Are EMRs Killing the Clinical Narrative?

Could your new electronic medical record system be missing vital information the old paper-based system captured?

Even the most seasoned technology champion has to stop and ask that question, if for no other reason than the new medical record looks very different than the old one. To put it in classroom terms, today's EMR is often multiple-choice, not essay.

But almost as long as there have been doctors, the preferred way for them to communicate has been through a narrative—a story.

EMRs may introduce gaps in that narrative, says Philip Resnik, professor of linguistics in the Institute for Advanced Computer Studies at the University of Maryland.

Since 1999, Resnik's been studying the limitations of entering clinical information into discrete fields and checkboxes in an EMR. At the recent South by Southwest conference, Resnik described the dilemma clinicians face: to embrace the EMR with all its limitations, or to push ahead for new technologies such as natural language processing that rarely see clinical use today.

Resnik illustrates the problem with a sample narrative of a woman complaining of shortness of breath. In a slide he highlights snippets that are easily entered into EMRs, such as symptoms and actions taken. But he also underscores much text that helps tell the story of the patient's encounter in the ER but doesn't readily map to fields in an EMR.

"The doctors in the ER were trying to figure out whether the shortness of breath in this woman was due exclusively to her failing heart, or was there a problem with pneumonia," Resnik says. "People who have pneumonia do not respond promptly to [BiPAP] treatment. But she responded promptly. This gave them information."

Resnik bets that few point-and-click EMRs have a check box or slider control for how quickly a patient responded to a treatment.

Text fields in EMRs can capture this information, but in a busy exam room, with doctors trying to point, click, and enter EMR data during the exam, while also trying to maintain eye contact with the patient, how much time will be left for text entry?

The dilemma compounds when you realize that any data entered in text fields will resist analysis. Database analysis works best with discrete numbers. So even if we get doctors to enter the portions of their narrative that don't fit in discrete data fields, we've lost the ability to really analyze that data.

As an experiment, Resnik and some other researchers took 20 cardiology dictations and went through them manually, highlighting the info that could be placed in discrete fields, without having to type into a text box.

"Then we took two cardiology experts and said, 'Let's pretend this clinical record is somebody a doctor across the country referred to you as a case,'" Resnik says. Researchers had highlighted info that couldn't be placed in the discrete fields, and they asked the cardiologists to rate how severe a gap in the record the highlighted information was.

In half the records, there was at least one thing the two doctors independently concluded should have been in the patient's record.

The researchers did another experiment where they assumed that the EMR, which happened to be in use in British Columbia, could capture more of the narrative with some extra engineering.

The cardiologists still found a severe problem in one out of four records, Resnik says.

Another issue with EMRs is the advance of medical science. In the early 1990s, a higher-resolution CT scanner was introduced. Radiologists started discovering semi-opaque nodules in the lungs which indicated a much higher probability of lung cancer. But older medical records simply offered the choice of "opaque" or "transparent" and had no way of expressing the newer notion of "semi-opaque."

Such examples must abound in medicine as it advances. How valuable will today's EMRs be in tomorrow's realities?

The traditional clinical narrative also has another set of nuances not present in the typical modern EMR. Narratives may say that something is "suggestive of" a particular condition without that condition actually being present. Patients may deny the presence of a particular condition, such as chest pain, but the EMR may not allow for such a denial to be a structured part of the record. In another example, doctors may agree that a particular pilot-as-patient should not be recertified to fly without undergoing a particular procedure.

"I have a feeling 'Don't recertify patient to fly without this procedure' is not a check box that is easy to put into this medical record," Resnik says.

Symptoms also change over time, and EMRs may not be nearly as good as a narrative when expressing this.

On top of all these concerns, a generation of older clinicians who are used to simply narrating their records creates a recipe for a mass exodus of personnel on top of growing doctor shortages.

Resnik worries that with the current stampede to meaningful use, all these considerations are being ignored.

As somewhat of a salvation, work continues on natural language processing. Resnik, who consults in this field, notes that machines are making strikes in learning to read, parse, and code narratives, partly because of the recent move to "big data" and advances in machine learning such as IBM's Watson project.

In other fields, including marketing and advertising, big data—the sophisticated analysis of very large data sets—is a big deal. Healthcare tech seems to be late to the game. Too many of today's EMR solutions seem to be based on the old-style client/server technology of the 1990s.

In Resnik's opinion, doctors shouldn't be checking boxes while they're trying to do a narrative. He says there are ways to "engineer the ergonomics" of the system. He, and I, think it's time we do.