Welcome to the site of accents podcast. Where we explore emerging ideas from signs, policy economics, and technology. My name is Gill eappen. We talk with woods leading academics and experts about the recent research or generally of topical interest. Scientific senses at unstructured conversation with no agenda or preparation. Be Color a wide variety of domains red new discoveries are made. and New Technologies are developed on a daily basis. The most interested in how new Ideas Affect Society? And, help educate the world how to pursue rewarding and enjoyable life rooted in signs logic at inflammation. V seek knowledge without boundaries or constraints and provide unaided content of conversations bit researchers and leaders who low what they do. A companion blog to this podcast can be found at scientific sense dot com. And displayed guest is available on over a dozen platforms and directly at scientific sense. Dot? Net. If you have suggestions for topics, guests at other ideas. Please send up to info at scientific sense dot com. And I can be reached at Gil at eappen Dot Info. Mike yesterday's Dr Mark Hoffman, who is a research associate professor in the University of Minnesota Against City. He is also chief research inflammation officer in the children's Mussa hospital in Kansas City. Kiss research interests include health data delayed indication sharing initialisation Boca Mark. Thank you for inviting me. Absolutely. So I start with one of your papers Kato you need the use by our system implementation in defy date data resource from hundred known athlete off my seasons. So Michio inflicted. Data aggregated for marketable sources provide an important resource for my medical research including digital feel typing. On. Like. Todd beat to from a single organization. Guitar data introduces a number of analysis challengers. So. So you've worked with some augmentation log and in almost all cases be used. Data coming from that single macy's listen primary care behavioral. Or specialty hospitals and I always wondered you know wouldn't be nice. Get a data set. That sort of abrogates data from the radio on-ice. Asians but a lot of different challenges around that. So you wanted to talk a bit about that. I'd be happy to the resource that we've worked with. Is primarily a called health fax data resource. It's been in operation for almost twenty years. And the the the model is that organizations who are. Using these Turner Electronic. Health. Record. Enter into an agreement was turner they agreed to provide data rights to sern are. The identifies the date of affords aggregated into this resource. And certner provides data mapping, which is really critical to this type of work. It also the aggregate the data. And for the past probably six years. Then, they provide the full data set to especially academic contributors who want to do research with that resource. And I've been on both sides of that equation Lead that group during my career there, and then now I have the opportunity to really focus research on that type of data. So before we get into the details smog so e Itar Systems. So this is. Essentially patient records. So he gets dated like demographics out family history, surgical history hats, medications, lab solves it could have physician nodes no snow. So it's it's a combination of a variety of different types of data, right? A couple of things on the examples you gave it includes demographics. Discreet Laboratory results Medication orders. Many vitals so If access the blood pressure and pulse data. It does not include text notes because those can't be. Automatically identified consistently. So. We don't have access currently to TEX notes. Out of an abundance of caution. That his Hobby Stephen, physician writes something down they could use names they could use inflammation that could then point back to their. Patients Makita Perspective been the data's aggregated, the primary issue shoe that date has completely the identified, right? Correct. So. So yeah. So the data that we receive there's eighteen identifiers. Hip requires be removed from data. And those include obvious things like name address email addresses are another example One of the. Things. That is also part of the benefit of working with this particular resource. The. Dates of clinical service are not allowed to be provided under hip. White is done with this resource that allows us to still have a longitudinal view is. For any given patient in the data set the dates are shifted by A. Consistent. Pattern that for any given patient it can be. One two three four five weeks forward or one, two, three, four or five weeks backward. But that preserves things like day of the week effect. So for example, you see -nificant increase in emergency department encounters over weekends and you don't WanNa lose. Visibility to that. but it also allows us to receive. Very, granular early time stamped events in so. We can gain visibility into the time that a blood specimen was collected, and then the time that the result was reported back. And so we're able to do very detailed analyses with this type of resource. Right right and I don't know the audience our market is fragmented. Tau himself e Amorebieta providers out there. and so two issues. One is sort of. Standardization as to how these databases are designed and structured and others even that standardization that the actual collection of the data. In itself is not standardized played. So vk CAV vk potentially lot inability coming from different systems. Correct and that's part of what the paper that you mentioned Evaluates so. Often, night you out in the field in conferences you hear. Comparisons kind of lumping all organizations using one. Vendor lumping all using another together but as you get closer to it, you quickly learn that. It's not even clear. It's within those. Vendor markets. There's variation from organization to organization in how they use the e Hr and so. Because the identities of the. Contributing organizations are blinded to those of us who work with the data. We have to be creative about how we. Infer