement. Quantifying. The types of environments, that people are in. Assessing. How often they engage, in certain risky, areas. Looking. At life course exposure, assessments. Disease. Dispersion, and tracking, these. Are all areas that. Exposure. Science has, and health science has really taken, and, I think embraced the spatial Sciences. In. Terms of intervention, I think there's still quite, a lot of work and quite a lot of areas, that, could. Be. Expanding, they use the spatial data into. I've. Seen, a lot of papers that look at tracking, people and then if they leave certain safe zones or if they enter, certain risk zones and an intervention, is triggered. And. Then. A lot of things that recommend, nearby health related opportunities. I think. Also for, exposure. Science a big challenge here is you. Know what do we intervene on and so. We. Can intervene. On the environment. That. Is often, very challenging, and very costly. And very difficult. That. It is not to say that it is not incredibly.
Important. But. I do think we're moving, more, towards, these interventional. Based interventions. Where, it, may. Be, less. Costly and easier, to try to help. Someone navigate, the environment versus. Trying, to overhaul, the, entire environment, itself, and. That, has a lot of implications to in terms of how we deal with social justice issues and. All kinds of other stuff that's I'm not going to get into in this talk but I think, it's just it's issued where I see the science moving and whether, or not that's a good or bad thing is I think definitely up for debate. So. Um. Oops. Lost one there so. What. Are the technologies, that explosive, research can employ um. Right. Now in. Terms, of measuring space the, vast, majority, of studies that I've seen are utilizing, GPS data there. Are a number of, other spatial technologies, that can be used and that. Probably. Would, take some work to implement but. These. Are a lot of these are technologies that are coming out through. Like. The googles and the apples. And whatnot and a lot of them are starting to use much more specific, spatial. Data. In order to quantify and identify, where their, customers are and I, think it might be really, important, for us as researchers to, start considering, whether. We not we want, to get, into slightly, better. Spatial, accuracy of our technologies, so. Things, like Wi-Fi compass. Beacons, access, point networks, distributed. Antenna networks, these are all. These. Are all infrastructures. That are being developed by commercial, giants. But that there are things that we can piggyback off of a lot of especially. Google they, really do like, to open, up and let others, use some, of this infrastructure, that they develop. So. In terms of measuring place we, have GIS data and you. Know collecting, and cleaning, and getting GIS data as a whole can of worms, but. There's just, an incredible wealth, of data out there in terms of how we talk about or. How we measure environments. We. Can look at social media or network data and, I. Think, there's a lot, of space to. Develop. Better ontology, of, how we take, GIS, data and then develop. Keyword and, methodologies. To harvest the data that fits our needs, and. Then we have historic, data which can, also be very messy but very rich. And, then in terms of measuring behavior. And. Health obviously. We, can use sensors both, after them passes and I.
Mean. This is a really, huge emerging. Technology. And or, space right now for us to be exploring. Social. Media networks again I think that's an area that we could do, better at harvesting, some of the data that's coming out of that type of data. EMA. So, you're. You're just asking someone, as they go on their day you're asking them questions throughout their day through, either smartphone, or other technologies, and then of course we have traditional questionnaires, and, biomarkers. And you know biomarkers, really are kind. The gold standard, um. But. With, the just, with a vast amount of data that's coming out in other, formats. I think. There's a lot of space to try to figure out how we can. Take. Biomarkers. And, link. Them back to other types of data sources, to. Just get richer, information about, people's behaviors. So. A brief. Word, about GPS. This. Is this is more personal. Interest. Of mine but, there. Is no I so, I work with a lot of researchers, who, utilize. GPS, then, they look at the GPS data and then, they get annoyed. Or frustrated, that it's not as accurate or as clean as they wanted it to be and, you. Know the original purpose of GPS was, military. And outdoor based applications, and, a. Lot. Of people always ask me well is it reliable and. And. I always say well if it's good enough for the military it's probably good enough for you. But. That being said you, have, to test GPS systems, and various environments and they're. Meant for outdoor applications often. Open. Outdoor applications. Using. GPS and built urban environments, can get Messi's it can get the signal can get lost and so. It's important to really test out how good your GPS, might, be, in the environment that you're looking, at using it in. Then the other, question I get is the act is it accurate, and I. Would, say that in ideal, conditions, yes it is accurate but, again, you need to test out your conditions, and. I. Just I see a lot of people implementing. The use of GPS devices in areas, before, even testing, you know how good of data they might be getting, and. Then once it does come in it's quite a bit messier than they were expecting. So. If. GPS. Is not good enough for your application, like I was saying earlier. If. You, look into computer, science and you look into. Smartphone. Based applications. In the computer science literature there. Really is a whole wealth of other technologies. That are coming out that people, are utilizing and, that. There's considerable research, and applications. Being, considered. At this point it's. It's. Advanced, enough that I would be very comfortable using these these technologies, in a Health Study so, things like using Wi-Fi off of smart phones come. Post technology, beacons. Are, incredibly. Accurate. Access. Point networks all of these are areas, that you can look into if you do need something more accurate, than a GPS, that, being, said, I. Often. Work, with people who want, the absolute, Premo of accuracy, but. You know when you actually take a step back and say well, how much detail do you really, need do. You just need to know like generally. Where someone is beyond. The home or do you need to know that they're exactly, in this building, at you know this exact time and so, really taking a step back and understanding, your health question and then how.
