The Future of Personalized Health Care: Predictive Analytics



good afternoon everyone so excited host you all for our webinar today I'm Molly I run communications at Rock health we're a full-service seed fund based in San Francisco that funds and supports digital startups on top of our startup support we also host events and conduct industry research including digital funding reports and special topics our latest report and the topic of this webinar is predictive analytics the future of personalized healthcare which was released earlier this week though personalized medicine has yet to deliver on the promise of its powers its precursor predictive analytics has proven effective in many industries and is now tackling healthcare dozens of new digital health products have hit the market and 1.9 billion has flowed into the predictive analytics space since 2011 but what does it take for an algorithm to accurately and reliably impact care to talk about all these things and more I like to introduce the authors of the report rock health managing director Emily Ghandi and our strategy manager Theresa wing and just a heads up please submit your your questions by clicking on the Q&A app and typing it in upvote your favorite questions and we'll answer them during the webinar without further ado here's Malloy and Theresa – take it away hi this is Theresa I'm the strategy manager here at Rock House and so today we wanted to start by taking a moment to remember the excitement about when the human genome has successfully been decoded there was so much hope and optimism that with the ability to sequence the human genome we would be able to personalize medicine to individuals genomic data it's been almost 15 years since dr. Francis Collins had described the promise of personalized medicine as the ability to be able to get the right drug to the right person at the right dose at the right time this is a deterministic view of health care where there's one correct medicine for an individual however we aren't yet at a point where we can say definitively whether or not a drug is the right one based on a person's genomic data and it seems there's been a shift towards this idea that today we can only try to deliver the best care for a patient identifying the best care relies on calculating probabilities to see which treatment well beyond just drugs will result in the most favorable outcomes this idea of personalized healthcare instead of medicine focuses that focuses on not just drugs but also prevention and intervention and leverages a much larger set of data ranging from clinical to claims to patient generated to patient reported data we view predictive analytics as a tool that will allow us to personalize health care moving into the landscape so what exactly does predictive analytics mean we define predictive analytics as the process of learning from historical data in order to make predictions about the future or any unknown and for healthcare this just means health this just means predictive analytics can help identify the best treatment or decision to be made for each individual we know predictive analytics can be applied to a wide range of things but our report focuses on how predictive analytics is directly impacting patient care analytics that are used for clinical decision support readmission prevention adverse event avoidance chronic disease management and patient matching are all like some examples where the technology directly interfaces and impacts the patient use cases of predictive analytics that fall out of scope of the report would be things such as marketing CRM business intelligence tools fraud detection and actuarial modeling well we know payers have been using analytics for rate and premium setting for decades we're not focused on this since it doesn't directly impact the delivery of patient care as data sources and technology advanced algorithms will be able to improve the odds that a certain treatment will result in a favorable outcome for a specific individual we want to use predictive analytics to more reliably predict the unknown but of course that prediction certainty changes with a type of question asked for instance asking a historical questions such as what did I eat today is one that you can answer with the high certainty since the answer is known however as you start asking questions for events that have yet to happen your predictions are in T decreases like asking how much weight will I gain or will I get diabetes it becomes harder and harder to predict the answers to these questions with a high level of confidence predictive analytics isn't new one thing that people often overlooked is that predictive analytics already underlies most of what traditional medicine and healthcare does today whether a physician does it in her head or uses guideline or guidelines laid out in protocols or c.d.s tools there are six steps in predictive analytics first and foremost predictive analytics companies must aggregate training data which often requires companies to learn to handle unstructured non standardized data when you hear people use terms like artificial intelligence machine learning supervised versus unsupervised it's almost entirely rated related to the second step relationship search or having the ability to search for relationships within a large data set these technologies mined the data to identify relationships and develop algorithm algorithms to find new attributes that are a predictor of a certain outcome these first two steps are the foundation and the next up is where we start to personalize health care collecting data points for the specific case in order to input into the predictive model which allows the algorithm to predict an unknown outcome such as a disease diagnosis or stratifying a patient by risk and the fifth step would be for the model to generate a recommendation specific to the case based collect case base collected on the collected data the last step of performance capture is essential to closing the loop so that the model can iterate and improve its algorithms by having a complete data set tracking the relationship between the recommendation once you record the results you have yet another set of training data to run your predictive model on to increase the accuracy and reliability of predictors as you can see this already occurs in healthcare today for acute care cases physicians observe and record a patient's symptoms and based on their internal algorithm from training and personal experience they match the patient's symptoms for a diagnosis and identify the best treatment in order to get the best outcome similarly for chronic and preventive care care providers identify risk factors such as high blood pressure or hypertension in order to stratify patients into a risk category and find the best intervention for that patient so that brings us to the question