Jonathan Teich MD is chief medical informatics officer of Elsevier.
Tell me about yourself and about Elsevier.
Elsevier is the world’s biggest producer of scientific and medical information. Traditionally that has come in the form of journals and books, and then ever increasingly over the past 20 years, more about electronic information. First as just electronic representations of those same things, but now more and more as specific electronic delivery of information for a particular need. It’s been very interesting to watch this evolution about how to turn information from these huge amounts that you have to go find into something that’s delivering what there is to you.
I am an emergency doc in one life, still practice at Brigham and Women’s, and an informaticist for the past 20-something years. I helped with a lot of the design and led the clinical systems charge at the Brigham, working for John Glaser over about 12 years, and then went into the industrial side to try and see if I could make an even broader impact.
I spend my time between working with Elsevier in an R&D capacity and a strategic capacity, as well as representing them and the field in government and industry conversations. I’ve also spent a lot of time working with ONC over the past three years as their CDS gopher, and a lot of interesting things have come out of that. It’s a broadly motley career that seems to be working out pretty well.
There’s a lot of information out there in the form of literature and reference material, but clinical decision support never seems to quite realize the promise of actually applying that knowledge in a manner that measurably improves measurably frontline patient care. What are we doing right and what do we need to do better?
I think you’ve really hit the problem. There are places that are doing it very well, places that haven’t quite been able to do it very well, and places that have given it up altogether.
You’re right about information. A company like Elsevier … I’ve been told that we produce seven million distinct pages every year of medical content. Books, journals, whatever else. As I’m going through it, I’m an emergency doctor and I’m seeing a new patient and I have a question — the answer’s in there somewhere. One of those seven million pages has what I want to know.
Clinical decision support has a lot to do with saying, “Where is that information? Can you get me that spot without me doing a lot of work? Can I get that information and then can I make use of that information?” Typically, that’s a wide range of things. People know about alerts, order sets, care plans, and pharmacy information. More and more, how do I deliver the intelligence that I need at a particular point?
It really has been an up and down situation. There were a number of leading institutions through the ‘90s and early part of the last decade that showed that you could do a great deal of change with preventing adverse events, reducing costs. Work we had done at the Brigham with Dave Bates and myself and others showed that we could knock off about 55% of the significant adverse drug events and the corresponding cost savings. We could show we could save a couple of million dollars a year on certain kinds of drug overuse costs and so on. Z
There’s certainly the potential for it, and certainly under some circumstances it works very well. But then as you’ve seen, when it comes to bringing it out into the open and having 6,000 hospitals and all the ambulatory practices use it, many places have been able to use it very well, other places have not.
I think that a lot of this has to do with two things: culture and information delivery. I think the culture, in terms of places where I’ve gone to see what hasn’t been working, have often led to problems with communication, problems with not involving people in the clinical decisions before the decision support goes live, problems with not getting everybody to see what’s about to happen before it happens. I think that’s been probably one of the biggest issues on that side.
On the information delivery side, some of this information is just not in its most usable form. If you try and build this 6,000 times at 6,000 hospitals, sometimes it works well and sometimes it doesn’t. There should be a way to pool everybody together to get the best delivery systems and the best information to be used by everybody.
It’s interesting the Brigham’s BICS rules that Eclipsys bought were very sound, but not widely used. Is the challenge that the underlying data just isn’t there in a way that can connect the rules to the real world? Is there a gap between what rules could do vs. what information is available to allow them work?
I haven’t had anybody ask me a question directly about BICS in a while. That’s good to hear.
The BICS rules were using data that was unique in its time, but I think it’s not unique anymore. I don’t think the problem is that we have insufficient data to get these things done. I’m sure I could construct rules that are making use of obscure data, but a great deal of what I need to know to handle basic quality measures, to handle Meaningful Use, to handle accountable care, and to handle just good practice are things that are are generally available. Most of this comes from medications and laboratory and problems and so on.
I don’t think it’s a matter of data. I do think that there hasn’t really been a good systematic way of showing somebody else at the next hospital what I’ve done at this hospital. I think that maybe some of the things that we did at the Brigham have been ported to other places that use the same IT team. But, it’s very hard to convey this in, say, a research paper and have that go along. I think that if I was going to put a technical finger on it, it’s that it’s been hard to share the techniques very well.
