Healthcare leaders today are faced with increasingly complex and unprecedented challenges. With COVID-19 taking the world by storm, the need for an intelligent system of insights that can proactively deliver actionable and real-time knowledge on patient populations is imminent to providing better care. Multiple Health systems across the country, such as Indiana University Health, are turning to technology and leveraging investments in Data & AI to adapt to the need for readiness, preparedness and response to the rapid surge in patient volumes.
KenSci recently launched a Realtime Command Center for COVID-19 Response to support our customers during these challenging times. This solution converts near-real-time (NRT) messages like HL7 (ADT, ORU, ORM etc.) into a single common data model using the FHIR specification and provides health systems a real-time view into bed management and capacity planning along with important insights like overall ventilator use, surge projections and discharge planning insights. The Realtime Command Center is deployed on top of the existing KenSci Platform and infrastructure, built on Azure Databricks. Join this session to hear about how prior investments in Data & AI have enabled Indiana University Health to rapidly extend its pre-existing infrastructure and self-service offerings to react to the need for COVID-19 focused dashboards and insights. During this session, you will learn:
– Thanks everyone for having us. We’re super excited to be at the Spark and AI summit, although virtually. My co-presenter Tony and I are really excited to walk you through some advancements that we’ve been making in healthcare analytics, all around data and AI, and especially around a time like this when the world is accepting a new normal. My name is Rohan D’Souza. I represent KenSci today and we’re in the machine learning and AI space specifically in healthcare. And we’re excited to have everybody here for today’s session. So today, we’re gonna be talking about how a large health system instituted a rapid response operations, using data and AI at the underpinning and at really to come to the forefront to transform healthcare and healthcare delivery.
Like I said, My name’s Rohan D’Souza, and I have the pleasure to speak with you today on this topic. Today’s agenda will be covering three things. First of all, just talk about data. How we leveraging it and the AI investments that IU Health has made, especially that was tested at this time during the COVID-19 response. We’ll then talk a little bit about this idea of improving hospital operations by leveraging AI and machine learning to assist in the better delivery of care, ideally, at a lower cost and hopefully at some significant impact and outcomes. And then, the baton will be passed back to me, and I’ll talk to you a little bit about what we as a company have done around real time analytics to assist hospital administrators, frontline staff members to leverage insights and value from their data to help with bed placement and bed management, which became really critical when COVID-19 hit us, at least here in the United States. So with that, I’d like to turn it over to Tony.
And he’s gonna talk to you a little bit about IU health and all of the amazing stuff that they’ve been up to. Tony. – Great to be here. As Rohan mentioned, we think we’ve got some exciting information to share here on work that we’ve been doing, both internally with our data warehousing capabilities, and then in partnership with KenSci, and some of the AI and ML work that we’re doing. As Rohan mentioned, my name is Tony Pastorino. I’m a vice president within our information services organization at IU Health. IU Health is a system of of 18 different hospitals in the state of Indiana. We serve about a 1,000,000 residents, and we have an overall team member count of around 30 to 34,000 folks at IU Health. Largest health care provider in the state of Indiana. So we wanted to take you today through a little bit of background on our journey at IU Health, with data warehousing, business intelligence and most recently, our jump into the AI and ML space. We started out back in the early 2000s at IU Health with a sort of in congruent set of databases, database servers, WotNot.
Very minimalistic, probably not even considered officially a data warehouse at that point in time. Around 2010, into the teens, the decision was made to bring in a platform that we could build a enterprise wide data warehouse on and begin to leverage the data assets that we have at IU Health. So we used Microsoft’s analytics platform system, their APS platform built a fairly large data warehouse and began the process of curating data that we needed to run analytics on, building internal capabilities around the curation of that data, and also the business and clinical intelligence being generated off that data. And then another thing that we started that I’m gonna talk about a few times in here was what we call a pretty large self service offering over our data. We have a community of about 200 plus analysts, call them community, data analysts, community data scientists across IU Health, that we have partnered with and work with on a day in day out basis, to ensure that we have the data they need, and that we have the sources of truth that they need, and that we’re able to support them from a tooling perspective. And they work hand in hand with a portion of my team that does business and clinical intelligence also. About, I guess it was beginning of 2019 when we actually started the transition, but we made a transition from that on premise, APS platform, and moved all of our data and all of our reporting assets up to Microsoft Azure into the Cloud. This is obviously provided us with essentially unlimited compute, unlimited storage, very quick access to new tooling that becomes available out there, and has been a real help to us both from a supportability and sustainability standpoint with our data warehouse and our data at IU