Specific. You need to get it's, a good idea to start with. The other question, I get a lot is um how, long somebody needs to be monitored um and, I. Recommend. Looking at Stephen Matthews work he took quite a bit of research on this but, also this paper by Perkins, at all does, a really good job of looking at. Fine. Scale human movement, and how. Habitual. It tends to be and that, and that is important. To recognize, that people are very habitual. Beings and. So again was your application, how specific. Do you need to get and. The other question is of course the sustainability, of your study protocol. As we. Keep moving into smartphone, based. Applications. Of sensor, technologies, you. Know we're really going to be able to I, think we, already are and we will continue to be able to, monitor. People over years, which, is really exciting. But. If you're going to be putting you know full-on. Co2. Or you. Know and any. Kind of. Intense. Sensing. Device. Onto someone obviously you can't have them with that for a long long time so. Just. Keeping, that in mind of what your participants, can endure. The. Last kind of. Fine. Scaled thing I wanted to talk about was. Ethics. And. IRB. Issues this. Is another area I get, a lot asked, a lot about, GPS. Can. Be terrifying. In terms of how again. How accurate, and how, detailed. It, gets such as a person's, movement, patterns, when. We layer on to. That some, kind of sensor that looks at health and that's. Really powerful data, and so, it is really important, that whatever, protocol, you're designing is, secure, and has. Treats. The data with respect it deserves. So. I. Really. Recommend going, to core this. Is a UCSD. Project led by Camille. Nutburger and, if. You go to their website you just have to give them your email basically and, once, you do that you get access to this, amazing resource, library, and. A forum, and the, research library actually has, other. Researchers. Submit their IRB, approved research protocols, using. Sensor, and, mobile, and imaging, social. Media and location tracking technologies, and so you can actually look at other how other people have addressed IRB issues how other people have proposed, to house. Their data securely. Other. People need to you know plan out their research protocols, this, is I mean I can't, stress enough how amazing. This resource is so I really do recommend you go and, check. That out and sign up. Okay. So in the last, 20. Minutes here I wanted to turn focus, on a. New, area that I've. Been delving. Into and it's. Really, taking. All that I've been talking about over. The past, half-hour. And applying. It into how. We can actually predict. What, somebody's going to be doing based. On their. Environmental, context. And. My. Interest, is very much in health behaviors, but, again I I really, feel like this. Applies. To health. Outcomes or. Health. Events. So, predicting. You, know when. Someone. Is going to develop a certain type of cancer or predicting. Where. The, next asthma attack is going to happen I think I think, using machine learning with, the wealth of data that we have, can.
Really Start moving us into this exciting, direction, so. This all really builds from the, idea of context, and. Again I'm giving this definition again the circumstances, that form the setting for an event and the. Reason I'm so obsessed with this definition is because, until. We understand. The circumstances. That forms, a setting for specific, health event we. Can't intervene. On it and we cannot predict, it and so. I'm. Really interested in understanding what. Data do we need in order to really get. The most accurate measurement, measure, of context. And then. To be able to act. Upon that so. What forms, of context, and. There's. I mean there's so. Many things that can layer answer this question but. These are the ones that I've, been, really interested in so. Face, where. Have they been previously, where are they now how. Often are they in this place where, are they likely to go next and you'll, quickly see that as, we start developing. This idea, of context, this, is not in any way shape or form only looking at the specific. Time that, we. Have a data point or we, really want to know, what happened, leading up to that data point what's. Going to happen after that data point and then. You. Know what are the patterns that shape. The. Occurrences, that lead to that data point. So. Place. What. Type of place are they in so looking at all the different environment, meant, variables, that we can put into that information, the. Time what, time of day days week we do a month month here, how, long are they spending there how long do they have to spend there so trying. To understand our time constraints. Really. Getting into understanding. Who the person is and this, is really key I think to a different, really. Understanding, context. We. Need to better, understand, individual. People. And individual choices which. Is entirely possible with, the type of data that we can collect from. Smartphone. Or other fate, sensors. And then. Also the, the social context so who is around this person. Are. These usual people that they know etc so. These are just examples of how, you can start building up this, idea, of context, and I guess what I really just want to emphasize is, it so much more than just where. Are they what's. The environment like, there's. So many other things we can be asking, about specific. Context, to. Better understand, what, their health behavior, would be, so. In terms of moving from context, to prediction, and, the. Idea here is that if we can accurately identify. Consistent. Context, and that's really the key there is something that has some consistency. For. Specific, behaviors, or adverse health events, then. We, can actually maybe predict, those events before they happen. And, so this really gets at. Working. With n of one type data so we're just collecting and saying about some information about one person, but, then also combining, that with multiple. People's data and then. Finally, I'm bringing, in machine learning and deep, neural network learning to. Basically, be able to take all this data and try. To make, a prediction. So, one, example that I've, been working on and using this type of data and making these type of predictions, is I. Want. To be able to intervene. On someone, making adverse, eating, choices. Specifically. If, they already have underlying, or. Health, symptoms. We. Really want to prevent them from further. Is a spring their health condition so. We've. Developed this prediction model, based, on a, person, a. Person's. Data and this is taking data from. GPS. Accelerometer. And then. Photographic. Data so we take photos of what the person is looking at every eight twenty, to thirty seconds, and. We. Can get all kinds of information. From, that and then as well, as layering, GIS, data about the built food environment, and basically. As we develop. This prediction model we layer in the information, what, we can get out of the machine learning algorithm is. The. Low risk of an adverse eating, event versus, this high risk and, once we get into high risk we can look at. Uncertainty. And thresholds, and just say all right well. This, looks like a high risk but I'm not that sure so I might need a little more data or this.
