>> Mary Czerwinski: Okay. Welcome. It's my pleasure today to introduce Dr. David Woods from the Ohio University of Columbus, Ohio. Dave is a full professor there and a professor of cognitive and human factors in fact. Dave's skill, I've known Dave since like 1989 or '88, I think. We have shared pedigree because we're both perceptual cognitive psychologists an our advisors go way back. But Dave's expertise lies in the area of critical decision making and usually under uncertainty and pressure and in life threatening situations, although he's done work in all facets of cognitive engineering. I remember when I first met him at Johnson Space Center he was working very much in the nuclear control industry as well. So done lots of fascinating work with lots of different kinds of methodologies and now is going to talk to us a little bit about what he's learned about design, which will be interesting to hear. So welcome, David. >> David Woods: Thank you. So quick before I do the big zinger to bring in all the people sitting at their desk, we do lots of cool things. We're about a million, million and a half operation at Ohio State. And the talk is going to hit in indirect way several of the themes in the lab. How do we make systems more resilient, everything is brittle, right. To be more efficient, more lean, this is an overarching theme especially in the safety area, but all aspects of complex systems. What's rigor and information analysis synthesis? This is the data overload problem. How do we find relevant data. How do we not make mistakes which are often forms of premature narrowing or shallow or low rigor analysis. We're doing a lot of stuff with sensors and how do you integrate feeds from multiple sensors, be they as robot in a search and rescue kind of situation or surveillance networks and make sense of them, drawing on our perception and attention background. You'll hear a little bit about how we're trying to help synchronize activities in crisis management like situations. This is relevant to lots of areas. Healthcare as well. And in particular, we've like you, have been worried like a lot of these organizations dealing with new human interactive technologies, how do we innovate. The world is changing rapidly and we're constantly surprised by the what the actual effects of new technology are. So let's start the talk. Users are in revolt against their systems. All right. John Graham, colonel doctor has been travelling back and forth in the Middle East and this was his summary line of what's going on as we're deploying lots and lots of new technologies into soldier's hands, users are in revolt against their system. Well, that's old news to us, right. We know, right, users protest, the people introduce clumsy technology, it thwarts the purposes of people, it creates workload bottlenecks, knowledge bottlenecks, intentional bottlenecks so people work around them. Right. How long have we noticed problems like that. How many generations of theses have we go through to show our chops that we can go out and find these bottlenecks and fix them? But that's not why John said this. When he said users are in revolt, he meant a deeper thing, that they're in revolt against the design process. They're in revolt against outsiders thinking that they can anticipate and outdesign the difficulties and surprises and pressures of their world. They're in revolt because they feel with modern technology they actually have some design power. And they are going to go and create systems or modify systems to make them work for them in their critical situations. And that's really the themes I want to talk about. John's, another way we've summarized John's point is called we call the law of stretch systems. And the way I like to present this is from a previous -- this result goes way back. You could say Winegrad (phonetic) said it in '86, you could say Jack Carroll said it in '88, (inaudible) cycle, things like that. What are we talking about here? Well, we'll summarize really nicely in the late '90s on a major deployment of new technologies into a complex world. And the summary, the afteraction report set all this equipment deployed was supposedly designed to ease the burden on the operator simplify tasks and reduce fatigue. Instead these advances were used to demand more. Almost without exception, right, operators were required -- operation of the system required almost exceptional human expertise commitment and endurance. There's a natural synergy, right, between human factors, technology and tactics so that effective leaders will exploit advances to the limit asking people to do more and to it more complexly. The law of stretch systems, right. The systems under pressure, leaders under pressure will exploit improvements, new capabilities. The end result will be operational roles under greater pressure. So where are we? We're really pointing out that people are adaptive, goal seeking, meaning seeking, explanation building, tension focusing, learning agents. And this is really a contrast to most framing in HCI. People are limited. Oh, those poor, poor people, they're so biased. They will fall apart under uncertainty. But we will come and save them. Somehow it's only those other people who are limited, it's not us who hang out in research centers and innovation labs and know how to do all these fancy digital things, right, or user testing or whatever, great algorithm people or we're the latest sensor people, whatever the latest technology is, it's those other people, right, who are the problem. So we design artifacts as resources, and we want to -- the paradigm the stance we want to push is that our designs, our stimulants are triggers on adaptive cycles. And two general kinds of adaptive cycles happen, all right. In the law of stretch systems reminds us that we usual have expansive adaptations. If there's a capability that's meaningful, people will find it. Like they're exploring a niche in an ecology. We may see the early adopters, what are they, they're active agents exploring and discovery new niches in their information. Ecology, right? We've heard these analogies before from other leaders in the field, right? And they will take advantage of that capability and in that process they transform the nature of activity, the goals they're seeking, right, what's exceptional, what's typical, what are standard roles. Transforming activity, coupling, et cetera. Now, what we usually see is that most systems introduce bottlenecks, intentional workload, knowledge, bottlenecks, right. And so what do we see users do, various kinds of gap filling adaptations, work-arounds and things in order to get, be responsible agents to work around these new complexities, to achieve their goals. All right. So our theme is and what we really want to push in our organization and other related organizations is that we need to really move beyond that we're thinking about devices and objects and features, right. It's not enough to think about we're going to support engagement and experience and activities in the world, right, but instead we have to recognize and develop the foundations and the techniques and the concepts and the designs that look at our innovations, our releasing cycles of adaptive behavior and how do we model those, how do we trigger those, how do we understand those processes? We're trying to amplify the adaptiveness and resilience of human systems and we're not playing these games about overcoming human limits. So the question is what are the concepts, measures, techniques that we can do, use and develop, what do we need to do? So here's the quick summary. Design triggers adaptive cycles, there's two kinds of adaptive cycles. That means we have design activity -- opportunities at three levels, right. So people could do expansive adaptations by taking advantage of a capability to grow expertise in a role. I'll give you a couple quick examples. To better synchronize activities over wider scope and ranges, right, hyper connectivity and, three, expanding systems potential for future adaptive action, resilience, how do we make systems better able to, right, deal with prize and change in the world and still accomplish goals or even redefine what are the goals that are meaningful to accomplish. Now, I'll illustrate that with a study of crisis management that we've done recently. I'll give a quick introduction to the new work that's emerging on modelling resilience of systems. I will try to spend more of the time focusing on if you take this adaptive stance what does it tell us about how to model technology change and talk a little bit about our linked expansion constriction model and we'll demonstrate that with the new study we just finished on the electronic intensive care unit and how that's been introduced to supplement actually care units. So that's the plan. So what we're saying is if we take an adaptive stance it means technology is going to trigger that to cycle. You can think of it as three cross link levels. Growing expertise in a role which partly is related to how a role participates in joint activities with other groups, all right, and so we have our distributed work perspective synchronizing activities, but synchronizing activities is related to the ability to expand adaptive capacity, right, as an emerging property resilience of a system. What