Health. Kind of talking about– what I just talked about there was our journey, sort of the building of the data warehouse and the movement across different platforms and really getting us to that modern data warehousing, environment and platform that I’m sure many of you have in place and is obviously essential to be able to gain intelligence and gain meaningful insights into your data. The other thing that we have invested in, as I mentioned, was the self service component of this, making sure that we are able to support a large group across our overall health system that does data analytics. There is no way that I could ever support in a appropriate fashion, a group of analysts within a centralized business intelligence hub that has the expertise in every one of these components of data, whether it be clinical data, data from our ERP systems, infection prevention data, there are folks that are obviously tight with that data in each of those areas, and are the true subject matter experts. And so what we wanna do is curate the data, put it in a place where they can join it with other data and make it available for them and make sure they have the tools available to them to do analytics on that data. And then the last piece, and this will be the sort of second half of our presentation here, where Rohan will take you through some of the work that we’re doing, from a machine learning and advanced analytics standpoint. One of the big areas that we are currently involved with is putting together predictions around length of stay that our care managers can leverage in their day to day work. So a little bit behind the echo system that we sit on at IU Health and not going to get in a big technology conversation here, but wanted to show kind of how data flows into our data warehouse and how it gets utilized. We bring in all sorts of data both structured and unstructured. I’ll show a slide here in a few minutes, that kind of shows the breadth of the data that we have. But all that data comes in and runs through an ETL process, leveraging some technologies from Informatica Cloud, and then Azure Databricks to bring that data in, process that data and get it into a data store, an Azure datalike, where we can leverage that data. We then go through a process, I’m gonna talk about some of the various layers of data that we serve up in a moment here. But that all happens in that this ingestion, prep, and then model and serve space here. And we serve that data up in a usable fashion up into Azure data warehouse. More than we connect to it with tools, obviously, to visualize that data. The main tool that we are investing in at this point is utilization of Microsoft Power BI on the visualization front. So I mentioned a second ago, the different layers of data that we bring in. So obviously, there’s a lot more to this, as everyone knows than, you know, copying tables from a source system into another database, and then connecting to those. We bring in that raw source data. Again, you know, some of its structured, some of it unstructured. IoT data is starting to become a part of that fold. Right now we have 31 different sources in production, we probably have 20 plus sources being built out right now to add on top of that, we refresh that data every 24 hours and we keep about 10 years of history for those various different data sources. We then take that data and we move it into what we call our integrated layer or our semantic layer that allows those analysts that I spoke of earlier to get in there, and actually leverage that data and tie it out and join it together and visualize that data. That integrated layer or that semantic layer gets refreshed and rebuilt every 24 hours, and a bit of a batch situation. We also have a whole set of cubes, about 10 different cubes that we’ve built, that allow for even easier access into that data where we’ve pre-joined data and rolled that out in tabular models for folks to be able to get in there and access it (murmurs). And then we’ve got a third and fourth layer that I’ll put there kind of on the same level. Reporting layer allows that community set of analysts that I spoke up earlier to get in and add some of their own data, join data in specific ways that they need it, and really support some of the departmental work that they’re doing.
And then lastly, we talked about ML analytics, we’ve got a layer that continues to be built out off of that integrated layer where we de-identify data so that we’re able to leverage it in some of the AI and ML models that are being built in IU Health.
So, obviously, everyone in their own way has experienced COVID-19 over the last few months. And obviously as a health system, we were impacted and really had to bring the whole organization together very quickly to ensure that we were able to serve our patient population in our community during the pandemic. The community of data analysts that I spoke of earlier, really allowed us to do this. It was really exciting to see something that we’ve been working on and forming over the last few years come to fruition and have a probably the craziest use case that we’ll ever experience put it to the test. Like I mentioned, we have about 250 analysts across the organization. We quickly formed a group within my area to pull those folks together, worked on a daily basis with our executive leadership, trying to understand what views of data that they need to see. Was it you know, patient traffic? Was it PPE availability? Various different things and quickly within probably a six to seven day period of time, we we’re able to spin up a dashboard that probably has, I would guess, at this point about 30 different reports and views of data in it. There’s a little snippet of it on the right side of the slide here. But that has become an absolutely we could not have gotten through this without the insights of the data that we were able to put out there for our executives in some very difficult decision making that they had to go through.
