May 2010

Radiology Today Interview: David S. Mendelson, MD — VR, Workflow, and EMRs

Radiology Today
Vol. 11 No. 5 P. 12
In addition to his duties in pulmonary radiology, David S. Mendelson, MD, FACR, is chief of clinical informatics and director of radiology/information systems at Mount Sinai Medical Center in New York City. In an interview with Radiology Today (RT), he discusses Mount Sinai’s implementation of voice recognition-based reporting, as well as how such reporting fits into radiology workflow and digital information sharing.

RT: Tell us a little about Mount Sinai and the informatics situation there.
Mendelson: Mount Sinai is an 1,100-bed academic institution in New York. We certainly take care of patients, but there is also a medical school here as well. It’s also a quickly rising research institution in terms of NIH [National Institutes of Health] funding and grants. Our radiology department does about 360,000 exams a year. The main campus is Mount Sinai in Manhattan and there is a smaller hospital, Mount Sinai Queens. The sites have an integrated radiology department. We have GE Centricity PACS and Centricity RIS deployed at both sites, which handles both our inpatient and ambulatory imaging. We implemented the Commissure RadWhere reporting product in the spring of 2006.

RT: How did that come about?

Mendelson: We purchased and implemented PACS in 2001 but at that time, voice recognition transcription solutions were having a lot of difficulty in the real world. So we decided to wait until the technology was a little more mature. At RSNA 2006, my department chair walked around and saw that there were more options out there. He was also hearing success stories and decided it was time to explore.
As part of our PACS installation, GE had agreed to interface with any major voice recognition vendor who was willing to interface with us. Years after the original installation, we ended up selecting Commissure. We were going to be the first site to integrate Commissure with GE PACS, which we did over a two- or three-month period beginning that spring of 2007. We did a pilot over that summer that involved six or seven radiologists. Some of that pilot was working out integration between the two products. Some of it was just seeing how our radiologists responded to it, which was pretty well.

We decided the pilot program was a success and that we were ready to go ahead and finalize our selection of Commissure RadWhere as our voice recognition transcription product. GE upgraded our workstations to handle technology as part of the original PACS agreement. It was January of 2007 when we took the rest of the department live, which added another 65 to 70 radiologists (including residents) in a more massive deployment over 30 days. That went smoothly, all told. Nobody was banging on the chairman’s door saying they refused to use it. Some people responded very positively. Most people were neutral, shrugged their shoulders, and said, “OK, if that’s what you want me to do, that’s what I’m going to do.” There was a small minority, maybe about 10%, who were unhappy but didn’t fight too hard and required a little more hand-holding. Over that 30-day time frame, we basically became a 95%-plus voice recognition RadWhere dictation site.

The last 5% were a variety of other things, including mammography, a couple radiologists with heavy foreign accents that we decided to go slowly with, and one diehard who resisted going to voice recognition as much as he could—even though he could have been a poster boy in terms of his recognition rate. He just had a philosophical difference, but several months later when we said we wanted to end all transcription services, he was given no choice. And we have been a 100% voice recognition department since that time.

RT: As these transitions go, that sounds pretty good. The classic argument you hear against voice recognition is that the radiologists put in most of the work and effort, but they’re not the ones who see the benefit. How did you approach that? What were the keys to making it work for your radiologists?

Mendelson: One, it was discussed in advance, publicly, within the department that we were going to do this. Two, the chairman stood behind it. In fact, he was the one encouraging it. So, politically, it was set up correctly. Three, I think we had a very reasonable implementation approach to people. When we were ready to go to that person and implement, we would say to them on day one, “First, we’re going to train you.” We would have a trainer, which was an in-house person from our radiology IT staff, sit there and go through the basic operation of what we believe was a good product and a fairly intuitive product. We’d say, “We’re not teaching you to use all the bells and whistles; we’re teaching you to generate a basic report and use some of the template functionality.” The trainer would sit with the radiologist for a couple of hours, then they would say, “OK, I think you’ve mastered this; spend the rest of the morning trying to use it as much as you can. But if you’re becoming frustrated, your digital dictation is still available. You can switch to that, no grudge held.”