why now if healthcare has already been using particular predictable it expression care why is there so much excitement and interest around it and our answer would be big data and algorithm production with new technologies to sequence our DNA collect continuous monitoring data and patient reported social media data the amount of healthcare data is expected to grow to 25,000 petabytes in 2020 that would be the equivalent of more than six thousand two hundred and fifty Library of Congress's over the past few years we've also seen new technologies including many open source ones emerge to help process and manage all of this data interest in the promise that predictive analytics can help deliver better care while reducing costs has been increasing since 2011 over 1.9 billion dollars of capital has been raised to fund companies that claim to use predictive analytics already in the first three quarters of 2014 alone funding has increased by almost four and a half times since 2011 we took a look at these companies that receive funding between 2011 and the end of q3 2014 and what we found was that companies are focused on enterprise customers mostly providers these analytics and tools have designed to interface and impact the health care professionals workflow but with their growing list of companies that interact with both the provider and patient however there are a few predictive analytics companies today that are directed towards patients only while the end-user is only slowly starting to shift we see that new data streams including patient generated such as continuous monitoring data from patches and wearables and patient reported such as me logging mood logging and social media are being increasingly used for analytics based on the stated or advertised data sources on each company website we found that an overwhelming majority of 71 percent of companies purport to use clinical data whereas only 42% of companies disclose using claims data this is likely due to the fact that claims data has been mined by companies for decades and those have received funding prior to 2011 so where are we heading with all of this we've observed three categories of predictive analytics that have been used and applied in other technology services which are beginning to appear in healthcare here we show how Netflix has used all three themes of predictive analytics enhance our TV entertainment watching experience the first perhaps most basic form of predictive analytics is creating cohorts based on similarities and providing a recommendation based on those commonalities for example if viewers who like movie a also like movies B C and D then you would you will probably also like B C and D because you like movie a and overtime these recommendations can be improved as Netflix collects more data to see whether or not date they work and try to find more predictive attributes next Netflix added in additional context to help further personalized recommendations by creating individual users within an account historical viewing became better tailored to specific individual with reduced noise from multiple users not only was a model better able to predict what movie would enjoy based on an additional data input such as time of day type of video watched Netflix can also tailor suggestions to your viewing habits and the final type of predictive analytics we view is one that enables the end users to take action house of cards which most of you probably have already seen and loved is a perfect example of what a good data set and predictive analytics can do for you as Netflix continued to collect more data there viewers from there viewers they were able to use this large data set to identify viewer preferences and what would appeal to a large audience instead of the traditional TV model of piloting a show first to gauge interest Netflix was able to purchase the entire first season of house of cards with confidence that it would match the viewing preferences of a large audience Netflix was able to take action based on this data and analytics Netflix is great and all but we're more interested in how we can apply this type of technology to health care in order to personalize care if you've ever gone online entered your symptoms and received a diagnosis you've interacted with Health Care's version of a recommendation engine based on the symptoms and related factors you entered the algorithm is matching you with others who have had similar inputs to your symptoms and what diagnosis were most common this type of analytics can also be used to help triage triage patients and high acuity situations identify hidden and quorum of eight morbidities stratify and identify identify potential high-cost patients and help best match physicians with patients based on their preferences the importance of having context around an individual's historical data is also so crucial especially in healthcare where were you where we are still learning what healthy or normal is hence Google's baseline study textbook textbook guidelines often provide ranges that are integrated into clinical alert systems without any context for each patient for instance data collected from the Cincinnati Children's Hospital Medical Center and Children's Hospital of Philadelphia show that 14 and 38% of heartrate observations and 15 and 30% of respiratory rate observations would have resulted in false alerts based on textbook definitions Lucile Packard Children's Hospital at Stanford has implement implemented these findings around early early warning algorithm systems thus we see it's key for these predictive models to be able to adapt their warnings to each individual other use cases where context will matter a lot in predictive analytics include decompensation preventing readmissions and behavior change people always say there's no silver bullet for behavior change and the reason why is that context has been lacking so many attempted interventions as we think about how to best use analytics to result in taking action genetic screening is the case example work that allows parents to plan ahead for example a couple planning to have a family can send their DNA samples to Council to understand the probabilistic risk of their child inherit inheriting a certain disease use cases expand beyond just family planning with large smart datasets you can run analytics to take action before you even know you need such a disease prevention population health management early intervention and even treatment selection predictive analytics can be used to improve the certainty of a prediction personalized care can emerge from high confidence algorithms that can predict actionable interventions that improve long-term health outcomes how we're building models that are able to reduce uncertainty and personalized care is not an easy task the requirement many factors such as the ability incorporate