makes for resilience of a system, well better synchronization; for example, cross-checks, we've done a lot of studies in how to make cross-checks work better in health care. If you don't have good cross-checks what do you have? You have coordination surprises, you have a system that is much more brittle and likewise with expertise in a role. Now, the backdrop for this is a very old concept in human systems. And I think going back and looking through old, old writings, you know, things even older than I am, right, 40 years ago instead of 30 years ago, we have not taken seriously a notion has been around since the origin of human systems, research and human factors which is the concept of fitness. Now, if you take an ecological perspective, you have to wrestle with what's fitness. You can't -- if you take an adaptive or co-adaptive system stance you have to define, not to define fitness, you have to have an operational measure or approach to fitness. In fact people can define an adaptive system of something that's struggling to maximize fitness. It never achieves it in a long term sense because in achieving a certain level of fitness, what happens? The world adapts. In human systems effective leaders will take advantage of your success, so in health care safety issues, patient safety discussions we don't go back and say oh, you're terrible in health care if you just copied aviation you'd be much safer, we come back and say no, because you're successful, be right, you actually create new problems. So because you're so successful in making certain healthcare procedures work better than they ever have, what have you done? You've created a system where if you have a certain disease that fits one of their silos, you get the best care ever. But if you don't fit the silo or you have a chronic condition that crosses silos, this health care system is highly fragmented. So that second level becomes the critical one. We're really expert in a role, we're really poor at synchronizing or maintaining continuity of cross roles in health care. So fitness, so I just want to run through really quickly because it ends up being a foundational idea that we have to take -- we all do it, we all talk about it indirectly, we got to take it seriously, we got to better model it, better understand how to measure it. What's fitness? The search and struggle for fitness. It's another way to think about design, shifting away from usability testing and saying what I'm doing is a kind of fitness management. How do I explore, how do I change definitions of what's fit? So what's fitness? The simplest notion if you go back of fitness is that there are various kinds of demanding events and changing situations in the world, all right. As those events occur, people develop behavioral responses, all right, to respond to these situations. Pick any field, right, we're learning agents as we confront various situations, we have experience, what do we do, we learn, we take responses, we get feed back, we learn. So we can in effect plot these sort of different proto-dimensions and then say oh, wait a minute, the match between those is a kind of definition of fit, how do those relate to each other? Notice fitness is a relational variable. As you learn more in a local way, what happens, right, you get the set of behavioral responses that become more fit relative to the experienced dimensions of the situations. As long as you're experiencing them, getting good feedback, right, you will learn, no matter what. You're a learning system, you will learn something. We can turn it off but only under very extreme conditions, right? We are explanation building learning systems. What happens though? What we learned about the world, given what we've experienced is challenged, right, the world's complicated extra things happen. Things go beyond the normal boundaries. Surprise occurs. Now, that's a fundamental factor, right? Lucy Sessman (phonetic) pointed out to us with respect to the limits of procedures, there are always going to be for practical and theoretical reasons situations that challenge the boundaries of procedure system, make the procedures more complicated, we create new situations in which following those procedures will break down because of the variability and potential for surprise in the real world. So the potential for surprise means any algorithmic system has limits. Actually we were warned about this by Norbert Wiener back in the late '50s, early '60s. He warned us in a variety of stories in the human use of human beings even though he's one of the fathers of modern computer systems about being aware of the dangers the literal minded algorithms, literal minded agents. Why? Because there will be surprises that occur, the agent will continue to act as if these are the classes of events that occur when it's really in a different situation. So as we confront these kinds of complicating factors, these mismatches and breakdowns occur, we start to develop coping strategies, we develop extra adaptations. In some sense this is already a definition of resilience. A resilient system is one that can respond effectively not in terms of building this base level of fitness, but rather respond when complicating factors or surprises challenge the normal ways they behave to handle the normal range in variations of situations they experience. How well do they invoke extra adaptations to take care of these surprises and complicating factors? Learning something new about the system. So what I end up with is playing with our Hal Escher (phonetic) image here is this mantra. In design we either hobble or support people's natural ability to grow expertise, synchronize activity or adapt resiliently. All right. We stimulate their adaptive capability or we undermine their adaptive capability. There is no neutral. There is no neutral. That's an odd thing. You're either getting in their way and they're going to work around you, or you are giving them resources that will participate that they can seize upon in them adapting relative to their goals and pressures and the systems they exist in. And that's a really different stance. So a couple quick examples so you know there's really specifics in our field about all this. Expertise in a role, well, we can run through things like Paul Feltovich in Spiro's cognitive flexibility theory how do we escape oversimplifications or reductive tendency in their work, how do we recognize boundaries, how do we avoid premature narrowing information analysis, how do we broaden search, how do we revise assessment, how do we reconceptualize or reframe if you look at sense making work from Karl Weick and Gary Klein. In all of these, all right, expertise in a role is more than just being expert at one piece of things, having autonomy, autonomous action on one skill, one piece of things, but connecting those together and being able to revise and switch. Synchronizing joint activities. A lot of interesting things going on. How do we integrate diverse perspectives? We talked about that quickly today and last night. Anticipated reciprocity of professor -- I don't know if you ever took a course from her, she was there when you were there, Elinor Ostrom at IU, all right, in psychology department, has some fabulous work on reciprocity as a model of trust and it's redundant to put anticipated in there, but I want to emphasize that, reciprocity is anticipated, right? I will do something in my role that may risk or consume resources, risk outcomes in my role, but I'll do it because it helps you in your role overcome difficulties or it takes these opportunities and together we will better achieve overarching goals in our system. And if you analyze things like tragedy of the commons you'll find that anticipated reciprocity is a fundamental requirement for effective joint activity where we don't fragment and spin off into you doing your role well, me doing my role well and the total system going to hell in a hand basket despite that. This turns out to be very important in accountability models in patient safety. For example, healthcare areas. Directing and redirecting attention, judging interruptability, a lot of these are classic phenomenon, you can put different labels on them in CSCW kinds of work, collaborative joint activity. And our favorite one is enhancing cross-check. So that's what we spend a lot of time trying to study and enhance how do you make more effective cross-checks to enhance these collaboration and synchronization. Healthcare crisis management and layered sensing are huge opportunities for us right now. So let's do a study. We've been using crisis management as a natural laboratory. It's our standard method. We go out and work with real people who have to do significant -- responsible for carrying out significant risky tasks. It's neat because people will spend money when bad, really bad things can happen. But we do it in part because it's a great laboratory for getting data because these people really have to work at what they do because they know bad things will happen maybe to themselves or to people close by if they don't make good decisions. One of the studies we just finished is a critical incident analysis of a major metropolitan fire department somewhere between Boston and Philadelphia on the East Coast. Wonder where that could be. And