 So, we started 13 hours ago this morning, and I'm either a dessert or the bitter end. Thank you all for staying with it, it's Macy's, this one down here. And it's, so my talk is at the, as Monty Python would say, something completely different. So, we're going to talk about the clinical genome sequencing at scale end of the spectrum and future genomic medicine for a few minutes. And address the issue of how we represent what we learn and what it means in terms of actionable things that can be done in a clinical setting, specifically to support learning healthcare systems, and that was a topic that came up in this afternoon session quite a bit. So just to remind you, if you're not a clinician, this is how DNA results currently get reported into clinical settings. This is one of my favorites because not only it was a PDF printed on a printer, but it was then faxed. So it's a set of black and white dots that only a human being that can read English would know how to decipher. Here's a more fancy color up-to-date version of the same problem where we do large high throughput complex screens, and then we report a bunch of words on a piece of paper electronically. What I would submit in this topic is about is that if you wish to ensure that genomic medicine will have little or no impact in healthcare, you do it in such a way that you rely on clinicians reading and remembering both clinical reports and a published literature that is completely out of scale for anyone to keep up, even a specialist, but certainly not the non-specialists. So what are the problems of treating genomic analysis the same fashion as other professionally interpreted data? For example, X-ray data comes as some technology, generates a picture, and then you get a professional interpretation of it. Well, so in this era and this arena of DNA and genomes, you're doing what informatics people call lossy compression. That is, you reduce things by throwing away data that you cannot then reconstruct. So you observe a lot of DNA features. You report only a few clinically relevant ones, and you either keep for your own research or you throw away the rest of them. The interpretation is itself inextricably bound together with the primary observations in a document format, and that format is not amenable, however good a computer person you might be to taking it apart for purposes of being able to do automated machine interpretation of what's in the document or decision support. And the big one here that has informed everything we've talked about today, and that obviously genome science is an area where a lot more is unknown than we expect to be known over the coming months and years, if not centuries, and the science is changing rapidly. Almost three years to this week, there was a NHI, a National Heart Lung Blood Institute meeting on incorporating genomic data into electronic medical records. A lot of people in this room were co-authors on this manuscript of technical desiderata for how you should get this data into clinical systems, and there were seven of them. The first is that you do need to be able to sort of dehydrate this data and get it smaller because a lot of it's highly redundant, but you need to do that in a way that causes lossless data of compression so you can rehydrate it, so to speak. We know that the methods are changing constantly, so the lab results have to carry with them the methods by which they were derived, since there are no perfect methods and they all have sort of blind spots. It's clear as well we need compact representation of clinical actionable subsets. The Informatics literature tells us that the average clinician takes about one quarter of a section to get the next idea when they see something on the screen in a clinical workstation. We need to simultaneously support both human-viewable and readable formats with links to interpretation for non-specialists and formats that are interpreted by decision support rules. We need to separate the primary sequence data, which presumably remains true if it's accurate as a laboratory observation from the clinical interpretations that would be expected to change with a rapidly changing science, anticipate the boundless creativity of nature. Maybe you know the germline genomes, not quite as stable as we thought it was, and that we actually have multiple genomes, both somatic and germline over our lifetimes, and lastly in this area support both individual health care and the discovery science that we talked about. So all seven of these desiderata are failed by a PDF. Let's focus on one in particular, this support for both human-viewable formats and interpretation by decision support rules. So here's the opportunity for NHGRI, and it is this idea that along the way of building the knowledge base about clinically relevant molecular variation is to create a scalable national capacity for genomically enabled clinical decision support and do it in a way that's actually quite unlike most decision support has been done for the last three decades, and that is to support creation of a Wikipedia-like, and I've emphasized closed loop, I'll tell you what I mean by that, computing infrastructure for guiding clinical care based on variation, that has the structural property of improving whether or not the clinicians accept the advice. Cool idea I think, and it's something where the data actually becomes the instrument of improving the system rather than the judgment of the people that use the data. Now before we get too far down this path, whenever computer people talk about rule-based systems, a certain fraction of clinicians go and hit the ceiling because rule-based systems don't mean providers need to follow rules. What rules are in a bioinformatics context is a systematic computerized approach to identifying a set of characteristics, and if they are present, then doing something. So examples of clinical interventions could include educational prompts that might occur when a clinician who doesn't know the literature happens to see a patient who has a genomic variant. They could be data-gathering prompts that says, given what is known about this patient, if you get this additional testing, it would improve the management. You could improve the certainty of diagnosis given the data currently available, and then the next case is the one most everybody thinks of, that is give a drug or modify its dose based on something about the molecular variation, or just give information relative to prevention and prognosis. So all of those are the spectrum of things that could happen when the decision support rules fire. Here's an example of a