new data types and sources to have reliable consistent predictive models timely data transparency around the prediction and convenient convenience and contextual recommendation and a close feedback loop that allows the model to rapidly learn and now I'll hand it off to malay gandhi our managing director thanks Theresa I'm going to cover some of the challenges of the space and some considerations as we look at moving forward but before we get to that and wanted to remind everyone that they can put their questions and into the right my colleague Molly's tracking them all we'll cover cover some of them here towards the end of the presentation so Theresa did a good job of setting up this beautiful world in which you know health care is deeply impacted by predictive analytics and it's just as easy as Netflix right well when we went and looked at the companies it's actually non-trivial and very very challenging and there's one core challenge that actually comes up time and time again that's not represented on this page and you'll see the threads of it but it's really related to adoption and workflow and we're gonna come back to that I promise towards the end of the the presentation and talk about the real problem which is how do we get people to use predictive analytics if we believe it kind of have such a large impact within within healthcare but as we looked at it we profiled a handful of companies here that are really showing differentiation across each of these key stats and what you see is a big difference between sort of a basic company and those that are getting pretty advanced within each of these steps now many of them use everything here that's presented around predictive analytics but they're really differentiating themselves in some core areas so as we look across them you know particularly training data aggregation what we've seen in the past is people are only accessing very siloed data sets the data is already structured for them and in the advanced model the data sets are much more disparate and they're using unstructured data as a release to relationship search it's the idea of using novel data so how do we do something that hasn't been done before versus the old traditional data the data collection I'm sure you can understand with continuous monitoring devices we're looking for real time versus lagged when we're characterizing cases people want to know how they're characterized how these algorithms work instead of being instead of looking at a black box and when we're providing recommendations they need to be personalized to the individual versus generic so sort of bottoms up instead of top-down and then finally when we're looking at outcomes how do you actually close the loop and prove what you're doing is truly working versus kind of laying out a recommendation and walking away from it so let's go through the cases with Flatiron you know the company was deeply profiled in Forbes magazine and this is one of the key learnings that we had from it and we also spoke with with one of their customers to better understand their model but ultimately when you look at oncology data you need to aggregate so many different disparate pieces of information not only from the electronic health record system of which one they they purchased and acquired as part of their last financing but also laboratory and billing data as well and and even clinical trial data so taking an investment from lab core was of great benefit to them ultimately because they're able to have access to all these different data types but it's not just about access it's also about cleansing it's like the willingness to go in and look at unstructured notes from the electronic health record and structuring them so that the data becomes usable in the old world you would just avoid data like that it's not clean it's not structured it's not labeled you know and you would avoid it instead you're working with things like claims that are you know pre designated fields and they're built built around structured data models the frequency also matters quite a bit as well you know in healthcare people many people are used to doing monthly loads or you know bimonthly loads and really with flatiron they're loading the data and nightly so a clinician coming in and looking for point of care information on a patient literally has yesterday's data at their fingertips this is a this is a big challenge obviously again clinical workflow comes up but managing and processing these new types of unstructured data is a non-trivial problem and it seems that one that flatiron is is tackling going into the relations relationship search as Teresa identified when you hare sort of the buzzwords artificial intelligence and machine learning unsupervised and supervised learning we're really looking at this category so you have this tremendous amount of data and now you're saying well what's really related to other types of data so but the goal here really needs to be to identify these new data sources that act as better predictors but it's not solely important to find the data it's who can find the data the cheapest you don't want to create something very expensive in terms of a data collection process you want to have the best prediction at the lowest cost so how do you conveniently acquire the data that has the most predictive reliability and so we looked at our own company well-framed to see how they were doing it and what we found was the engagement modality is so much different when you're able to send a patient home with a mobile app so every day you know over two times a day they're checking in and providing new data that's giving you a great prediction of whether the person is likely to be readmitted following a cardiac incident they're acquiring that data you know basically in real time and then the algorithm is prioritizing who's at the most risk in delivering that right to you to a care coordinator within a dashboard where they can manage hundreds of patients this isn't again trivial whatsoever identifying these novel data points that can better predict outcomes you know that's the first step but again the second is how do you do that really cheaply and expensively the other thing that's come up quite a bit is you know once you start searching for relationships and data you can own let's correlate anything with anything so ensuring that you don't have spurious relationships is also important and then finally you know we've talked about workflow time and time again but really how do you get this into the hands of people who can use it and intervene on on patience so with relationship search what we're really looking for are great predictors that you can acquire cheaply and well frames a good example of that in terms of data collection this is the point where you say okay I have the algorithm now I know the predictors but how do I collect