they made available some of their firefighter injury and death cases. When we break out some that just happened because this is just dangerous stuff and you start looking at the coordination surprises we start to see how there's a variety of things they do to prepare for these episodes that create cohesion but the inherent difficulties they tend to approach, remember fires are relatively spatially contained, at least initially, so they approach and initiate their activities effectively, but the inherent difficulties of communicating what we find is they start to work at cross-purposes. And working at cross-purposes is one of the classic signatures of a synchronization breakdown. So o posting lines, so one group is hose line is driving a fire towards the other group, venting can do it, so if you vent inappropriately, you can actually increase the intensity of fire towards -- do you remember the movie Back Draft, towards the other group. And in a huge percentage of the cases these kinds of working at cross purposes play out, someone gets injured, they reconfigure they have a crisis -- a new crisis within the crisis which is to deal with a rescue an injured responder, and eventually they have some kind of resolution to that situation. So some very interesting kinds of things you can look at about the synchronization. But I want to tell you about a different study. It also illustrates the kind of way we use natural labs. So in this case, it's the new capability is you can put sensors out in the world. And you can understand remotely before you couldn't -- if you were a commander, for example, in a crisis, disaster, chemical plant fire, say, you were remote, you weren't on the scene, there were people on the scene, you were trying to share information, you had different roles, well with new sensor technologies you can be on the scene even though you're in the command center. One of the ways is UAV flying around, you get the video feedback in the incident command post. People love it. And if you've been in any of these incident command posts or military command post with these UAVs everyone lost the UAV feed. So we have this new capability, people adapt to exploit it, in fact they are so captured with it they overutilize it, over-rely on it. And everybody who sees this goes, I wonder if this is creating a danger. It's changing the roles creating the success and creating new vulnerabilities, new forms of breakdown. So they said let's study it, how do we study it? Well, we have a bunch of reasons we would worry about over-relying on one data channel, right, premature closure, framing. You're hypothesis generation and is framed by the image you may be unable to revise effectively if the data comes from a different information channel. So what do we do? Well, we create a staged world, right. We want to have a realistic rendering, but we want to be able to repeat the problem and run it for a variety of real people. So we were able to track down eight different real incident commanders in the Midwest who do this with reasonable amount of experience. We set up a command post setting, that's pretty easy to do, because what are they, rooms with lots of paper and a couple computer screens. The only thing unrealistic is we didn't put as much noise in the background as would really be going on as people came in and out and talked in the background. What did we do? Well, we flew around, took pictures of a chemical area and we created digitally a fire and so we simulated a UAV feed. Not quite as realistic. So then you design that, how do we do that? Well, we take in this case we took a real accident and it happened a few years ago, in England and we modified it a little bit and we did a garden path problem. And so on a garden path problem the imaging channel suggests initial plausible diagnosis, later subsequent data comes in that's contrary to the current diagnosis and it comes in outside the imagery the UAV video feed, all right. Additional give them a second opportunity to understand what's going on. All right. So the question becomes do they revise? And so what do we end up doing is seven of the eight commanders went down the garden path, right, they were stuck in the initial plausible diagnosis, they over-relied on the video channel, and when you look at the details of what they did, again, only one avoided the trap, but the ones who were caught were using fewer data sources, had limited cross checking between the data sources, and had a variety of anomalies in the data that they did see relative to their hypothesis that they didn't recognize and follow up, that their hypothesis really didn't account for everything they had seen either. All right. So they were doing a poor job of understanding what was going on in a dynamic uncertain situation. Now, let's pause for a moment. In some ways you can say that's a classic situation we're in. New technology has an upside and a downside cognitively and collaboratively. Here we're identifying that upside and that downside. You don't want to throw out the baby with the bath water and say don't use this stuff. They will use it. It creates a new challenge for us. Let's design better visualizations, bad balance, help people balance diverse heterogenous information sources. Some will be more compelling under some circumstances or most of the time. All right. Some may be, you know, some abstract data sources may be less compelling. Sometimes they may tend to believe people on the ground, maybe they don't believe people on the ground, maybe they can't translate what they're getting on the ground because it's in a verbal form versus a sensor, a data sensor form that's reading certain readings about contaminants versus a concrete visual. Very diverse formats. How do we integrate and balance those? Great research questions for us, fabulous for the visualization community, very important, very relevant, we're being reactive to an adaptive cycle. Now, we're trying to influence the adaptive cycle. So in some sense I'm coming back and saying what do we have to do? We have to get ahead of these adaptive cycles. How do we look further ahead, how do we start to understand what's going to trigger what kind of adaptive cycle? What's going to dominate? Is the expansive adaptations going to dominate in the video feed example, the UAV example? Are the bottlenecks going to dominate? In this example we can predict pretty much what's going to happen, right, the new capabilities are highly attractive, the deficiencies only happen when you actually do this, right, and so you don't usually do this, so we don't notice you do it wrong. And when you do it wrong, there's always a simple excuse, right, it's human error. What do we need? More technology to overcome those erratic other people. Not as smart as we are. Instead of recognizing, right, that these are fundamental challenges for any human system, human system because the system however automated, however many fancy algorithms, however many virtual technologies in the end is about human purposes, right. We're trying to have a safe world, an effective world, an energy efficient world, access to equity and access to healthcare, et cetera. So what we're really at is that the time to take advantage of all the advances in ecological and adaptive system modelling and start to bring it in to understanding and designing human systems. That means we have to start thinking about that third level. How do we expand adaptive capacities? How are systems resilient and brittle? How do we see hidden dependencies in coupled systems fundamental and software reliability, software dependability issues. How do we track side effects in replanning? We adjust, we have all this great new information, we can see that situations are evolving rapidly, we can take advantage of that instantaneous information, let's take directions to take advantage of this or to cope with this new information. What happens, we miss side effects other effects that are associated with the changes in plan and direction and activities and so we end up with that as a typical kind of failure. Rigor and information analysis. We have access to so much data we think we have enough data to make a reasonable decision on for example a space launch. Is a phone impact a safety of flight risk or is it just a throughput risk that we have to handle this on the turn around for the next launch of the space shuttle? So we have low rigor shallow analyses just to find, right, it's only a productivity issue, why, because they're under high productivity pressure. And you have these rationalization discounting processes going on. No one sees that it's in fact a low rigor analysis. In fact if you looked at the real data, right, these are more energetic than they had any actually models to predict. They were striking parts of the vehicle that no one had the -- that weren't the usual parts that they were worried about structurally, so no one knew the structural limits of the leading edge device on the wing where it actually struck in Columbia, so you end up with a situation where there's these key warning signs, right, you have a shallow, actually zero rigor analysis. You move ahead and all of a sudden, boom, you're surprised by an unlivable vehicle. Complexity costs of creeping featurism are a great example. Some people might