those implementation details, and so with this paper, we describe a couple of methods that We think move things forward towards that goal. Yes. So I'm not really familiar with that. So you mentioned a couple of things here. One is the the merge network. So this initiative including electric medical records and genomics network and pc off net the national patient, centered clinical research network support. Decentralized analyses that goes disparate systems by distributing standardized quotas to site. So this is a situation where you have multiple systems sort of. Communicating with each other and this net folks at allowing to sort of quickly them In some standardized fashion. So In this type of technology, there's janitorial core models. One is the. Federated or distributed model, the other is a centralized data aggregation. So there are examples including those that are mentioned in the paper where. Queries are pushed to the organization and. They need to do significant work upfront to ensure that there are standardizing their terminologies the same way. And once they do that upfront work than they're able to perform the types of queries that are distributed through those. Federated Networks. With. Okay. So that just one click on so that the police have standardized. So all on the at Josh site, then they have like some sort of a plan slater from from Stan Day squatty do all the data structure. And in many cases, they work through an intermediate technology. that would be. In general, consider it like a data warehouse. And so the queries are running against the production electric. Health record. That has all kinds of implications on patient care where you don't want to slow down performance. By using these intermediaries They can receive queries and then Follow that mapping has occurred. Than, they're able to to run those distributed queries. Okay. And the other model is You know. You say the g through the medical quality, improvement consortium and sooner to the health facts initiative. So this says in Sodas case, for example, in swags. This is essentially picking up data from the right deals, clients and Dan standardizing and centralizing data in a single database is that that is correct. One benefit of that model is that Organizations who for example, may not be academic and don't have the. Resources to do that data mapping themselves by handing out over that task over to the vendor you get a broader diversity of the types of organizations so you can have. A safety net hospitals you can have. Critical access rural hospitals, and other venues of care that are probably under represented in some of those. More academically driven models. And clearly the focus on healthcare about I would imagine applications in pharmaceutical out indeed to right I. Don't know if it s use and bad direction there has been some were performed with these data resources to. Characterize different aspects of medications, and so it does have utility in value. In a variety of. Analytical contexts. I was thinking about you know a lot of randomized clinical trials going on into Kuwait context and One of the issues of dispatch seem development toils that are going on that one could argue the population there are not really well to percents. it may be number by Auditees, men, people that deputy existing conditions. and. So he will serve at my come out of facedly trial. granted might work for the population. Tried it minority have sufficient? more largely. So I wanted this type of well I guess we don't really have an ID there right. So clearly, you don't know who these people are but they could be some clustering type analysis that might be interesting weight from It's very useful for Health Services Research and for outcomes research for you know what I characterize digital phenotype being. they can then guide. More, more formal research. you know you can use this type of resource to. Make sure. You're asking a useful question and make sure that there's likely to be. Enough patients who qualify for given study. Maybe you're working on a clinical trial in your casting your net to narrow you can. Determine that with this type of data resource. And is the eight tiff date who has access to it typically. So for this data resource on, it's through the vendor so. You need to have some level of footprint with them. which is the case with our organization. They're definitely a broadening their strategies. So they're. Gaining access into health systems that aren't exclusively using their electronic health records so. It's exciting to be a part of that that process. and to again work with them to. Analyze the data. I think. To the example you gave a formal randomized trials. In key part of what were growing our research to focus on is because this is real world data. You learn what's happening in practice whether or not it's well aligned with guidelines or formal protocols. And doing that there's many opportunities for near-term interventions that can improve health outcomes simply by. Identifying where providers may be deviating more from. Best Practices in than taking steps through training and education to kind of get them back towards those best practices. This data is a fresh on a daily basis. It's not. It's because it's so large and bulky? Typically we've received it on a quarterly basis in since it's retrospective analysis that's not been a major barrier. But. mechanistically, on onto soon aside is data getting sort of picked up from this system that it's harvested every day and then it's aggregated bundled and distributed on A. On a different timescale. Okay okay. So. From again, going to the, it's our system designed issue and implementation You say many HR systems comprised of more news at specific clinical processes or unit such as Pharmacy Laboratory or surgery talked about that. But then then people implement them this of fashion right they they implement modules by that can be a factor or sometimes they may want. One vendor for their primary electronic health record, but another vendor for their laboratory system. and so that's where you