the data I need for my particular patient and in the old world it was very much based on what you could collect at points in time typically when somebody checked in for for care came into the office or it was lagged in claims data today again with mobile apps with continuous monitoring it's much easier to collect real-time data and a continuous basis and for Wildflower weight is a key predictor so simply having the patient self-report makes a really big difference they're able to catch high-risk pregnancies in the first trimester which is a really hard problem for health insurance companies today that are dealing with with claims data as everyone has heard I'm sure target knows that people are pregnant long before your health insurance company does there's some core infrastructure issues here for sure electronic health record systems payer administration systems none of them are really designed to collect real-time data they're not accustomed to the idea of millions of data points coming in you know per year instead of just a handful on each patient so there are some things that you have to build from an infrastructure standpoint really work within the healthcare system and honestly as Theresa pointed out you know in the contacts slide earlier most of the healthcare system doesn't even understand what's normal and what isn't until they start looking at continuous data so even the care system itself doesn't know how to handle this many data points on this board that really thinking about now I need to match my patient to this particular algorithm so how do I dump them into a cohort that says they look most like this other person so thus I want to lump them in and then make some type of recommendation about what to do them and that this is really what's going on with clinical decision support and today you have such a limited set of protocols and guidelines and even the guidelines that we do you have you don't even understand how much evidence there is behind them a recent review of clinical guidelines for you know cardiology which is a pretty well studied area show that only 19 percent of the guidelines were supported by our CTS so we really need just greater breadth in terms of what's the current scientific research what's the current clinical practice data telling us when we're employing CVS tools and then being able to adjust those guidelines and on a more continuous real-time basis guidelines traditionally a traditionally lag actual scientific research by many many years and then the final area and what we would say we leotta has done a pretty good job with is providing some transparency into how these knowledge graphs algorithms etc are built at the end of the day a clinician really wants to know why am i making the recommendation that i'm making they're accustomed to reading a paper incorporating that evidence into the practice of medicine but when you just provide them an algorithm with the answer there's some challenges that come come up with adoption and so visualization makes a big difference here and you know in the recency of the the data makes a big difference here for our next case on contextualization we're really lucky to have christine Lemke from the activity exchanges actually here with us in the room so let me let her explain what's going on over there thanks Malayan Theresa the activity exchange is really targeting or really targeting the problem of there's thousands and thousands of what we call digital interventions in the market whether they be mobile apps or clinical wearables or even fitness devices that people are using more and more to track and maintain some of their health behaviors and so when you have data coming from all of these disparate devices when you have that combined with medical data claims data demographic data that providers have how can you create use predictive analytics to create a very personalized experience for consumers that will motivate them to change their behavior and sustain that behavior in general and so what that requires on the infrastructure side is a system that can collect all this data from all these different api's or SDKs or different modes of these providers sending data have a consumer authentication system for consumers to get permission to access the data and then lunge all of these disparate things together in a way that makes sense so that you can run predictive algorithms on top to really create a very highly tailored program for a large population of people so it's really taking the many to one or the one-to-many approach where each person has kind of the right prescription in a way self-directed prescription for behavior change that's sustainable in the long term thanks Christine we're so lucky to have you here what has today to walk us through the activity exchange so in the last step here right so think about this long process we've gone through we aggregated all this really interesting data we had to cleanse it and find it in every unnatural place put it together we search for relationships using a variety of technologies that Teresa outlined before open source tools to mine it find this relationship and then somebody presented themselves we collected data about them they put him into a cohort we characterize them and we make this really great recommendation for them but so what now we need to know what happens was your recommendation good or bad was it right or wrong and what was the outcome of it and all of that has to go right back to the beginning to go back into your algorithms to say did we characterize the person correctly did we put them in the right cohort was this the right type of intervention for them so a company you know like Google Christine pointed out to us when they do a search engine result you type in something and they know within one second whether they did well or not because you could either click the first link or the second link or the third link they know how well they're doing predicting what it is that you want and in health care it's just not so easy you know the outcomes take some period of time to capture and so you don't know whether what you're doing is actually working or not a lot of the things we're trying to do in health care really about the long term especially when we're talking about prevention or chronic disease management so to do this you know and to actually be able to record your outcomes to control the full chain here you really need to be deeply embedded within healthcare organizations so we looked at this company you know health catalyst and once you build this type of closed-loop system where you're rapidly iterating and testing and the activity exchange has sort of done it as well and you're able to look at your outcomes that's really where the long-term defensibility comes from and it's literally the most important thing we look at for analytics companies