And then the next day, the same trainer would visit physicians and ask whether they had any questions or problems and address those and say, “OK, we’d like you to use it again, voluntarily, more today than you did yesterday but, if you’re frustrated, you can switch back to the digital dictation.”

On day three the IT staff would offer the same speeches in the morning, but what basically happened was that people were comfortable enough that they did their entire day’s work on the voice recognition product—and that was their separation point from digital dictation. That may sound like a fairy tale, but it worked pretty well for the large majority of our users. The proof was in the pudding; we had 95% of radiologists using it routinely at the end of the month—not switching back and forth.

We then filled in the gaps. We put in stand-alone dictation stations using RadWhere for mammography; it was not integrated on the mammography digital workstations. That’s not a basic PACS product; we didn’t have the integration accomplished there and we still dictate that way. We have a stand-alone PC with RadWhere running next to each mammo station. What you lose is full integration. You would log into PACS and you would have to separately log in to RadWhere on the same workstation. Once you had logged in to both, they share context of the patient. And we put that into Centricity when you’re ready to dictate, you click one button and it moves the patient name and ascension number into RadWhere. All things told, it was a fairly gentle go-live.

RT: What did you learn that would be good to share with other people who might be going down this path?

Mendelson: We learned that you should not give up on people with accents—for reasons that may elude us a little bit. Some accents do quite well with voice recognition; others are more problematic. Unfortunately, I can’t give a clear philosophy of which ones. I think that if you have somebody with an accent, don’t have a preconceived notion of the outcome. Some folks with accents do quite well.
The other thing we learned is that the particular product we were using had a lot of tools in it that really help to increase the accuracy. And so, without endorsing a specific product, I would tell somebody shopping for voice recognition to look beyond the speech engine itself to whether or not it has tools that when the voice recognition consistently makes an error, the product has tools that let the user substitute words. For instance, if I say, “Mount Sinai,” and it prints the word, “Montana,” for instance, there is a tool there that can make it just print, “Mount Sinai,” even when it’s misinterpreting the words you’re saying.

RT: Like autocorrection in a word processor?

Mendelson: Yes, it is an autocorrection, but that is on top of the native autocorrection built into the speech engine.

RT: What was the biggest challenge in getting the system to work at Mount Sinai?
Mendelson: We needed to interface this to our RIS. To get the most out of voice recognition, it needs a bidirectional interface as long as you’re using your product separate from the RIS. Some companies sell both RIS and voice recognition and may have a very direct integration. If you’re doing a best-of-breed installation and you pick a voice product that is separate from your RIS, you need to understand the field mapping and make sure it’s correct so that there are no errors. Then you can really leverage the information that’s available in each.

RT: How did the change impact workflow initially and then after everyone became familiar with the system?

Mendelson: Interestingly, after the three days, it was really, for most people, a time-neutral experience. It takes a little longer at the point of care when you are dictating your report and you sign it right then and there. That is what we encourage because we wanted to leverage the rapid time turnaround. If you dictate and sign right away, it really shortens the length of time between when an exam is done and when a report becomes available to a clinician. That is our goal, so we encourage signing right away. But if you sign right away, you must edit your report right away and that lengthens your dictation process.

Signing when you complete and edit a report saves the radiologist time on the back end of the process compared to people who are back-signing their reports at the end of the day. That is where radiologists make up time; they no longer have to do that end-of-the-day mass signing of reports. My experience is that the entire process is ultimately time neutral for most people. For some people, it’s a little slower. There are some people who are a little quicker.

The big gain that we appreciated early on was the clinicians coming down congratulating you that your reports are available before they ever get to their desk. You have no idea what that positive reinforcement does for the radiologist and what it does to make you look like a genius to your referring clinicians. They’re very appreciative and that goes a long way on both ends. That availability of the report expedites patient care tremendously. That’s the biggest gain.

There are also significant financial gains when you don’t have to use a transcription service. You have to consuder the cost of the voice recognition product, but there’s no question it’s cheaper than dealing with a transcription service.

RT: How long did it take the doctors to get up to speed, so to speak? You said, ultimately, it’s time neutral. Roughly how long does it take to get there?

Mendelson: That really varies with the user, but if people were staying an extra hour, I would have heard about it loud and clear. Within one week, most radiologists were not lengthening their work day.