don't see a hundred percent usage of every module and every organization. And detailed number of different you know sort of noise creating issues in data one. This is icy speech over from ICT denied ten. and I don't know history of this but this was supposed to be speech with sometime in twenty fifteen. That's correct. So there is A. You know. There's a date in October of Twenty fifteen where most organizations were expected to have completed that transition. When I see with researchers who aren't as familiar with the you know the whole policy landscape around `electronic health records that? you can imagine researchers who assumed that all data before that date in October is is nine and all data after that date would be icy the ten. While we demonstrate in this paper, is that that transition was not Nearly, that clean and it was a much more, you know there are some organizations who just It the bullet and completed in twenty fourteen, and there are other organizations that were still lagging. In. Two Thousand Sixteen. Potentially because they weren't as exposed to those incentives in other things that you know stipulated the transition so. Part of why were demonstrating with that particular part of that work was that. you know these transitions aren't always abrupt. Yeah and and and so that is one issue and then you know a lot of consistency inconsistency issues fade. So we see that in in single systems and one of the items note here as you know if you think about the disposition code for death. you could have a right your race supercenter, right? It's a death expire expedite at home hospice, and so on. if this is a problem for a single system, but then many think about aggregating data from multiple sources this this problem sort of increased exponentially. Absolutely. So one of the challenges with documenting and and finding where you know if a patient has A deceased that. There's just multiple places to put that documentation in the clinical record. The Location in the record that. We have found to be the most consistent is what's called discharge disposition. By as we show in that analysis, that field is not always used document that and so if you're doing outcomes research and one of your key. Outcome metrics is death. And there are organizations that. Aren't documenting death in a place that successful. You should filter those out of your analysis before moving forward. And so part of what we wanted to promote is the realization that. That's the type of consideration that needs to be made The four. Publishing. Your data about an outcome metrics like death that. You're not. If you're never gonNA see that outcome it doesn't mean that people are. Dying in that particular facility, it just means it's not documented in the place that successful. Right. Yeah. So you know you on your expedience. Unique Position Mark because you you look at it from the from the vendor's perspective you're in an academic setting you're also in practice in a hospital. What's your sense of these things improving the on a track of getting getting this more standardize or it's camping in the other direction I think in general there is improvement I think The. Over the past eleven years through various federal mandates, including meaningful use and so forth. Those of all incentive organizations to utilize. Standard terminologies more consistently than was the case beforehand. I think there's still plenty of room for improvement and You know it's it's a journey, not a destination, but I think things have improved substantially. I was wondering there could be some applications of artificial intelligence here to In a clearly TATECO systems and you'd like the most them pity human resource intensive Yvonne to get it completely right. So one question would be you know, could be actually used a Dick needs to get it maybe ninety nine percent white. And that the human deal with exceptions I definitely think that that's an exciting direction that You want those a algorithms to be trained with good data, and that's a big part of what's motivated us to. Put this focus on data quality and Understanding these strange nuances that are underpinning that date has so that. As we move towards a in machine learning and so forth. We have a high level of confidence in the data that's training those algorithms. Right. Yeah. I think that a huge opportunity here because it's not quite as broad as NFL, not natural language processing it is somewhat constrained. that is a good part of it. The back part of it is that is highly technical. and so. you know some of the techniques you know you can have a fault tolerance in certain dimensions such as you know, misspellings lack of gambling and things like that. But as you have Heidi technical data, you cannot apply those principles because he could have misspelling the system may not be able to. Get, sometimes, and that's where you know I think. It's totally feasible to use. Resources to you know when you're dealing with. Tens of millions of patients and billions of detailed records. Using a I'd even identify those patterns of either. Inconsistent data or missing data it's also very powerful just to. kind of flag in identified. Areas that need to be focused on to lead to a better analysis. Greg Wait Be Hefty. Use that information somehow did is a belt of information that you know and so it just filtering into decision processes that the are really losing it. So hopefully getting improving in that dimension I've jumping to another paper bittersweet interesting. So it's entitled rates and predictors of using opioids in the Emergency Department Katrina Treat Mike Dean in Young Otto's and so so this is sort of a machine learning exercise you have gone through to locate you know coup is getting prescribed. OPIOIDS water the conditions for the Democrat not Nestle demographics but different different maybe age and things like that gender. and and then ask the question desert has some effect on addiction. In the long term rights. So that project To great example of team science though. We. Assembled a team of subject matter experts