patient-specific decision support application. It's been up and running since 19- since 2010 at Vanderbilt, and so it's a pharmacogenomics app that has behind it what clinical informatics people call event monitors. That's a computer program that just spends its whole day looking to see whether something has just happened in the real world, such as a prescriber wishing to prescribe this drug clopidogrel, for someone who is sitting across from them in the exam room, unbeknownst to them the patient actually has been genotype and does have a genotype for which it is relevant that they're known to be potentially a poor metabolizer of this drug. So up pops and a little alert that gives a straightforward way of solving the problem, or you can override it if you wish. But this did not require that that clinician either read the literature or even understand what a haplotype or a genome or a SNP was. How do you go about doing this? So the building blocks of this infrastructure, kind of in a Lego block mode, that NHGRI could very much advance, would be development of the standards for what I would call decision support packages. They have three elements to them. The first is the recognition logic for phenotype, genotype, or both that means something is in the space for which there is guidance available. This is the kind of work that eMERGE and the other consortia have been doing over the last five to seven years. The second element is, well, when that rule is satisfied, what do you do next? Who gets the message and what's the message or what's the guidance for the target users? It doesn't have to just be clinicians, could be patients or families. And then the important closed loop is another set of recognition logic. In the top of the rule fires, that activates the rest of the rule to watch for something happening in the future that represents either a good or a bad outcome or a particular process or outcome variable that is measurable and observable. And then that needs to be combined with ways that mere mortals can use these authoring systems that could import these rules so that people can understand what the effects would be in order to implement them in their electronic medical record system. That needs to be linked at probably a national level, if not an international level, to a public library of decision support packages of this type that could be thought of as a Wikipedia where you have contributing experience, real world experience, doing this by lots of health care systems. The event monitors are a necessary component, but it turns out they're actually required by the new federal incentives for having electronic medical records. So the engines are there, you just need to run the program through the engine. You have the system generated alerts at the teachable moment. And then importantly, you close the loop by automated tracking, and that gives you a learning health care system that is driven by the data. Now, well, how do you get the data back? I would propose that the way you do it is have a simple quid pro quo, which is if you use the public library and you download it, you do it with an agreement that you will subsequently monitor, you know, use that same event monitor to look for the downstream effects of whether the guidance was accepted or rejected by the individual clinicians. And then, you don't have to upload every single, you know, it's not a HIPAA issue. You could upload the aggregate local outcomes. And as you gather the experience of institutions having their own experience across dozens, hundreds, perhaps thousands of patients, then you get a true learning health care system that's driven and basically learns from every single decision support event whether or not anybody paid attention to the guidance that was recommended. So to echo at the end of the day, Eric's opening of the day, why? Well, it fits the strategic plan. It would, it's resource generating. It would certainly advance the technology. It's clearly scientifically and medically relevant. And it is a natural consortium requiring opportunity. It certainly fits this complex high volume evolving science, you know, of genomics where we expect five years from now the views of the importance of these variants is going to change quite a bit. It certainly would make an NHGRI trailblazer. And it's the observation in clinical arena that although the driving spear would be genomics, that if this really works, they'll start using it for all kinds of much more mundane clinical care to systematize and learn from provider behaviors. It's important to start now because this is a big project. It would take many years to scale up to full adoption. So it really needs to start as prototypes. And then to meet Eric's last point, it's something you're not doing now. So you should be doing it. So with that, I think, you know, what it recognizes is this, the fundamental truth of this cartoon, not to scale, but it has some sensibility to it since it's appeared in IOM reports and such that as you have structural and functional genetics and proteomics, you get an awful lot more facts that bear on any particular clinical decision in a kind of stutter-stup fashion that exceeds this long-known human cognitive capacity of people's ability to deal with maybe five to seven covariates. And after that, they just start extinguishing variables in order to keep problems simple. And so it's clearly, nobody knows if it's 100 facts or 10 or 1,000, but it doesn't really matter because it's above the bounds where people can read and remember and do the right thing and only the right thing and do it every single time, as we would like to do in 21st century health care. So that's my story and I'm sticking to it and I'd be happy to answer any questions. 