It’s hard to measure success or failure because when a clinician accepts the guidance, they may not enter the order and you don’t have anything documented as to why. On the other hand, then they override 95% of the warnings, you have a record of that and the implication is that warnings in general weren’t really very useful. Can decision support work without allowing clinicians to tailor their desired levels of messaging and without using more patient-specific information, making it less about interruptive warnings and more about guidance?
I think that’s a lot of it. If clinical decision support in a given institution relies on alert after alert after alert, then it’s simply not going to be something that’s accepted. Quality of care is important, but getting your work done in a timely fashion is also important. If you start getting hit with 50 alerts that are taking time out of what you’re supposed to be doing, you’re going to find a way pretty quickly to game that system and get around those.
You need to have a couple of things. There need to be ways to measure what these things are doing. I think you need to be able to understand upfront that this month, this year, we are going to make diabetes care better in our primary care population. You need to be able to be a cheerleader and do the personal side, and tell people, OK, it’s three months later — we’re getting a little better. It’s six months later, we’ve kind of planed out. Nine months later, we’re getting a lot better. I think people respond to knowing that what they’re doing is having an impact.
I also think that you need to get away from kind of doing alerts for everything. In the newer edition of Improving Outcomes with Clinical Decision Support: An Implementer’s Guide that we just published through HIMSS, we’ve said that there are 10 different types of clinical decision support. They include data displays, order sets, analytics, and they also include providing information. A lot of the things that people use that should be clinical decision support is simple information to say, how do I figure out what to do next? How do I figure out where I’m going?
I see a patient in the department. They’ve got a problem I’m not familiar with. What happens? I go off to the computer and I go look up things. I go look them up in MD Consult, or Clinical Key, the new version. I go look it up in other references. You see that all around our department, people are looking up things. But it takes time, and it’s hard to get exactly what you want. A lot of what decision support ought to be doing is giving you the knowledge that I need to get through the next task. I say that clinical decision support is all about telling me what should I do next.
At Elsevier, that’s a lot of what we’ve been doing with the development of two things. One is Clinical Key, which is the complete overhaul of the MD Consult framework. It is designed to try and filter down your questions. It’s based more on the kinds of questions that we know people to have asked in the past and tries to do as best it can in terms of funneling down the information to match up with your question.
We do that, and that’s been lying on top of the framework that we call Smart Content now. Smart Content is our effort to put semantic tagging under almost everything we do. Books are tagged. Journals are tagged to the paragraph level. The order sets are tagged. Care plans are tagged. The idea of that is that I need to be able to jump from one thing to the other, because my first task is going to be, what am I supposed to know? That may lead me to, OK, I’d better order that. That may lead me to, OK, I’d better do a procedure on that. I want to have some way of connecting these things together.
You need to use a variety of different types of decision support for different situations. They need to be really focused and task based.
There’s a philosophic underpinning to how people view decision support. On the one hand, people think, “All those other doctors need to get these warnings, but I don’t, because I’m smarter than they are and I don’t have to worry about it.” But on the other hand, the guy who’s getting the warnings says, “I don’t need them either.” Everybody seems to want the other guy to have constant oversight via clinical decision support. Should we trust clinicians to know when they need help instead of constantly trying to find reasons to warn them?
There’s a balance. You need to have something that’s usable and friendly and acceptable to gain acceptance. Physicians and others are smarter than the baseline, but not quite as smart as we think we are.
The history of something like order sets is that whenever someone implements order sets in a hospital, everyone has this big clamor for personalized order sets. So it’s, “I’m going to do these things differently, so I want mine to look different.” People go along with that at first because that’s what you need to do to build acceptance. Very often, about two years, later the Pharmacy and Therapeutics Committee comes around and says, “We’ve got all these things that were invented two, three years ago and they haven’t been touched and they haven’t been reviewed and they’re using things that are now considered dangerous.” They eventually decide to abandon personalized order sets. That’s one example.
In the area of alerts, should I say that I shouldn’t get a certain alert? I think that it depends on the criticality. I think I should be able to put away certain less-important things or things that I’ve seen repeatedly. I think there should be systems that do smart things like, if I’ve already heard something once on a patient, I probably don’t need to hear it again on that patient during that same admission. I would stop short of saying that I should have a switch that says, “Don’t tell me about this.” But I might have a switch that says, “Don’t tell me about this too frequently.”