in neurology pain management. And Data Science and. The neurologist and pain management experts. Identified an intriguing question that we decided to pursue with data. In their question was. Based on anecdotal observation and so we thought it'd be interesting to see how well the data supported that. Observation is that. for youth and young adults Treated or admitted into the emergency. Department. With a migraine headache that. All too often they were treated with an opioid. And so we Use the same day to resource that we were discussing earlier. To explore that. Question. And using data from a hundred and eighty distinct emergency departments. We found that on average twenty, three percent of those youth and young adults were treated with. An opioid medication while they were in the emergency department. In general, it should be almost zero percent in general. There's really Better medications to us, four people presenting with a migraine. and. So this fits into obviously the OPIOID crisis it. it demonstrates the. Scenario describing that. You know using real world data. You can identify patterns of clinical behavior that. Don't match guideline. And the good news is that the? correctable and so through. Training and communication there's great opportunity to. To, manage this. Really. Striking. So fifteen thousand or so inevitably the encounters. And nearly a quarter of this encounters you say involved inoculate. and these are not just Misha and Congress right. It is not filtered down to migraine encounters. Okay. Okay. So these fifteen thousand just might in encounters might vein being repeating disease So once you. If you make a statement and. This or not Easter conditioning issue here. So you get your pain, you go to an emergency department and you get treated with an opioid you get quick tactical relief. From pain. auditing condition expect that in the next episode. So you can say we didn't pursue that particular question, but that is Definitely key part of. Managing the OPIOID crisis is that drug seeking behavior and so Part of our goal was to quantify that and use this as an opportunity to educate providers that. You really shouldn't be treating migraines with an opioid in there are better alternatives and. So we we felt that this was an important contribution to that national dialogue, but we didn't specifically pursue the question of whether the patients we analyzed. Within. Encounter show up Subsequently. With the same symptoms. Right right. Yeah you it develop into period when problematic patterns of drug use comedy. FEST MERGE THE PREVALENCE RATE OF OPIOID misuse estimated to be two to four percent and debts in each goofy just young adult drew from overdoses are rising. and. You say that literally prescribe IOS has been slumping loose future opioid misuse by thirty three percent. Betas Mehta say really huge number. I think just validates the importance of this of this work. Interesting mark. I don't know you exploded on data. Last the question if you look at the aggregate data, it'd be flying opioid. Misuse. what percentage of the total number. Actually started from. You know some sort of medical encounter has mike or some sort of. related encounter that could be completed otherwise was three a bit opioid. in that encounter documented resulted in that misuse. So what so If you look at the active misuse problem that we have today. do you have a sense of what percentage of that goal is actually started I? Think the exciting thing about this type of research is for everyone questioned that you pursue you have. You have ten new that you can pursue. We haven't. Delved into that specific area, but it's It's very ripe for further analysis and A considerable part of where I end my colleagues and our time as. We do this type of work to get an initial analysis published. And then You know in my leadership role I just WANNA. support people like my colleagues on this paper Mark Connelly Jennifer Bickel. in in using data to. Support their research into identify those follow. I mean, he tests policy implications. So it's sweet important work. and. If you find it direct relationship here than you have to ask you know from from a medical perspective what is right intervention? maybe is not just added of care just best practice but clearly should be the bay You know things should be looked at you say you're American Academy of Neurology has included avoidance of using opioid to treat gain one of stop top flight choosing wisely recommendations. For high-value duck in this gives Really evidence to to support that. The other thing that's really intriguing is this level of variation from site to site in. Some Sun facilities are very much aligned with the guidelines. Others are at the you know well, above twenty three percent. And that gives an opportunity for a really precision. conversations about you know, where does our organization stand on that spectrum? Yeah that's a that's an interesting avenue to right. So you know one could ask he says some sort of push sliced Intervention if we can fly goal of patients who who had gone an opioid sexually don't have an addiction problem. that as you know Anna, the kofoed does. if you can fly those type of patterns than you can think about. A customized within electronic health record systems. There's. The ability to provide decisions poor. There's certainly phenomena called pop up fatigue were physicians. You know they don't like having so many pop up windows but at the same time. It's Within the capability of an e e Hr to do that if then logic if patient has. migraine medication order equals opioid. encourage the provider to pause and reconsider that. Right, right and so this is supervised machine learning type analysis where so you have. you have number features that comes directly from each else. So each sex race ethnicity. insurance type. Encounter prostate suggest duration. time of the year and so on. and you have labeled data in this case I guess you have able tater because you would know if op- inscribed on trade. Okay and so are the two questions here. One is to ask the