The only other thing I’ll comment on this is this has been such a success that we are now forming a group internally, representation from various different functions at the executive level, to take this beyond COVID, and really get this to be our enterprise day to day operations dashboard, tied to our goals and tied to our day to day work that we do. So really exciting stuff, really good to see all that technology that I talked about earlier on in here, really being put to work and really allowing us to meet the needs that this pandemic threw at us. So next, I’m gonna flip this back to Rohan, and he’s gonna start to take you through some of the AI and ML modeling and platform setup that we have been partnering with them on as we look at how to improve our patient flow through our facilities. – Tony, that was great. Thanks so much. Great to see all of the amazing work you guys have been doing. So much amazing stuff happens in the Hoosier State. So really awesome stuff, to see what can be done when the call to data comes through. So I would like to switch gears a little bit and talk about you know, going back a couple years when we started off with IU Health. We were asked to do things around how can we improve hospital operations. And you know, the target always ends up being the symptom versus the actual problem, right and the symptoms typically end up being executive level dashboards that show really long lengths of stay, maybe poor discharge planning etc. And we had the opportunity to work with Tony and his team and lots of great collaboration between clinical users, operational users and data users to try to put together this really complex puzzle of hospital operations and hospital throughput. Where one, you’re working in a very noisy environment. There’s lots of things going on, and there’s lots of things that are that are changing, very stochastic environment. Two, you’re dealing with folks that tend to have the risk of alert fatigue. So there’s lots of initiatives that need to be carefully thought through before we establish an AI strategy for hospital operations. And then last but not least is, everything is interdependent, right, there’s tons of stuff that are happening between the emergency department to the case management team, to the triaging team, to the providers that are admitting versus the providers that are discharging and from a data science perspective and an ML perspective, you might think of this as a pretty simple regression problem. But it’s a lot more complex when you layer on the actual healthcare meaning of this entire end to end process.
So one of the things I’d like our audience to learn is hospitals are basically, you can think of them as a linear flow of patients safely moving through the care continuum, but with a risk of a lot of bottlenecks that create a significant cascading effect that prevents an entire hospital system from choking up. And more specifically, things that could be pretty dangerous, like actually going on diversion and asking people not to come to the hospital, and eventually making the local news to say that my hospital actually was not accepting patients. Nobody ever wants to be in that situation. And it’s usually because of things that could be controlled, far more upstream. And we believe that predictive analytics could help solve for one of those challenges. So here you can see that this is a simple flow of how information can get choked up, and we were bullish on this idea with the predictions that we could do in the top right of the slide around, could we predict how long we think a person was going to stay in a hospital? Do they have a risk to come back within 30 days? Where are they likely going to be discharged? Could really provide clinical frontline staff with certain actions that they could be taking in order to avoid some of those bottlenecks that will create a significant problem for the health system. So how did we do this? Right, the first thing that we did is we took advantage of the amazing infrastructure that Tony and his team had been working on, for gosh, maybe a decade, right? All of this data that was now coming in from these various different data sources, were in a place for data scientists to now come in and do some pretty incredible analytics. We started off by first doing this in an offline setting. And this is your traditional machine learning training process, where you take a bunch of data, you run it through a bunch of experiments, and you interrogate the experiments. And most people are excited about the results because you believe that a model could actually improve care. But then when you put it into a production setting, it becomes a little bit harder. And so what you see here is a view in into the dashboard that we provided IU Health’s care management and case management team, where they can walk in into their morning huddles and leveraging the Microsoft Power BI platform, we’re able to very quickly present the prediction endpoints. So now case management knows that there are certain alerts going on for their patients that they can take an action on right within the point of care. So right when they go in into their morning huddles, or their stand ups, or whatever ad hoc meetings that they have to help plan for safe patient throughput.
And some of the learnings that came out of this is, this is the beauty about putting in an online learning system is things start to get better, right? As the model starts to see things in the data based on actions that people take, it really starts to improve itself, it starts to compensate and compromise where required, but more importantly, it takes very safe imputations to handle for abnormalities that exist within a system. Take, for example, when COVID-19 hits a health system and they shut down elective surgeries. But what is it due to a length of stay model that needs to be factored in. This is not something that’s sitting in the background that’s completely offline. And then you know, over time, we’ve been able to improve a lot of the performance both from a pipeline perspective, where you’re taking a massive amount of information and can generate a score in just a matter of minutes, and also see that the model is performing for the bulk of encounters that the hospital typically encounters on a day to day basis. In our case, about 85% of these encounters were coming in within a 1.6 to 1.5 day range of an actual length of stay, and some pretty incredible results that could only be possible with an online tuning system versus an ad hoc deterministic model. So what did this teach us? As a company we started to get asked the question, well, how do we take this information and make it more real time? How do we present this much closer to the bedside and we were working on this solution for the last couple of months. And you know, serendipitously COVID-19 hit and the need for real time information became much more of a must have versus a nice to have. So we were very quickly able to spin up our infrastructure and take a couple of things and put it into a pipeline.
Number one, we empowered health systems to identify what was going on with their community and look into areas of opportunity for outreach coordination.