RT: You mentioned financial savings. How long did it take to begin seeing the savings?

Mendelson: We realistically saw some savings when we were doing our pilot with six radiologists. But the big bang was at the end of those 30 days when we converted everyone. At that point, it was almost instantaneous. If your accounts payable lag by about 90 days, 90 days later you’re going to know the difference about what you’re paying out to your transcription company.

RT: What objections were raised, and once you started rolling, how did you address them?

Mendelson: Instead of force-feeding it down people’s throats, what you should do in a large practice is a pilot using your early adopters, the people who will like to do this. They’ll just be less resistant to change and then they’ll set the tone for everybody that follows.

RT: And, presumably, that buy-in helps everyone else?

Mendelson: If you have a couple of successes without somebody shouting, “This is a waste of my effort,” then there’s no negative feedback. There’s only a positive effect and a positive environment set. And so the next round goes to people who you think are more likely to accept it than the ones you suspect who are going to be problems. You identify your potential problems and move them to the back end of the transition.
My other piece of advice is to not install every bell and whistle at the beginning. Just try to accomplish the basic task. If you are moving from digital dictation, just see to it that you can get a report out at the end of the day. There are fancy templates you can generate but start with the basics, then take some training to learn how to use them once you have experienced users. It’s a matter of using the old keep-it-simple principle. It’s fine to have some templates available, but make them simple templates. Don’t make things too foreign to people.
Create a transparent migration from holding a microphone and talking nonstop into the digital dictation system to holding a microphone and navigating the screen as simply as possible—that’s the biggest change in a speech recognition process. So as much as you can make it resemble the prior workflow, at least initially, the quicker your adoption of this is going to be and the better the response you’re going to receive. When everybody is reasonably happy, begin to introduce some more sophisticated utilization of the tool.

RT: Mount Sinai is a sizable facility with, I presume, substantial IT support. How much IT hand-holding was involved in getting this up and running?

Mendelson: This was all supported by our radiology IT subsection. One person was assigned to this full time and was responsible for doing most of the training. His supervisor, the radiology IT administrator, provided some additional help throughout the entire process. But let’s assume that it was like another quarter FTE [full-time employee]. So we did the 30 days with 1 1/4 FTEs.

RT: So now you’re running pretty well. How do you see all that fitting into the future of informatics, images, and records?

Mendelson: Today the major gain is reduced cost and increased quality of care because you deliver that report much faster. Turnaround time is much shorter than it used to be; that’s a quality issue. Those are the two major gains in this time frame.
Looking ahead, these products have the ability to handle macros and templates basically enable structured reporting. Now we’re just at the beginning of that. But structured reporting, if you believe the mantra, someday will provide an improved way of providing quality metrics because if you force certain elements to always be present in a report, there’s a quality improvement.
You should also be able to data mine structured reports better than you can data mine free text. We expect to see a lot of quality initiatives downstream based on this idea.

You can also provide better decision support based on structured text. That should prove useful to the clinician. I’m not talking about decision support on order entry—that’s another field altogether in which there is great interest. I’m talking about recognizing certain buzzwords in structured text that can trigger events with an electronic medical record. Ultimately, that will be very useful.

RT: Where is Mount Sinai on the path to electronic records?

Mendelson: Mount Sinai is about two thirds to three quarters of the way through implementing Epic EMR in our ambulatory environment. We’ve contracted to use Epic in our inpatient environment, too. The implementation go-live date is targeted at the end of 2011 on the inpatient side. We are fully interfaced for our reports to get into the EMR when they’re ready. They get there now, actually.

RT: Are many of your affiliated practices and doctors with privileges getting involved with the EMR? The end game is implementing an electronic system that has all of a patient’s information. Is there much of that or is it still pretty fragmented?

Mendelson: Most of the world is pretty fragmented. I wear a second hat at Mount Sinai; I’m the chief of clinical informatics. Some of the issues I’m dealing with right now happen to be interoperability between disparate EMRs. We’re just at the beginning of that. From what I can tell, with the exception of some very few, limited pilot projects across the country, there’s very little integrated data sharing across enterprises. The stimulus package was really meant to foster getting over that in the next few years, but we’re just at the beginning of that.