95 mile an hour fastball. I think one of the profound things that has happened in surgery is the adoption of checklists that came from the medical community consulting with the aviation community. As a pilot, I will tell you that if you're going to do this, you should do it like the FAA does because they deliver to pilots information of this range and complexity every day to safely transport millions of people. Yeah, so you're a pilot talking to another pilot and I have the aviation version of the same talk, but actually what I found is that the FAA does this as a centralized database and you're going to download it into your GPS NAV unit and all that stuff. It's a wonderful model for them providing a source of ground truth that then manufacturers of navigation computers make your airplane safe doing that. But there was a lot of feeling like, oh no, the federal government shouldn't own the rules. And so this one, you know, the FAA doesn't do it as a Wikipedia, but it certainly is the case that this would constitute, in essence, healthcare autopilots that know a safe path. You can still override them because you're pilot in command. You have the authority to say, no, I'm not going to follow that guidance, but at least you'd have a default that was known to me best practice initiated. So it's, yes. Dan? Yes. So I'm a big fan of clinical decision support, as you know. If I were Eric Green and listening to your talk, I would wonder to myself, how does this fit into genome institute and is there some other funder? Is there some other partner I can find that will do this? Because it's sort of, it's a little bit outside genome science, but those of us in, you know, in meshed or drowning in eMERGE will sort of attest to the sort of fundamental need for this kind of capability, but it's not analyzing genome. So how do you deal with that? Well, so I think that genomic medicine has this special place of awe and mystery among most clinicians, which is they don't understand it, they know it vaguely must be important, and they sure know it's complicated and there are a lot of facts there. So NHGRI actually has the privilege of sort of building a platform infrastructure that actually could have been built for any aspect of healthcare. I mean, we're still operating on this Ocelarian model that somehow the apogee of the profession is the learned professional and there was way too much to know 25 years ago in a lot of arenas. It's just this is the one where everybody acknowledges everyone's cognitive capacity is insufficient. So I think, in essence, it's that general awareness that this is the most complicated use case for this that gives NHGRI both the mandate and I think the privilege of being able to advance it. When other institutes would have a number of doctors to say, I'm smarter than your damn rules, I don't need them. Even though the literature says that we aren't as good as we think we are when we try to be consistent in the application of best practice. Eric. Just to follow up on that. In this domain, it may be wrapped in mystery and all, but we don't actually know most of the answers. So wouldn't it be a good idea to do it someplace where they kind of knew the answers and there was a lot of data and nobody disagreed that if you knew the right thing there were just too many things to keep track of, you could get it right. We're sort of figuring out this stuff along the way. There are variants of unknown significance everywhere, combinations we don't know and building a decision support system when we don't know the right decision might not be the best way to convince people. Oh, but I think that the fact that there is so much unknown of unknown significance that is expected to become known over the ensuing decades is exactly the reason why you would build it here because there's already enough data that is not widely appreciated by practitioners even in just pharmacogenetics. So I actually do think that although you see it as the place where you wouldn't wish to start, I say that's exactly the reason why you'd start here as opposed to a very mature area where people would say, but I already know these. But there are great diagnosticians and then there are not so great diagnosticians with regard to internal medicine and if you could get a decision support system that made a mediocre diagnostician into a vastly better diagnostician, it would have huge impact. And you could imagine doing that today. Well, except there actually is a 35-year history of doing this with systems like Oncasin and Octobarnet at MGH, the de-explained system, which was published in the New England Journal in the late 80s, that it could perform as well as expert clinicians at grand rounds of the mass general. And they've never taken hold and the reason they don't take hold is by and large because they are kind of a threat to clinician autonomy, who they believe they're actually smarter than the rules and they sort of believe they're better at managing mature bodies of knowledge than they are because human beings are really good at detecting patterns. They're not really good at working with long lists of interrelated things like drug-drug interactions, all that stuff. So this, I think, speaks to this fundamental human fallibility of our believing that we are more consistent and better than we actually are when we try to do something as complicated as this. And to bring it back to aviation, they actually recognized this quite a long time ago because the classic here is the Tenerife accident, which was the world's most perfect pilot, generated the worst air disaster in history. He was the chief of aviation safety for KLM, pushed the throttles forward and killed 500 some people. And it was because of human fallibility and they realized that human beings can't basically be trusted with very, very complex bodies of knowledge. You've got to have some systems infrastructure to make them more reliable. All right, one more and then I think I'm way overdone here. Never mind. I would just say as a clinician, I agree with Eric Lander that I think doing this in an area where we have no idea what the right answer is has the danger that all you'll get is a Wikipedia of opinions and that I'm not sure that that actually results in high quality medicine. And so I might, I agree with everything else you said except that it seems to me where these in absence of data, I don't really see how crowdsourcing opinions gets us anything but who likes Instagram pictures, but not actually high quality healthcare. But this isn't crowdsourcing opinion. This is crowdsourcing things that actually occurred. So I mean this is a set of observations that are systematically acquired. So the power is in the data, not necessarily in the interpretation of the data. Yes, yeah. Systems from the late 80s and I think I would avoid the word rule based because I think one of the problems there was the whole idea that the practice of medicine would be reduced to rules. And I can understand why there was a clinician revolt back then no matter how good or not good the rules were. I think we can provide decision support without casting it as a rule based. Yeah. Yeah. And I actually had that on the slide that the rules actually relate to the recognition of the scenario not to the fact that you have to do something that represents a rule. Just my opinion. Thank you for your attention. All right. Thanks very much everyone for staying to the end. We will convene again at 8.30 tomorrow morning.