is what is your position in the value chain to collect your own outcomes to quote unquote close the loop and really understand what you're doing is working so if you make a diagnostic a treatment recommendation you know whether that was right or wrong so you can further chain your algorithms with the goal obviously to be delivering many of these things through software over time and not involving human beings or involving lower-level you know folks who are less trained so how do you get there well first you need access like I said you need to span the spectrum you need to be able to test you need to be you know iterating rapidly there's a large push and I think it's a very good one to be able to randomize patients at the point of care there's so many things we don't know whether they're working or not working we don't know I prescribed one statin over another or one anti asthmatic over another so why not randomize patients at the pointed care to see whether our our interventions are working again this is also another thing the activity exchanged as well and then the timing how do you really measure performance on a little bit you know faster basis versus running you know long-term studies doing retrospective observational reviews this is really really hard I don't want to trivialize how challenging it is to actually understand whether the recommendations you're making are truly working within healthcare and clinical workflow integration is actually deeply meaningful here as I said to randomize to test but also to have a place to deploy you know your algorithms so those are some of the cases that we looked at let's wrap up and talk about some of the consider stations going forward here there's a huge opportunity for predictive analytics and this is ultimately what matters the most to us if we can get to this point where we have this personalized care being delivered delivered by clinicians but supported through algorithms there's a tremendous opportunity to take costs out of the healthcare system when we looked at it probably about 350 billion out there just by reducing treatment variation improving our care delivery models and fixing care coordination so let's take them one by one over treatment there's so much care that's delivered today that's outmoded it's driven by who the specialist is and frankly justice news science altogether we need to start restricting treatment and interventions to the patients who will actually benefit from them Theresa said it at the beginning that was sort of the goal of personalized medicine right drug right dose right person health care is so much more than drugs today we have many many millions of patients where most of the cost lives and chronic disease management and lots of companies frankly selling crap claiming that's how you manage chronic conditions we really need to start looking at what things are working and using predictive analytics to best match the interventions to the patient's failures of care delivery so we're not continuously studying what works for whom and in what context we're not scaling the best practices for bread for preventive care Theresa talked about the cases of early warning systems I mean there's so many people that were we're missing because we're not doing the things that we already know that that work so it's not just about over treatment but there's things that we do know that work and we need to implement and scale those and having more control at the point of care will allow for that and then finally you know this lack of care coordination and how do we really stratify outpatients readmissions being sort of a classic example of this understand their risk and intervene early on them and make sure people are cared for between visits and following hospitalizations and this entirely comes down to a predictive model we have limited resources we have to stratify patients and really focus on the ones that are at the highest highest risk to ensure that they don't end up back in the hospital or they don't fall and fall in between the cracks between visits ultimately the industry matters here a lot we need payers providers pharmaceutical companies to help implement these use cases for for predictive analytics and I think they can really go after some of these big areas and capture some of the value from them so with payers what we're interested in is the idea of much more personalized medical policies how do you decide what is and isn't covered instead payers today have global medical policies deciding you can have this and you can't have that well why shouldn't that really be personalized to actual individual members based on the context of their of their care and their health status as well as the benefits how costs are shared by parties reference based pricing is one of the first areas where you can see the payer is making a decision about benefits based on what they believe is best for you so if you want to go to a high-cost provider you are going to pay for out-of-pocket and the same should be true of things that we know don't work in medicine if you it's we run it through an algorithm and you're not a good match for it then maybe you should bear the cost for it if in fact you want to pursue that type of treatment every payers got wellness chronic disease management programs that aren't working today we got to start matching the interventions to individuals as exactly as Christine had identified earlier that the activity exchange is doing if we want to have any chance of scaling behavior change programs a one-size-fits-all the quote/unquote silver bullet for behavior change doesn't exist we really have to start personalizing these programs for providers you know it's important to provide access for the doctors the so-called green button initiative that's you know some folks at Stanford are relief working on to provide point of care access to historical data many situations in medicine today it's ambiguous what the treatment should be what is the diagnosis what is the treatment and doctors should have access to you information on patients that look just like your patients anonymized private sure but you should really understand what's going on and what's been done historically to try to identify the best course of treatment for your individual patient and that should be available at the point of care the so-called green button reducing treatment variation we talked so much about that and providers today have the same issues of as payers as it relates to managing populations with biopharm er there's a great opportunity for increased productivity within R&D to better match patients to the treatments that almost benefit them thinking about services that wrap around the pill and there's also studying that needs to happen in a post market context how are people responding to already approved drugs whether that's for safety