say they're a great Seattle example. You know, we have -- we put out systems and we keep adding things. How do we recognize when the complexity costs dominate the incremental fitness value, right, what do we need, a fitness space. Complexity costs are looking at an aggregate over the fitness, shifting our fitness definitions from very local, right, to more global things that require take a broader stance on what people are trying to accomplish in the world. And so we have to think about how systems are resilient and brittle and develop models like that. Now, adaptive capacity, all right, that's been our theme all along today is to say that we need as partners is different research organizations to take this adaptive stance seriously and develop these new tools. And one of the interesting things about adaptive capacity is that it's about the future. Measures of adaptive capacity are about your ability to respond to future surprises. But the only way we can assess that kind of adaptive capacity is to look at how the system responded to past opportunities or disrupting events. So we look at how you adapt in the past, right, in order to assess your potential for adapting meaningfully to future surprises, even though we don't know what the future surprise is going to look like. So adaptive capacity becomes a generic kind of thing, generic kind of system capability that we could measure based on what the system does now or in the recent past but it's important variable because it tells us how the system is likely to behave in the future even though we can't predict the exact disrupting events. In fact, that is the definition of surprise, right which is, right, that future events will challenge in smaller or larger ways current plan full activities, right, how likely are your -- if you just behave according to plan routine typical contingencies that you do -- that you've done in the past, how likely is that to work in the face of future events? So what's a great place where this plays out? Emergency medicine is a great place to think about this. And in emergency medicine, you can think about systems and you can say oh, emergency medicine is designed to handle varying loads. Well, one of the interesting things is there's a report from two years ago from the Institute of Medicine saying that the emergency department is the brittle point in the national healthcare system. Why? Well, ask the head of Emory, the Emory ED. The Emory ED handled the Atlanta Olympic bombings. So an Olympic year. What was the casualties there? Two dead, 12 wounded I believe something on that order? I haven't memorized the exact statistics. The Emory ED head is very proud of how they handled that mass casualty event, and he has stated publically that if the Olympic bombing happened today his ED could not handle it, much less anything larger. Could not handle it. Erosion of expertise, important rigid environment, inability to shift rules and authority, lines of authority. Depleted physical resources. Right. EDs are under-resourced relative to increasing demands, more cases, more diverse cases flowing through. A recent studied shoed that EDs start to sacrifice normal quality of care indicators almost as soon as patients show up at the door. So normally you think of an ED would start sacrificing things as surge levels got high, it turns out you see some things being dropped as soon as you get a couple difficult patients. It doesn't take much to challenge these things. So what do you see in an ED, you see a lot of adapting. As loads go up, you can learn a lot about, right, cognitive strategies and how they change, how they utilize physical resources in new ways, how they change patterns of teamwork, assignments, roles, communication strategy, all of these things adapting in order to keep the system intact. What's a failure? Well, we have interesting things going on. If you can't handle -- if you're in charge of the ED and you're starting to run out of capacity, what are you doing, you're anticipating that ability -- can I handle the patients, right? So logical thing from your point of view, say I'm getting too close to capacity, I might mishandle a patient. So we'll diverse patients to other EDs, other hospitals. Patient dies on the way to the other one or waiting at the other ED, what happens? Scandal in the paper, right? Who is to blame? This hospital. What does your hospital director do, gives you a new policy, you can't divert. So is diversion a failure or is diversion an adaptive strategy? Both. Right. Now you can't divert so how do you handle patients? Well, now it's your mistake, not the hospital's mistake if you can't handle all the patients you have to deal with. Surge capacity modelling EDs is one of the hot places where we're trying to understand, again, a place that needs this and can doubt as you innovate new technologies how do they support adaptive capacity, the ability to change and escalate these strategies as loads escalate to stay resilient. On the other hand, it's a perfect natural laboratory for us to model and understand these processes, all right, and how do we design technologies and designs to support it. So this is the big thing. How do you manage systems resilience? How do you assess brittleness? How do you do this kind of stuff? We can draw on ecological and some math modelling kinds of large-scale complex systems. There's a variety kinds of basic principles that we don't have time to go into today. How are we on time? >> Mary Czerwinski: About 15 minutes. >> David Woods: Okay. Let me skip that example. I already hit this. Let's skip up to the technology change issue. So if we're taking this adaptive stance, then what this tells us is we have to think more about models of technology change. So this is an old animation we did a long time ago, the black box of new technology hitting ongoing fields of practice, creating a variety of reverberations of change. All right. Ongoing field of practice defined by artifacts, various kinds of agents in collaboration given demanding strategies in the world, the black box hits it, the processes of transformation occur, new complexities, people adapt in various ways, failure breaks through, usually gets blamed on human error but in fact we see new capabilities, new forms of coupling when we take advantage of those new forms of tempo, new complexities for people to take. How do we understand, anticipate the side effects of these changes? Well, the classic model of technology change is technology adoption, early late adopters, the Roger's S curve. Now, there are some other models that are starting to move beyond this, but this is the classic one. Now, this is enormously weak model for understanding what's happening with computer based and digital technologies. Jonathan isn't here. This is the part where I would talk all about him. >> Mary Czerwinski: Say hi to him. He's on vacation. >> David Woods: Yes. In Ohio of all places. And so one of the great examples that you can't explain with this is systems that fail. And you can't explain systems that fail due to workload bottlenecks, right, so the classic phenomenon that we see over and over again where in your role -- in my role as an administrator I get all the benefits if you take on new workload associated with the introduction of this new technology. But you're all under the workload pressure. So you get no advantages but you're supposed to do these things so I in my role get advantages. And normally those things fail. And you can read this large in many of the medical information technology failures. So for three decades they've been trying to put in computerized medical records. What often happens is exactly Gruden's law (phonetic), right the people who pay the workload penalty don't get benefits for their role, they are under workload pressure or even workload saturated. So naturally these systems don't go very far. They don't explain adaptive expansions, right, which is successes don't take the form of what designers anticipated but users create new exploit the capabilities to do new things that no one predicted, law of stretch systems, co-adaptive processes. So what did we do? So we start let's take these adaptive things. Even in a simple descriptive way we've got to start moving forward, sort of taking advantage of these other areas that have developed, different ways to do the multi agent co-adaptive simulation and things. It's a similar way to think about a role in the world as a role is like a performance track and the performance track is defined, at least when a role is well practiced relatively stable environment, stable background system for funneling people into the role. Think of it as a nice geometrically regular track, performance track. The performance track is defined by facets over four dimensions. We've got our workload dimension, right, Gruden's law, we have to have a workload dimension. We have an expertise dimension, right, technology change can introduce new capabilities for us to exploit in terms of new forms of expertise. We have an economic boundary, right, efficiency, productivity, economic gain. We introduce these things people want to get, right, see productivity or economic gains from it. And then we have, we've been debating whether to call this a safety or a risk or an adaptability. Let's just say for