question given a new patient and those features. you could assign a probability that that patient will be prescribed will. Definitely. Impress the data from that predictive Minds. Right and then can you so that data definitely tell you if the patient is going to progress into some sort of an addiction issue. So. Earn Predicting Substance Abuse. So. Yeah. Yeah. Yeah. There's additional diagnosis codes that document. whether a patient has a history of substance abuse disorder. and. So it would be feasible to. Identify the with those diagnosis codes in than really look at their prior history. Of What other conditions were they treated for? What medications were they give in? to develop that model. One of the things in this case that helped with this study is that just in general, it's not advised get. So there are other things that are much more of a gray area. Or whether opioid is as useful, but in this case. The really not. Considered. To be helpful for migraines compared to other options and so that help us have a fairly clear cut scenario to do this work. Yeah. This this won't be the data like you say once you do something like this, you have been other things you could. You could stop asking. So unquestioned that that been to my mind as you know, how did they hugged the actually prescribing opioids? Is it the patient asking for it all so? Off that was another scoping thing with this project is focused on what happens within the emergency. Room. So it's it's. Really, medication order in administration that happens. In that emergency room setting. Whether or not the patient. was. Requesting that you know if they came in and said, this has worked for me before. Can I have it again? we don't have visibility to that. Right. Right. And so from a practical perspective So the the analysis that you did slightly ended up with the Family Clyde power we think it is. Compelling. Pretty compelling. So as as a new patient gets into e D either high. and what I mean by that probably is if there is a history of substance abuse property. the physician has really think twice about. The use of may be the well, and in this case, even without that history. Just because it's not considered to be an effective treatment. You know encouraging them to pause in that decision making. In this particular case is as effective as wall. Right. So looking forward. In if you think about both of these issues, one is the data quality data aggregation data standardized recent problem in the the right of Utah Systems have did that the talked about? And then if we can get to a level that we can look at cross a large data set. Beacon, ask. More. US specific questions, treatment. Optimum treatment type questions. subpoenaed. US The mark big think B be hunting. Certainly, the volume and variety of data that we're able to work with will be even greater I, think the. Opportunity To. Look, holistically at how upstream data capture. Effects Downstream data. Analysis. example I frequently give is if we have a Aggregate Data said we identify. Ten patients whose way in that data such shows up as being. Something that's completely infeasible. let's say they're documented is being. Fifty year old person who weighs two pounds. Clearly air. What's important is? Creating the process to communicate that back upstream. Because that clinical decision. Support. Many drug dosing things are evaluated using weight based logic and so. That same logic that's Evaluating the appropriateness of dosage. It's going to be running against an incorrect value in that may or may not always be visible. So I really am intrigued with that holistic opportunity. In it I am I remain just we have three or four additional papers coming out. About other examples where Provider behaviors not aligned with Best Practices and I'm just excited about you know when you compare that to how long it takes to develop a new drug or how long it takes to. To a really long term research. This research has the opportunity for a pretty quick turnaround on an effective intervention. A really that. Other so much that right. Providers. been taught in a no, but they're. Not always using that in practice and so to help them. Identify, those topics in just modifying behaviors is. In the scheme of things, it's a very straightforward way to improve. So. You know the entire spectrum from essentially getting the data. Right or cleaner like you know Missa mischaracterized or miss input data like wait or something like that. To to get. Better diagnosis better treatment modalities. policies there and from a femme perspective clearly inflammation therefore clinical trials. I was even thinking about drug interaction type. Inflammation. I haven't been involved in the former de for awhile but. Typically, this type of data doesn't get back into automatic processes that fast but I think that is all I know there's strong interest in Pharma in. Working with this type of data there a again looking at real world behavior. This is an excellent resource for off label medication use at. you know where Pharma's Always interested in repurposing existing medications the. Regulatory Processes, much more straightforward for that because the safety is already been. Evaluated and so. The. Significant Opportunity With this, there's also just exciting. Patterns of you know. What are those unrecognised correlations? That's where the machine learning opportunities are really exciting where. You know we're not always asking the right question. And the data can show us what we should be. Yeah exactly. So if the machine a sort of red flags something or create hypotheses. that Cubans have missed sometimes, those types of things are extremely powerful. because maybe that sometimes it's countering tutor. and so we all look at data with an Incan bias. The beauty of machines that at least on the surface began deploy Michigan. This volume of data. Techniques like machine deep learning can recognize those subtle but consistent associations. Wait quite. Excellent. Idea this has been great mark Thanks so much time with me. I enjoyed it very much. Thank you. But