Ambulatory clinics that need to be spun up temporarily in places that had a higher risk or a higher precedence of high risk patients. Right, the early warnings that were coming out of Wuhan, and in Europe, were telling us that patients above a certain age range with comorbidities are likely to suffer significantly more than ones who did not have those characteristics. How do we need to enable health systems to identify those cohorts within the population? The second thing that we did more on the traditional and BML side and a book out of (murmurs) was to run some forecasting models. How do we provide health systems with the ability to forecast what is likely going to happen with supplies? Fixed supplies, such as PPUs, which were in the news so much, but then also things that are a little bit variable like bed capacity, and could bed– could units in hospitals be turned around to support the predicted surge of patients that are likely to come in. And then last but not least, we introduced a mobile based application to assist people like Tony and help system administrators to be able to have real time information into their data, so they could make on the cuff decisions, and to help identify where patients need to go and do discharge planning more appropriately. So obviously, none of this would have happened if we did not have a strong technology underpinning and a platform to support the scalable movement of data from its source to the consumer that was dependent on this information to make very critical decisions on the fly. We are fortunate in that we were leveraging Databricks’ platform and capabilities very early on to help with distributed computing capabilities, especially around data movement, ETL, model training and both models scoring. And this fit very nicely into Tony’s existing architecture that he had already driven and established within IU Health as they made the move, you know, the monumental move from an on premise system into the Cloud. We continue to push Databricks to its limits. And so far we haven’t had a tip over, at least in a production setting. But I’m sure when that day comes, we have the right support to help us with that. And along the way, we’re also empowering Tony’s team to start to build predictive models that are very specific to their use case. They’re one of the largest employers in the State of Indiana, and a very prestigious institution, not just in Indiana, but in the Midwest and beyond. And they probably have 100’s of ideas around machine learning experiments that need to be put forth, and tested really hard by experts in the field. And our technology underpinning supported by Databricks and Azure really helped make that a reality in the time to value significant. – All right, so Rohan, thank you for taking us through that. I think as we wrap up our presentation today, I know, I’d like to give one takeaway and learning for folks to take out of here with them.
And then I think, Rohan will jump back on here and provide one himself, and then obviously, we can go into some Q&A following that, but I think for other health systems, and quite frankly, I think this would apply across to other industries also, other data teams, other AI and ML teams, I really tried to focus in the areas that I spoke about, on getting data into your expert’s hands. There’s no value in a data warehouse that has steel bars around it and nobody can get into it and get the data out of it and put it to work. So I would encourage the collaboration with analysts and data smart people across your organization, to find pout what do they need.
Focus your attention around servicing those needs. Focus your attention around figuring out how to quickly curate data from every possible source out there and make it available for your end users who truly are the experts in the day to day use of that data and can put it to work and get some real, meaningful and usable information out of it. – Yeah thanks for that great takeaway, Tony. I’d like to give one to all the data scientists that are out there a couple of sort of tips and key takeaways.
My advice to most data scientists in healthcare is, you know, Andrew Yang the very famously said that “Data science 80% of it “is really data and 20% is science.” In reality, in healthcare, it’s a lot more than that. Don’t take for granted the semantic connotations and complexities with the data and Tony said it beautifully, “Find yourself very close to people who understand “the deep context in the data.” Attach yourself to the data analysts on one side, where you really need a deep understanding from a data perspective. And on the other side, you know, most data science projects never see the light of day in a production setting. So you will need to find yourself very closely attached to your engineering peers, your MLOps people as your DevOps people to help take your model that you will build out likely in a static environment and support it through the lifecycle in production and gracefully degraded down. Because you don’t wanna be on the front lines when that call comes in, when a machine learning model goes haywire at 2 a.m. and there’s potentially a patient safety problem. So governance, be humble and pay close attention and respect the complexities and challenges that healthcare data presents itself. There’s not much else on the science side, that’s going to trip you up. So with that, I’d like to end, I’d like to thank everybody for coming and listening. I think we’re gonna open it up for questions now. So, curious to hear some thoughts and questions for Tony and I.
Rohan D’Souza is the Head of Product at KenSci. Rohan is accountable for driving the innovation and product strategy in bringing machine learning solutions to some of the largest and most prestigious health systems in the world, aligned to achieving the quadruple aim of healthcare. KenSci has pioneered a suite of machine learning models that are tuned to indicators that improve operational efficiencies, clinical quality, and financial profitability on a highly scalable, elastic cloud framework. Prior to joining KenSci, Rohan was instrumental in building an industry leading population health solution at eClinicalWorks from the ground up. With over 100M patient’s data flowing through the platform, he helped some of the largest and most successful ACO’s in the US to understand opportunities within their data and build intervention strategies to drive change. He is also a leading voice for the open health data initiative and was responsible for pushing the agenda on EMR systems adopting an open API framework for healthcare interoperability. Rohan graduated from the University at Buffalo with a double degree in Biological Sciences and Business Marketing, focused on the intersection of health policy with biostatistics. His research helped guide the University at Buffalo to become the first public University in the state of New York to go completely tobacco free.
Indiana University Health
Tony is a Vice President within the Information Services organization of IU Health. He has responsibility for Enterprise Business Applications, Data Warehousing & Decision Support, and Health Plans IS. Tony has 25+ years of IT experience across multiple industries.