right now think of this as a risk kind of intention and a variety of facets where we are anticipating that things can break down and go wrong badly, all right. And we do a variety of hedges and adaptations so that we don't have major failures in some sense. Okay? Four classic kinds of things, all right. This role as an individual could be a whole group. An ICU could be modelled as a role, all right. An ICU dock could be modelled as a role. An emergency room could be modelled as a role or an individual subteam within that could be seen as a role, the cardiac group, right. And we have these dimensions, right. So what happens when you throw new technology? Well, you can have constrictions or expansions. So we have a simple way. Our track gets, right, bottlenecks thrown in. You can't go on the straight path. You've got to go around. Or we could break this path and say you've got to fill the gap, there's a hole in the path, you've got to reconnect the path in order to continue to do your job. Simple visual metaphor for the work around, right, gap filling. But you have the possibility for expansions. So think of expansions like niches. Now, notice with an expansion it's different. With a constriction you run into something. If you just keep going, you run into something. So you've got to go around. You've got to deviate from the usual things to make it work. In an expansion, you can keep going in the old way, but a new niche has opened up. That's the way you would think of it in ecological systems terms. A new niche has opened up and people explore who are thrown off the routine track start to discover, wait a minute, here's something new I can take advantage of. What happens over time, all right, is, right, we renormalize because we experience these as regularities in our environment, we take advantage of this, that defines new routines, new practices, right, we have standard ways to work around, we learn these, we develop these, we transmit these to other people come in, boom, boom, boom, we're now into a regular performance shape again. A regular track again. And you see this very nicely as people develop new experimental procedures in healthcare, they start to move them out to other settings and so it becomes a routine accepted and paid for by insurance kind of measure. All right. So let's look at this happen and so the place we went to look is the electronic ICU. So what's electronic ICU? It's a remote facility meant to support the physical ICU. What's physical ICU? That's where the actual patients are. So electronic ICU has access to nurses and doctors in a remote facility who are looking in doing vital sign monitoring, other forms of looking and communication. For example, they can have video, remote cameras they can control to look at the patient, look at the equipment set up around the patient, maybe a telemetry on vital signs monitoring, they can help out the nurses. Why would you want to do this? Well, the ICUs are under a lot of pressure. And you can see that as economic pressure, you can see it as expertise pressure. All right. There's a shortage of experienced nurses. Qualified people. In rural environments there can be a shortage of specialized expertise for different conditions in the ICU. You can have a small hospital that has a general ICU but may not have enough people to cover all of the different specialties issues that may arise for a particular patient. So you can expand the access to specialized expertise, a lot of different things. So what did we do? Well, we did the classic kind of stuff. What is it you start with, you go out and look. So we got access to it actually ICU and we hung out with them, observe what they did. Based on that, we started to do a cognitive task analysis of what are the different ways the EICU is being designed and operated to support the physical ICU? Well, what was really critical was they were logging their interventions, in other words whether did the EICU intervene to help the physical ICU or maybe the physical ICU didn't think of it as much help, but there was some interaction and since we had these logs, we had the ability to take a longitudinal study. What kinds of interventions and how were they changing over time? Well, if we're going to look at adaptive systems, we've got to start adopting longitudinal methods, we can't just look once, we can't just say how do people react at one point in time in a learning curve, right, how are we going to start to recognize the value and adapt that into their situation? We started out with worried about, well, gee, these things are going to support anomaly recognition especially, so monitoring help. Nurses like to call it extra eyes on the patient. Sense making kinds of functions. For example we saw issues where the people in the EICU could step back and take a big picture approach, right, revise, wait a minute patient wet or dry. Are we late or early, are we overreacting to something, driving, overdriving them into the wrong state kind of thing? The people in the physical ICU, right, have a lot of physical tasks to carry out. They deal with the patient family, variety of attendings and residents coming in, a lot of interruptions and things, interruptions, right, so all of a sudden you can start saying hum, the EICU is valuable because they're a more stable environment relative to monitoring. A lot of potential pluses. Specialized expertise, that was a key one we thought would be valuable. What we didn't best practice reminders, maybe that's sort of a sense making kind of thing but basically giving feedback to the physical ICUs of various interventions they need to be doing. In some ways this is stepping back from the detailed flow and saying there's other tasks you haven't carried out yet that should be on your priority list. One we didn't realize would be there was mentored learning. The EICU nurses were tended to be fairly experienced nurses and in the kinds of ICUs that needed an EICU had tend to have more junior. They were licensed but they were more junior in experience, they were very junior in experience levels. Yeah? >>: What was the mechanism to communicate with the EICU to ICU? >> David Woods: They had television linkages. They had video -- well, there was two-way communication. They end up calling, all right. So they had dedicated lines to call and talk to them about what was going on. That's one of the predicted trajectories is the EICU can now become a source of interruption and inappropriate interruptions. If they don't have effective common ground, how do you get effective common ground, is the video interplay critical, right? Not just to see what's going on with the patient but also to see what's going on in the physical ICU and the nursing loads to know when to interrupt and when to interact appropriately given what is going on. So you can think of it as being an ICU under a variety of pressures, variety of constrictions, economic constrictions, expertise constrictions and workload constrictions. What do you see? You see people taking advantage of we can avoid risk, we can adapt in a variety of ways by taking advantage of these remote monitoring capabilities. You see a transition and now we've moved from the -we started this study two and a half years ago and the number of EICUs in the country have doubled or two and a half times from when we started the study. So we weren't at the very beginning, but we were early in the migration, and it's been taking off. So this is a success story. People are adapting to take advantage of this. Different ones are configured differently because of their particular geographic area concerns, and we're starting to see it normalize into a new pattern of activities. What are we interested in is predicting what's going to happen next. Well, we saw one of these things happen because it was a longitudinal observation. There was one thing on that list I didn't mention. Billing. Billing. So what happened? All right. Hospital administrations under various -- these are notional, don't take these as the literal point from the data, but notional, they're under various pressures that create constrictions and from the hospital's point of view finances, right, are a big one. What are they looking for in looking for ways to be healthier and here's this EICU which is in their environment is a resource that they can seize upon and use to adapt hospital procedures. So what do they do? They start introducing workload into the EICU saying EICU your job is to monitor the physical ICU to make sure they enter things in a way that maximizes our ability to bill, not improve patient care, no more -- not access specialize expertise. What do you see? You see this thing where a success can start evolving into a not quite as success or the grounds for an accident that we could then come in and investigate and go oh, look at how this was a system failure. Look at how the origins of this accident happened, begun years before in organizational decisions, how this is an organizational accident not human error. I don't want to be there. I've been there enough. I want to prevent those things from happening by being proactive. So we saw this trend happening. So part of our analysis is trying to project adaptive trajectories and how do we do that, well, without going into all the things, I just wanted to