issues or for narrowing and a use case to targeted populations the last couple areas I want to cover are just related to regulation which we see is one of the key constraints for this area and the second on adoption so first with regulation FDA has sort of punted for the time being right now we were expecting a little bit of work to be done as part of the middle mobile medical applications guidance but the guidance specifically does not address CTS there's a lot of pressure from Congress right now for the FDA to pass the you know baton I'm regulating the space to another agency and I don't know what's going to happen we know that the protect act the software Act those aren't going to pass and FDA won't be stripped of its regulates regulatory authority but they will be providing guidance on c.d.s soon covering everything from the alerts reminders warnings that you know Teresa talked about companies that are doing computer aided diagnosis any type of treatment regulation treatment recommendation excuse me and then ultimately you know it's going to be agnostic to whether where that data comes from it can be patient generated it can be manually inputted it can be collected at point of care they're going to be looking at everything that's using software to influence clinical decision making there's a couple frameworks that have been put out there most of them are advocating for a risk-based approach which is already the one the FDA takes today you know and we'll see what ends up happening in terms of interagency collaboration because many of these areas are going to fall under oversight from multiple agencies so you could have to be not only dealing with the FDA but the ONC or FCC as well depending on if you're using a mobile you know wireless device our big question honestly as we looked at this space the FDA's mandate has been to protect human health and safety that's totally understandable and they look at these algorithms I think there's been you know a couple hundred deaths that were reported um over the last ever the last a couple decades as it relates to health IT and it's really important for them to protect human health and safety but there's almost an inversion happening today we were speaking with a company that recently received its an FDA clearance for a software based diagnostic essentially it was diagnosing a medical condition using only a software algorithm and the FDA cleared them and in studies they had conducted they only had a three percent error rate so basically they took a thousand reads they had specialists look at the a thousand reads and they had a sixteen percent error rate while the algorithm only had a three percent error rate so when are we gonna reach the point where the FDA starts regulating medicine because it's actually less accurate than software and where physicians clinicians and healthcare professionals at large are required to use software to support their decision making processes because it turns out the software is more accurate than them and I think that's going to be an entirely new space and role for the FDA where they have to start regulating people and not technologies beyond regulation the most obvious and largest constraint on the use of predictive analytics in healthcare today is clearly with with adoption the power dynamics here are so dramatic when a patient is empowered to input their data into a predictive analytics engine that could return a better answer than a health care professional like the one that you can in the case that I was just sharing which you can buy this device you know over the counter you could make a more accurate diagnosis for yourself by providing the data to this to this device and it's going to report back to you what your your diagnosis is it really changes the power dynamics within healthcare and so the first and most obvious pathway which is where clinical decision support lives today is you know the patient provides the data to the doctor you report your history and your symptoms the doctor can put that into a predictive analytics engine it reports it back to them and then they give you your you know their their treatment recommendation but there's also going to be another pathway if we don't have you know physicians readily adopting hospitals healthcare professionals readily adopting these predictive analytics because they don't want to lose that decision-making power they don't understand the algorithms or they think they're going to be liable we're going to see patients going directly into the engines once the data is out there the cat is essentially these predictive analytics algorithms will exist whether it's open source or commercially and that we provided directly to patients they're gonna have their own adoption challenges there as well though I mean how convenient is it to access an algorithm if you have to go out and find your own open source tools versus being provided a mobile app there's a big big difference I think consumers will have you know trust issues as well what's the accuracy what's the reliability of the recommendations coming out of some random predictive analytics engine what we know and as Theresa said is that people love diagnosing themselves there is no more common activity on the Internet besides Facebook that diagnosing yourself you love typing in symptoms and learning more about like what's potentially wrong with you and honestly these algorithms are only going to improve and you're gonna see an emergence of WebMD 2.0 companies that utilize real data from patients and really personalize the diagnosis that they're able to offer to patients but I do think as that becomes more and more real and less and less of sort of a joke that anytime you input a headache that you have the possibility of you know a brain tumor they become more and more serious patients are gonna start to flip and really wonder about the accuracy and reliability of the recommendations there's always a privacy issue and there's going to be a clear regulatory burden of for any software company that's trying to provide a diagnosis or treatment recommendation they will be they will absolutely be regulated so one of the big things you know where we're interested in and maybe it's just me personally is as these systems come up and they're making recommendations the place that Theresa really started the conversation here is that medicine is shifting towards an entirely probabilistic model you're going to input all this data and what's going to come out of it is weighted treatment recommendations and how does a physician or a patient and decide which of those recommendations to take how they're presented if you think about design today design is about essentially manipulation it's how to get a user to do what you want so the presentation of this information to individuals to doctors and others matters a lot and we