The perfect decision support system is order sets. You’re repetitively using things that have been vetted and that keep you from doing anything too crazy. Somebody with enough of those could get rid of a lot of the standardized warnings about doses and drug interactions because everybody is following the same guidelines. Do you think there will be a point where order sets become so prevalent that we can move to the next level of decision support, where instead of saying, “What you did was wrong,” we say, “Here’s what you should be doing that maybe you didn’t think of?”
Order sets are excellent. One reason that order sets are so capable is, well, two reasons really. One is that they are helping you become more efficient at a task you have to do anyway. You have to write orders anyway. CPOE sometimes takes longer than the old way of handwriting. Order sets tend to make it much faster and bring that equation back even or even better. People like order sets because they’re efficient.
The other nice thing about order sets and why they are so acceptable is that you’re usually using them to support a decision and to help you with things before you do them, as opposed to changing a plan. Where decision support tends to be more onerous is where I’ve already made a plan and something comes up and says, “No, you’ve got to change your whole plan” Order sets are timed nicely.
Same thing with nursing care plans, which we don’t hear so much about. Those are timed nicely. They can help you as you’re making the decision. It’s the right timing.
Order sets are strong and I think they can be a lot stronger. Most order sets are giving you the standard ways of doing things. We’ve been looking at order sets and how you can do them better. I think a lot of that resolves around, “Can I fine tune it in certain ways? Can I help you get down to certain nuances, certain situational aspects that take you away from the standard of care?” Because the problem with order sets sometimes that they’re too big in their quest to support everything.
I think that you’re right. Order sets are a great form of CDS, and again care plans on the nursing side. They have all the right user aspects. I think they will become more common. Probably every hospital has some anyway, but I think that they will become more common. The next step is to say, “Can we make these a little bit more data sensitive? Can we make these a little bit more flexible? Can I share them from one place to another?”
Some people would say that what clinicians want is the same tool they would use to make other decisions – a smart search engine to help them find and sift through all this wealth of material that’s out there. If you had a single body of literature like what Elsevier publishes, you could just search the whole thing and have it somehow graded or weighted or personalized in such a way that it would return meaningful data without having to actually do any thinking.
In a sense, that’s exactly what we’re trying to do, and we are.
You’re right on target. People need information. They want to get it with as little effort as possible, which is perfectly human and perfectly reasonable. You need different information when you’re first assessing a patient than when you’re on rounds and when you’re preparing someone for discharge and so on. You want to be able to have smart filters that can give you information that is geared to a problem, geared to a set of circumstances, and geared to where you are in the workflow. Then you need to be able to get smart enough to deliver just that.
Among Elsevier’s book catalog, there’s certainly all the things you’d want in books and among the journals. It’s a combination of things that we publish and things that are published elsewhere, of course. But, really, there’s a hierarchy of what people want to know for different tasks. We are really attempting to do exactly that, which is to focus down on a given task process, a given problem, and try and deliver it, ideally down to the paragraph level, down to the table level. Whatever we can do that’s more focused, that’s quick, the better.
I’ve said in lectures that nowadays, given a choice between good information and quick information, people will take quick information every time. We have to make something that’s both good and quick, because that’s the competition.
It’s like imaging. Everybody says, “It’s not a diagnostic quality imaging,” but they are diagnosing from it anyway. In reality, people will settle for whatever they have available, especially in your field. In the emergency department, you can’t wait for perfection. Maybe asking a system to be perfect is not only not realistic, it isn’t even necessary.
I think that’s true. You obviously want to be correct. What you don’t want to do is put out incorrect or inconsistent information. But you don’t have to put out exhaustive information. Maybe this is the mindset of the emergency physician, where I want to do something that’s good, but something that I can do in the next 15 minutes to an hour.
There is a focused amount of information that I need for anything. I don’t need to know the entire pathophysiology of a given disease to treat it when my question is, “Do I need to do a CT or an MR, or do I need to include angiography?” What I really need is the answer and a reasonable amount of information that can help me justify the answer for the clinical purpose. But when I want to read about exactly the full history of it, let me make a bookmark and let the same system hang it up for me and I can read it when I go home.