Is High risk enough and I'm pretty sure about it and so I'm going to trigger an intervention, message and in, this particular application the, the concept is trying to prevent people from eating, and fast-food restaurants, so, when, I become relatively, sure through the or when the machine learn model becomes pretty sure that they are about. To enter a fast-food, restaurant. They. Would get an intervention, message, to. Either. That. Woody to provide the healthier choices in that restaurant or maybe provide restaurant. Options in the area that would be a better choice. So. Um, this. Is the analysis, framework that we developed for, performing. This of, this, model and basically, we're taking data sources from the. Body. Wore incense cam so person, worn camera GIS. Layers the GPS data at the accelerometer, we, process those data's in. This kind of crisis processing task and then, we extract a bunch of features out, of that data and then, we train a model, a machine learning model and in this particular example we train two of them to, compare them which retained a logistic regression model and a random forest model and, then. We, evaluated. That model and. We. Actually got pretty, decent. Prediction. On when. People, were about to go into fast-food restaurants, we were at. About the 75%, level. Which. We thought was pretty decent. And, if you look at the importance, of different features you can see that obviously the, Sun scans data so, looking at the photographic, evidence of what they're about to do is really really important. But. Then just, their location, and then the different types of GIS, environment, that they're in also. Are pretty important, so. Whether they are, close. To a fitness center whether they're close to, certain. Types of food environments certain types of parks all of, those things. Were. Taken, into the model and actually had some importance, on predicting, if, they were about to eat in a fast-food restaurant I bring, this up because we, took. An approach that was kind of like let's just, put every, there when we possibly, can and, let, them let what's a model decide what, is important, and I, think that obviously. Sometimes. Does not have value in this, situation, that really does have value because the. Point I'm trying to make is we don't really understand. Those researchers yet what. Context, matter, the most for certain health behaviors, and so that's, I think really where the power of this kind of machine learned. Methodology. Is is we can use data to, try, to help us uncover, certain. Things that really maybe, we didn't think mattered but. May, actually really, matter for certain health behavior, or a health. Outcome. So. Um, some. Final thought. That. I wanted to share um I. Would. Say this field is um incredibly. Optimistic it, has a lot of potential, and there, is some areas that we need to exercise some caution in um, so. Hopefully. Through the course of the talk you would agree with me that we can do some really cool stuff with, um place, place time and behavior data there's. There's, so, many applications out there and, I'm. Excited to try them all. Anyone. Who knows me knows that I really like to experiment. With anything, that comes across my desk, that's. Being said it. Is really important to remember that there's a lot of error and there's just a lot of noise in the type of data and so, coming at it with a very clear research question and a clear. Way of how you want to use the data I think is really important. And. That's. Again I think where maybe.
You Turning, towards machine learning might be a really exciting Avenue. Because. Then we can take that noise and, really start filtering. Out what, is important, and what isn't and, I. Think. In general we need to be mindful of what we are calling context. Exposure, and access and. There's. A lot of research out there that, equates. Measures. Of access with measures of exposure and, they really aren't the same thing and. The. Reason I think it's important to differentiate. Them is just because they have different health implications because. You can't, you, can't say, anything about changing. Some of health behavior if you're not measuring. The right thing um, and. So finally I guess the my last main. Point, is really as. What. Is the question that you're trying to ask, with, spatial, data and then, why, do you need specific, types, of data and. So really honing. In on what. You need and you. Know do you need to expend, the resources on. Better. Location. Data or, do you need to expend the resources on, better questionnaires. Or, better sensors. So. Yeah, just really making, sure that the data. Collection is matching your questions, um. That's. About, it that I have for you all I, really. Love to collaborate with people I love. Looking. At all kinds of questions that deal with spatial, technology, and health so, I'm vixx's go to contact me if you have a project that you. Are interested in collaborating on, and. Yeah. I think I am ready for questions. Great. Thanks, Martha. Fascinating. Presentation.
2018-03-18