wonder when design itself the actual user experience what the screen flows look like how recommendations are presented our probabilities shown how transparent are the algorithms when that's going to end up being regulated could it influence the decision-making so deeply in a way that's almost subconscious which is the goal of design that the FDA starts to look at that area as well so I'll leave you with that I do want to take one minute and thank our our corporate partners who looked at an early draft of this report almost two months ago gave us invaluable feedback and beyond them so many people more than I can even name today who've helped us along the way to really understand you know this this space so with that I think we'll take any questions that might be out there Molly my colleague has been tracking a few that are that are highly voted so we'll go ahead and take those so Daniel asks why is transparency and prediction so important is this related to levels of adoption well you know I think it absolutely is what you can't see you can't trust and what was pointed out to us I'm just gonna lean on Christine here who's helped us understand this space quite a bit is that transparency can actually be bad to you there's a point where it's actually too complex to be transparent we're dealing with something with very very large datasets and to try to unravel exactly what led to any recommendation being made we'd be entirely too complex and so some of the transparency and I put that in quotes really needs to be simplified because it's showing people everything could result in some problems as well Adrian asks are any companies you're looking at pulling social indicator social data into their models so things like income employment status drug abuse etc I don't know about drug abuse but we are seeing you know usage of social and demographic data I mean it's interesting once you start kind of colliding the traditional healthcare data sets let's say claims and you know kind of electronic health record or clinical data with just the basic demographic data that marketers and others have been you know using we put that under the category of patient-reported about 26% of companies said they were using that we heard great stories from the activity exchanged christine and her team come from one of a financial marketing and analytics backgrounds and they know the value of that data we heard it when we spoke with the folks at catalytics too that it turned out the predictive power of some basic you know demographic indicators and by base I mean things like zip code are better at predicting readmissions than every clinical model that had been derived derived and described in clinical studies let's see here what else do we have Daniel asks how knowledgeable our providers and health systems about predictive analytics seems like at this point consumer education may be a large adoption barrier well I mean I think there's a lot of education to be done for everybody and there's a lot of you know things that need to be that needs to be done to help educate everyone in the system about their value but also there's a lot of noise in the system as well it's not as if clinicians and others don't understand the power of predictive analytics but what I would say is the largest adoption barrier today is really about getting it into workflow so how do we make it part of people's jobs to employ more predictive analytics and how do we express the value of it at the at the end of the day we know what influences physician satisfaction the most and that's providing high quality care there will be no problems with adoption if there's strong belief that these predictive models will improve the care they can provide to patients and Melissa asks a very deep question who's expected to be responsible for teaching and informing health systems about predictive analytics I don't know I have no I have no idea I hope there's many health systems reading this report and listening to to the webinar what else do we have here this is a good question so how could a buyer substantiate you know the purchasing decision without if you couldn't provide closed feedback loops so what I would say is building on open source clinical models is a good start so there's a bunch of things that people are studying Teresa gave an example of that with the early warning systems that somebody's already demonstrated the outcome for so there's a lot of things before we get into all these new and novel data sets it goes back to this failure of care delivery there's a bunch of things that we know works or targeted populations and so how do we implement predictive analytics gravity's predictive data attributes from patients today and implement them because they've already been published in research to work so it's not always about closing the loop having access to to everything because there are a few things not even a few many many things that we can look at and and implement based on pre-existing research we like those models actually as investors we think about it as exploiting open source clinical protocols for for commercial gain let's see what else what else do we have this is a good one Vivian is asking how much can we rely on patients in putting their own metrics as a reliable consistent source of data some studies have shown that patient use of health apps drops off after a certain point in time so I'm gonna take the first part of that and I'm gonna hand it over to you Kristine who's gonna tell us a little bit more about what the decay rate looks like with usage of health apps I think when we look at sort of the patient side of the the data equation claims have their own set of inherent biases that you have to correct for so do electronic health records but if you think about patient generated and patient reported data for a medical device as a continuous monitoring we can generally trust it and with patient reported you're going to have to deal with biases and I think as the companies become larger they're they'll learn how to correct for them but a good example and you see it in Yelp right like people are more apt to complain about something so if you take patient reporting on drug side effects they tend to go go negative people don't report the drug makes me feel good I feel great it's working for me they tend to report when it's not working so that's like a very simple example but it's how it's like how we need to understand that patient-reported is also is also biased but let me let me see if I can hand it to Christine really quick and she could talk a little bit about consumer engagement with apps and devices over time you know consumer engagement with apps and devices especially in the what I'll call the fun wearables sector which are things like um fitbit's connected pedometers things like this RunKeeper etcetera are actually kind of dismal most