Some folks say it just needs to work as well as Amazon, which gives me what everybody else is reading and things I might want to order with a particular product. You’re not reading every factoid in a 20-year-old medical journal. Maybe you say, “Most of what’s in that journal is not important. You just need 2% of it, and we’ll make you smart about that 2%, but then you can go find the rest when you need it.”
We have to explore what new technologies are doing, particularly new social technologies. I don’t necessarily want to have everybody in the country writing into a medical textbook because that has to be carefully curated, has to be carefully checked and triple checked. But there is the possibility, for example, that you could use a social media tool to let people say to each other, “This is by far the best article on diagnosing a pulmonary embolism.”
Imagine residents in particular, who talk all the time and who rely on each other for their training and their information. Imagine if you could put up your catalog of literature, and people wouldn’t necessarily add to it, but they could say, “This is the place to go. This is the place that I like.” Eventually if 4,000 people say that, maybe there’s something to it. That’s the concept we’re looking at. It’s got its ups and downs.
I do a lecture on social media in medicine. Certainly there’s a lot of space on the curve between reliable information and well-shared information. But I think that you can use certain kinds of crowd techniques and social techniques to great advantage in this world, especially when sifting through all the millions of pages.
People are used to the idea of grading evidence, but maybe not grading each piece of literature. It seems that another alternative would be to ask each time that that warning, recommendation, guidance is presented whether that information was useful. If not, then downgrade it so it doesn’t come up as high.
Potentially. I think that you have to look carefully at, is there a difference between what someone wants to see and what someone should see? Usually those things line up, but you have to be careful about being so faithful to that that you miss something important because it’s inconvenient.
There’s also the challenge of how vendors implement the hooks into that information. The clinicians might say, “I’m a nephrologist. I’m tired of seeing serum creatinine warnings,” whereas the data vendor says, “Look, it’s not our fault. We’ve got the data. Talk to your systems vendor who doesn’t use it correctly and tell them to fine tune it in a way that makes sense to you.”
Very much so. As I’ve said a couple of times, the ability to share effective CDS across sites is really important. One of the reasons why we haven’t seen universal acceptance is that there’s too much rework going on, and the rework is inconsistent.
I’ve been working with ONC. I’ve been working with the Advancing CDS project that RAND and Partners did, and on how to make a practical way of taking the various types of CDS interventions and putting them into a form that can be easily shared, and that therefore can be easily integrated.
If I’m Epic and Siemens and Cerner, I may say, “Gee, I really can’t do this right now because I don’t what’s going to win, what going to be the national standard.” But if we can get enough agreement on how these things should look, enough to make a reasonable XML schema that corresponds to certain CDS interventions, then I can get the big vendors to say, “Now we’re confident enough that this is what’s going to happen that we can go and bring this in.”
I think that it’s really important. I think that integration of knowledge and CDS into data and EHRs should be more advanced, and needs to be more advanced if we’re going to fulfill our mission of best care for all the best people.
I had lunch today with a fellow ED doc who’s doing a small project. He’s a child abuse specialist. He’s doing a small project on building a system that allows you to document certain kinds of aspects of a child’s exam and then be able to come back to you with best practices, recommendations, referrals, and so on. He asked me, “Can I get this to work inside all the different vendor systems?” I said, “You know, today that’s a little hard to do because each one’s going to be different and even different implementations of the same system is going to be different.” I suggested that he probably needs to put this out as a service that his practitioners can call on independently. That‘s going to be a way to do things smoothly and a way to do things consistently, but I think if I was an EHR vendor, I’d want to be able to incorporate those.
Any concluding thoughts?
The reason we’re doing electronic health records, in my mind, is that they facilitate the efficiency and the quality of care and the safety of care. CDS has always been an obvious choice of something that can help facilitate that. If you just use the EHRs as data sources, that’s good, but if you can do it and also get recommendations on the right thing to do, that’s even better.
A lot of us, like myself, struggle to know what the right answer is in a given time. Anything we can do to make this more universal, more implementable, more valuable, is going to be utterly good. We really need this. I think we need to see this incorporated more deeply into systems.