illustrate on these how we related a variety of sustainability conditions. So that's one of our new themes here is you have a new capability that's seen as a niche, be effective leaders start to recognize the niche and exploit the niche and expand the niche. As they do that, they have a variety of effects, it's a linked expansive restriction, expanding on one dimension and the facets can create restrictions on other facets, other dimensions for that or other roles. So the issue is let's identify sustainability conditions and say if those aren't met, what are we going to see? We're going to see this erosion, instead of benefits on the expertise and quality dimensions or risk dimensions we're going to start to see benefits in financial ones erode those expertise gains and we'll start to see the risk of failure go up. So we can start to see some of these common patterns that our study reveal. So the extra eyes, right, better monitoring is great but that assumes the EICU isn't task loaded with other additional administrative or other kinds of tasks. If you've got a bunch of people sitting around, not much is happening, how you going to -- you going to load them more. How are you going to load them more, well, if you're monitoring three physical ICUs from one EICU, why not do five? Wait a minute, we can get more billing. Let's do seven. Whoops, wait -- all of a sudden we see economic interest of the entity running the EICU increasing the workload, either direct monitoring or indirect tasks gets -- other additional tasks gets transferred to the EICU all of a sudden that monitoring expertise goes away. Best practice reminders. Those are interruptions from the physical EICU's point of view. If you don't have mechanisms for effective common ground, what are you going to do, you're going to have bad collaboration. We've studied these things over and over again and in healthcare in particular interruptions are a very simple corollate of mistakes. High interruption, more mistakes in these routine activities. Administrative billing task complexity. The sense making. Your ability to step back and take a big picture approach, that wasn't one of the high frequency gains but it was potentially a critical gain relative to outcome of your patient, if your relative or you were in the ICU, right, you care about that a lot. Again, as tasking goes up, how are they going to be able to step back and take a big picture approach, make these high level judgments that may be difficult for the nurse on the physical scene because they have so many activities and interruptions there that are going on. Specialist expertise. Well, one of the interesting things about the EICU that nobody noticed is part of the reason the EICU is working right now is because there's a pool of available expertise. So older experienced nurses who no longer want to deal with the physical rigors of being on shift in the ICU. So you've got experienced nurses, older nurses, they want to work in a different style, they can have a better lifestyle or they can work shorter shifts part-time in the EICU than they would work in the physical ICU. Now, when that experience base gets consumed will some of these benefits still hold? Mentored learning, you can't have mentored learning if the people in the EICU aren't much more experienced than the people in the physical ICU. You're not going to have the same access to specialized expertise, this isn't simply the physical expertise, it can be also to experience nursing expertise, so a sustainability -- the sustainability condition is that there's an experienced pool available to draw on. So what we're starting to do is starting to say, hey, wait a minute, as a field we know a lot about these adaptive trajectories. We use them all the time in the way we design studies and the way we make recommendations and critique interfaces and run usability studies and innovate new designs. But we don't organize any of our knowledge around adaptive trajectories, we don't organize our knowledge and techniques around assisting and predicting and steering these adaptive trajectories. But maybe we should. >>: So the idea of the experience pool and perhaps all of those adaptations that you talked about, (inaudible) thinking of them as sort of directional, that the experience pool only benefits the EICU people with more experience than the ICU people? The flip of that, this may be true for lots of these different sort of adaptations is maybe the people in the original system in the ICU in this case can now provide additional learning for someone that might be at this remote location. I want that person in my Internet new by in my ICU but, yeah, sure they can over the shoulder in the EICU and be out of the way and get this sort of high level thing but still gets ->> David Woods: What did he do? What did he do? He just did an expansive adaptation. He said wait a minute if I want to train in nurses, all right, what do I do, I start bringing them through the EICU as a mechanism to get them up to speed before they have to deal with everything as a way to get concentrated and safe learning because in the medical world we don't want people learning on real patients. That's one of the trends out there, right, so we're doing more crisis simulation, other forms of medical simulation are starting to penetrate and spread. It's still relatively early even though they're available in most metropolitan areas at least one or two centers. But that becomes oh, wait a minute, another trend, another set of constrictions, here's a resource I can take advantage of, right, and so that's what we want to do is we're saying how do we start mapping these potential trajectories. Now, one of the other ideas I didn't put up on the slide is how do you monitor, we may not be able to absolutely predict ahead, but at least we can detect early emergence of these trends, right? And so your idea says I don't know for sure they're going to adapt in the training, and I'm not going to make a Las Vegas bet that their adaptation is going to be new training methods. On the other hand that's one of the potential trajectories. Notice we can come up with this by applying our general knowledge of human systems and new technology in HCI to these things. We don't have to know a lot about this to know, hey, that's one of the potential trajectories. Then the question is set up monitoring conditions. So now we say innovation monitoring conditions. How would we notice if that was turning out to be adaptively valuable to the user community? These human systems. Then all of a sudden you go wait a minute it turns out as training we can help you take advantage of that and to that in a good way, rather than in a poor way. Remember the cross benefits and weaknesses in our UAV for example for incident command, the technology capability created new vulnerabilities at the same time they provided new benefits for the incident commander. And so we don't want to get in these simple, you just have to take the bad with the good over-automation, human error, back and forth, games that don't get us anywhere in terms of overall system design, system cape ability. So we need to start being able to use our knowledge. It's already halfway there, at least, to be able to talk about emerging patterns across local adaptations. All right. That means we have to start integrating these multiple perspectives in different levels, right, so we had the hospital administration, ICU, physician, right, nurse, nurse training, all these different perspectives interacting. How do we anticipate adaptive traps when they're going to get stuck? Classic example we're all dealing with right now is corn ethanol. Great example of adaptive track. And it's also an example of an adaptive fluorescence. Can we recognize link sets of adaptive expansions where one adaptation -- there's an expanded niche, somebody adapts to take advantage of it, that creates another expansion which somebody takes advantage of which creates another expansion, that's an adaptive fluorescence. The emergence of corn ethanol was an adaptive fluorescence. It turned out to be an adaptive trap when you took a larger perspective that we're stuck in right now. That was the news driving over was cellulose-based ethanol, that will get us out of the adaptive trap. Notice the key theorem behind an adaptive stance. An adaptive cycles, yesterday's solution produce today's surprises that become tomorrow's challenges. It's not about right or wrong, good or bad, it changes the way we do the metrics problem. What we're trying to do is anticipate how people will adapt. When they adapt they will be maladaptive and in the sense of it will create risks from a certain perspective. And there will be benefits. That's one of the fundamental things about adaptive behavior. Locally adaptive behavior does something for those who adapted the behavior. From a larger perspective we can later come back and go boy was that strange behavior, right, and call it error, right, but we have to recognize and that's been 20 years of the new look at error which is error is fundamentally adaptive behavior, locally adaptive behavior that globally is maladaptive when you take a larger perspective and set of information. So the challenges I want to leave you with are can we project and anticipate the multiple unintended