people try them and drop off and then they try another one a few months later and then drop off of that so the cycle times on those things are fairly short there are strategies that companies can use like incentives for example that sustain some of that behavior where you can sustain their usage of Fitbit or RunKeeper things like that by using smart incentives and by using sort of smarts encouragement in what are some of our clients call key moments of influence generally speaking though you want to you want to employ a slightly different strategy for patient self-reported data you want a lot of passively monitored data and you want devices that people are going to naturally use for other things and health data emitted from those things are sort of a side-effect so the Apple watch might be a good example of this okay I have a question for you so Joanna Hass we mentioned the challenge of finding spurious relationships within within the data and you and I have actually talked about that a lot activity exchangers aggregating lots of data how do you deal with that challenge yeah it's really difficult I mean obviously there are statistical tools to help identify spurious correlations versus things that are a little bit more concrete and so definitely work with a provider who understands statistics really well but other than that you need you need variables outside of your controls outside of your experiment group to validate some of the findings that you're discovering and so from our perspective we set up you know scientifically valid experiments on populations to figure out whether the effect that we're seeing is something completely spurious or something that is actually cause and effect or related so you have to set these things up like true scientific experiments if you want answers to spurious correlations thanks Christine let's try to take in a couple more before we hit at the top of the hour here they asked you see outcomes as an output of predictive analytics or as predictive analytics the result of an outcome such a good question and so meta that's why this loop is sort of closed what you need is outcomes to feed back into the loop to improve the accuracy of your of your algorithm so not to believer this but this idea that you can feed the data back in and the faster you can see the back end and improve your algorithms whether that's for matching people into a cohort or making recommendations you know for for that particular individual like contextualizing the recommendations it's really important to feed the data back in so Daniel is asking apart from venture-backed startups you're the other major industry players now and in the near future EHR is Google misfit or or others so this is a really you know good question we tend to focus on the early venture backed companies as part of our research new tracking all the funding which which treats does a great job with um when when you look at this space even if you just go back to the top staff of of predictive analytics who controls the data at the end of the day becomes a really really important question if you want to become you know a big predictive analytics player gaining access to the data so electronic health record companies are certainly in a good good position to do that they're collecting a lot of data and I guess you know what we're deeply interested in is how how predictive um you know what importance will we place on patient generated patient reported molecular data that patients have access to you in future predictive models because then you can see control of data loosening up we don't know who own all of this data in the future there's many companies who are seeking to aggregate it today but none are really at scale so you have to Google so yeah baby maybe it's Apple through its healthkit platform but an individual patient is really provisioning all the access to that data through through Apple's platform so maybe the the patient or consumer is really you know sort of at the center of it we do see again going back to EHRs they do have a lot of clinical data but we think being at the point of care matters quite a bit too so one of the areas other areas we'd like to fund are people who can develop amazing tools for doctors that go around electronic health records to be honest with you because we don't see them innovating fast enough to deploy these types of point of care tools in varun asks because you can you speak a bit on how you see medical responsibility and legal strap how and how legal structure may need to change for predictive analytics outcomes um you know I love this kind of statement here clearly over time it'll be irresponsible to use current blanket methods over analytics you know just like self-driving cars it's so fascinating it's apt that you use the analogy to self-driving cars because you know we were we were literally discussing this in our office the other day which is we know we're gonna reach the point where the the number of data inputs is so so wide it would be impossible for a human being to process them and thus the algorithm will actually be better at doing it it doesn't mean we replace health care professionals it means we simply augment them with technology as we have in basically every other sector and industry but as with self-driving cars the day that their first accident occurs everyone will demand all self-driving cars come off the road even though there's millions of car accidents each day and we see the same challenges occurring with any type of health care predictive analytics so it's all great until somebody is harmed by one of them and we already know there's been deaths from you know how resulting from health IT software but they are miniscule in comparison to the number of medical errors happening each day so honestly I don't know where is some kind of tipping point is but we are deeply interested almost at a philosophical level of how you start regulating the medical professionals instead of this software when it's very clear that the software is more accurate than than any human being even a very very well well-trained one well looks like we you know got to most of your questions today so I think we're gonna go ahead and end the webinar I want to really thank you know our whole team Holly and Molly for getting everyone here to this webinar and my co-authors on the report Theresa Wang and Lauren DuBose and thanks to Christine also who happened to be here today I was gracious enough to spend the time with us and answer your questions feel free to reach out to research at ROC health dot org if you have any questions we hope you enjoyed the webinar and have a happy Halloween

1 comment

  1. the chick host sucks terribly, her voice makes it really hard to pay attention, she's stuttering and unfocused get rid of her

Leave a Reply

(*) Required, Your email will not be published