consequence of proposed changes? What methods do we have, are we using those methods, are we getting too local, are we being locally adaptive and globally maladaptive like in complexity creed? Can we forecast trajectories of adaptive responses? And can we discriminate what's promising so that we can actually participate in steering change in towards expansive adaptations? Back to Colonel Graham's comment, our users are in revolt. They are not waiting for us to figure it out. They are not waiting for us to create new capabilities, new visualizations, they are going to do it, they are going to adapt because they are under pressure to accomplish goals they appearance real risks of bad consequences. Whether we're talking about intensive care units, emergency departments, military systems, crisis management systems, people are adapting and in that adaptive process how do we find those open periods that we can help them, right, rather than being reactive or rather than just keep throwing new little innovations hoping one or another will penetrate the walls of adaptation that the adaptive traps they are stuck in. It's a whole new stance and paradigm. But the good news is we actually have most of the work already available, right. In the end, we already know a lot about how to release human adaptive power, it is through our design innovation process. But we have to harness and link that and model and do -- and change our techniques a little bit in order to connect it to our customers and the end users in the world a little differently. Thank you. (Applause) >> Mary Czerwinski: Any questions? >> David Woods: Comments? Arguments? Yes? >>: I've been using (inaudible) from crisis management trying to adapt it to long crisis management situations like life science, research which is kind of -- it's not actually -- it's a bit kind of crisis management situation but there are no human (inaudible) involved directly. However, it does involve high cost if you think about the (inaudible) so do you think (inaudible) situation is where there's no fire, nobody is dying but still it's very critical time-wise or financial? >> David Woods: This is an old, and old debate in some ways, right, which is oh, yeah, there's the sort of everyday experience, it's the business world, it's whatever versus there are these high risk critical specialized situations. And one of the changes that the technological world has produced for us has linked those two together, right? The stuff I used to do with mission control or with the nuclear control room are now happening in the business world. They are talking about resilience not because there's lives at risk but because the financial stability of the organization is at risk. If you hang around people who make lots of money, these high levels they don't act like it doesn't matter, right. When they lose out on their bonus, they get really mad, right. When the company goes under and they're out of work, they're not happy campers. So what's at -- there's always -- you know, when you say it's a human system there's always stuff at stake. And it -- the other way to look at it is be a parent because when you're a parent and you look at your early, the pre-teen for example or whatever and they are just obsessed with the crisis in junior high, in middle school and as an adult you're like oh, come on, this doesn't matter a bit about what clothes you wear or what style or which camp you're in, you know, or whatever, but to them everything in the world depends on that. So again, it's a classic example of balancing perspectives at multiple scales. So that's the paradigm and that's some of the other work we're doing is again combined knowledge from multiple disciplines, putting it together, we're trying to look at what's called poly centric control models and what are poly centric control models that actually comes from tragedy of the commons, it comes from social scientists like Elinor Ostrom at IU and what it is is a new way to think of supervisory control or managing adaptive systems and what you're doing is saying I've got partial autonomy at different levels so I've got multiple centers of control, each with partial autonomy and how do I balance that out? And so they have to be in sort of a creative tension across those levels. And that seems to be one of the key drivers to creating systems that are effectively resilient in the face of change and avoiding a kind of classic pattern of maladaptation which is called the tragedy of the commons. I don't know how many people know the tragedy of the commons. >>: (Inaudible). >> David Woods: Yes. Or you know, Enron or, you know another thing would be to look at Al Ross work. Another great example of this, he's the husband of one of the cognitive engineers, cognitive human factors people, always like to say it that way. You know, he's a famous microeconomist but he's the husband of a great human factors person. And he has a great paper on how facts unravel, sort of a synthesis on experimental microeconomics and again markets unravelling you start to see these multiple level operations going on. And so he's been doing this on transplant exchange programs, for example, or cross-bid systems like how do you place residents or medical students in residency programs. So cross-bid systems are a couple of the examples, transplant exchange systems, the kidney exchange system right now is working better, finding better matches because of the stuff he designed. So these are great examples that we can take these adaptive approaches and that we can do things that improve design. Now, one of the things is when I work with the designers at OSU is it's an interesting challenge because industrial designer, it's not the normal thing you design anymore. Right. But for HCI it's not the normal thing you design anymore. We're thinking of there's an object, there's a -- our system where it still has kind of an object quality to it, here's the visualization, right, you have a name for it. All right. And we can get that name in the system into the larger system because they make use of a fish eye lens or they make use of this other technique or visualization or gooey feature that we created. It's saying our design object is quite different when we take the adaptive system stance. Now, adaptive system stance has been around for a while. It's not new. What's new I think is that a whole bunch of people are saying it's matured, now is the time to move. All right. There's enough resources to draw on. And don't let the people who do adaptive system models do it by themselves because they'll screw it up when it comes to things like emergency rooms and intensive care units and genomics and all the other areas of hyperactivity. Everybody's experiencing brittleness in a highly hyper- interconnected, hyper-pressure world and they want to be able to respond more agilely, they want to have more foresight. These are the things that we're trying to do. Now, that's why I said they're cross link levels, right. Because if you're going to enhance adaptive power, what are you going to do, you're going to come up with better CSCW systems. What are you going to do, you're going to come up with better expertise systems like visualization aids and things that give you better feed back, better ability to sort through massive amounts of data, right? But the reason this matters is because how it helps synchronize and how synchronization matters is how it releases adaptive power, we've got to get those three levels interconnected. Yeah. >>: So sometimes it seems like you can be both (inaudible) and supportive. >> David Woods: Yes. >>: The EICU is a good example. I am no longer in the space, so I'm not hands on, you know, there's some things that I clearly can't get through a video camera, but at the same time you said there are benefits like taking the high level view. So has this just become a larger balancing problem now or perhaps ->> David Woods: Well, the multiple effects problem, that's why I use the crisis management example, because that's a really intense specific example and this was the first study that just gave empirical support to the observation, the anecdotal observation. And it doesn't have any design implications yet. We've got a better balance across the sorting the different information sources. But the mixed effects problem, right, that's what you're pointing out, technology change has mixed effects. And the adaptive stance says look for and anticipate how those mixed effects are going to play out and then if we're an effective field we ought to know a bunch about how those mixed effects are likely to play out. So when we come back and somebody says oh, this is going to be great, it's going to be a remote monitoring facility. We ought to be able to say, even though we've never walked into an ICU, oh, there's a variety of things that a remote monitoring facility might do that would be effective and there's a variety of traps that might happen in a remote -- because it's remote from the real world. So we would say what's the value of being on the scene? Is there something special about being on the scene? And so all of a sudden you start going what is it about being on the scene that's different than only looking at things remotely? Well, one of the things that comes out for example is this was shown pretty nicely in the movie Black Hawk Down or the novel actually we don't know if it's really true or not so I'll say the novel, it was really great which is the remote monitoring, right, the commander was monitoring the battle, all right, from feeds on the helicopters over it. And what is striking is he had no feel for how things were going downhill. All right. So in a -- so you remote monitoring. I know in advance that remote monitoring are more susceptible to late recognition and intervention when things are sliding downhill. They don't see the slide downhill early. We call it going sour. It's a going sour signature leading to an accident. So they're late at recognizing the going sour signature. Right. So there's something special about being on the scene. Now, it's not enough to just be on the scene because we've got lots of going sour cases where there was a resident on the scene who wasn't expert enough to exercise that things were going downhill until there was a collapse in the physiology. A good place to study this is intensive care units, operating rooms, anesthesiology kinds of places. But there's others. We've done it with aircraft automation, same thing happens with aircraft automation. So we know this is the case. We also know in highly automated worlds it's hard to recognize going sour signatures. Why, because the automation keeps responding to the early trouble hiding the trouble from the human supervisor until things having to hell and then you go, and then the automation goes here, dude, take care of it, and the guy goes, oh, no, what do I do now? And so there's a really cheezy Japanese reenactment of a real aviation accident where this happened, a near miss, where this happened, an airplane plunged 20,000 feet before the pilots regained control. So these things really do happen. So we know this pattern and so remote monitoring has this potential mixed effect. And notice the other reason I use the crisis management UAV example was notice when we pay the penalties for mixed effects can be different versus when we get the advantages. So the UAV feed gives us advantages right off the bat that everybody sees, they want to take advantage of that new opportunity. They don't want to give it up. The penalties are later, in cognitive and collaborative performance in an actual disaster where they get trapped in the wrong assessment. They have an incomplete analysis of all the data. So one of our advantages would be to help people, all right, see that mix. And then one of our things should be is how do we compensate for that mix. So we get the advantages that the opportunity represents because they're going to adapt to take advantage of the capabilities anyway if they're under pressure to do more. >> Mary Czerwinski: You have a question back here, Dave. >> David Woods: Yes? >>: I was just curious, is any work being done to investigate so the system that you described, EICU and ICU in the hospital I am in negotiating those together into a broader network of knowledge so to get that wholistic schedule especially in epidemiology, I (inaudible) so you can identify friends and ->> David Woods: That would be a great thing to get with a public health school who is connected with some public health departments in the area and work on. I mean, we've been trying to do that in Ohio and if we can get more funding or whatever, that's a great opportunity. What's the short answer? In healthcare, what do they think? They think that you can do this bottom up. Collect all the data, once you have the complete digital system for healthcare all the data somewhere in the digital system so then what do you do because there's too many data and you can't find anything meaningful then you must have data mining algorithms and the data mining algorithms will tell you, aha, here is the outbreak and da, da, you're done and it will tell you this and now you know that it was this pattern or trends and now you know what's going on. Will data mining do some things from massive healthcare databases? Sure. One argument is, a Google argument is that because we can have ultra massive databases on basic human activities all we need are bottom up data mining. We have massive amounts and we can find these little emerging trends that matter. I don't think -- yes, you will have some successes. Will it work overall, I don't think so. It's not really a wholistic approach. A wholistic approach would combine a bottom up approach or top down or middle out approach. >>: Okay so there's no real work being done ->> David Woods: I've been crying for 12 years now, 2008 we started patient safety movement in 1996. I've been trying out for 12 years saying guys, it's all about overcoming the fragmentation, you can't overcome the fragmentation if you try to save one big massive database will allow us to integrate everything. You have to have others kinds of mechanisms. Top down may be too hard to initiate given the current structure so the issue is how do you do middle on you, how do you do decentralized, how do you get emergent properties, what are the key levers that will give you emergent properties to create continuity? But that's the big, big target in healthcare. How do you get emergent continuity over virtual or distance interactions. And you're right, can we get combinations of perspectives? We've done it in the national aerospace, we've decentralized authority and we've created better coordination mechanism between airline dispatch and the strategic air traffic control so we know when to back off and how to back off when risk is introduced into the system and how to handle disrupting events in meaningful ways. We've done it on a fairly large scale system. But national air space system is tiny compared to the national healthcare system. >>: (Inaudible) applied to Southwest I believe it was for managing their main crews, luggage ramp, all that. >> David Woods: Well, we constantly run into people trying to transfer business world solutions which are related to efficiency and lean in hyper-controlled settings to really messy high potential for surprise settings. Healthcare is a high potential for surprise world, and things that depend on controlling the environment. So the classic -- applies to automation, right? If you increase the level of automation, level of autonomy, the automation in the system what's going to happen? I can tell you exactly what's going to happen, all right. There will be human roles developed that will -- whose job is to close, to align the context gap. In other words, their job is to make sure the world matches the assumptions behind the automation. As long as the world that the automation runs in matches its assumptions, it will do things that it really -- make everything look real smooth, hyper-smooth, hyper-efficient. The problem is if the context gaps brews, the automation will keep doing what it things what's right for the world it thinks it's in. That's actually that UAV crash I skipped over, that the example I use for this. 1999, a UAV has a problem, takes six months in 1999 to plan one UAV mission, Global Hawk. Obviously they're a little faster now. Many people were involved in it. Onboard failure, software contingency plan, return to base. Plans, stops, goes to look up the next thing to do, there's a bug in the software and it looks up nobody had evaluated the interactions across software modules, instead of looking up a taxi speed, the speed it looks up is the last to send speed. So it tries to taxi, to taxi and make a 90 degree left turn accelerating to 150 knots. Needless to see the physics doesn't work and it goes careening off and crashing in the desert. What's there. You have the automation does the right thing in the wrong world. Area 501, software blows up the rocket launch. Software modules interact. A test piece of software they didn't realize would actually be feeding live data or feeding non live data to one of the monitoring programs to say is it on course or not. So the monitoring program thinks the rocket's off course, it's not off course and blows it up. Literal minded machines, right. And that was what Norbert Wiener was reminding us of and warning us about is that danger. So people will be there to align and close the context gap. So again, what do we have, a trajectory, there's a technology change, there's certain things we know that are likely to happen or could happen. What do we monitor are those things happening, are the support features for that in place or are we going to end up again in one of these mixed-effect situations? The classic mixed-effect situation is where we end up with some successes and problems on one scale, some successes and problems on the other scale and we end up debating human error over automation. I went through this with the aviation industry, going through this in the healthcare industry. And you just point back and forth. You know, there's a screw up, well, doesn't matter just put more technology in. The screw up means that you have too much technology, stop the technology, right, just have people do it. And of course neither of those are stable, stable points out there. >> Mary Czerwinski: Okay. We should probably stop there just to save his voice. >> David Woods: I'm actually fine. I don't know why I've lost it here. >> Mary Czerwinski: Thank him again. (Applause)