The article about Pediatric Associates in CA has a nugget with a potentially outsized impact: the implication that VFC vaccines…
Joel Diamond 4/22/09
Why We Need Natural Language Processing
Chief Complaint: “Are my testicles black?”
Operative Note: “The patient was brought to the operating suite. She was propped and raped in the usual fashion.”
Granted, these anecdotes came from colleagues using much earlier versions of voice processing software, but they show the inherent need for voice understanding.
Oh, I forgot to tell you the actual dictation:
Chief Complaint: “Are my tests all back?”
Operative Note: “She was prepped and draped in the usual fashion.”
Most long-time EMR users will agree. For documenting the ROS and physical exam, drop-down lists and templates, particularly when the users have customized them around their own workflow, have proven to be incredible time-savers. The problem is that History of Present Illness and Assessment/Plan are best expressed using free text. Moreover, these are the most critical parts of the record, as they reflect nuances and thought process associated with the true art of medicine.
Since my typing skills have improved drastically, I continue to express these areas with significant detail. To me, the thought of using voice recognition software just seems like a less efficient move from dictation. I know that there are many of you out there who have enjoyed tremendous cost savings and efficiencies, particularly with the vastly improved newer versions of this technology, but I’m holding out for something better.
Natural Language Processing (NLP) is a method that transforms text into structured data.In essence, it understands text. Clearly our growing requirements for reporting and analytics will make this technology essential, yet it continues to be discussed mostly in academic circles. Significant advances have been made in this technology and I believe that incorporation would enhance EMR adoption. Without it, I worry that we will continue to add unstructured (i.e. unusable) data into the collective medical record.
I would love for other readers to comment on this subject. (Equally important: my immature and irreverent side would love to hear more anecdotal voice-processing gaffes like the ones above).
Joel Diamond, MD is chief medical officer at dbMotion, adjunct associate professor at the Department of Biomedical Informatics at the University of Pittsburgh, and a practicing physician at UPMC and of the Handelsman Family Practice in Pittsburgh, PA. He also blogs on interoperability.
Fully agree with “drop downs” good for Physical exam, but need free text for HPI, Impr/Plan. And for one time visits, dictation is likely very efficient. But for long term, chronic care of a patient (ie PCPville), I think that dictation is actually less efficiect than typing – ASSSUMING, that you are working on the model of copying forward the last note and editing it for today’s visit- which I think is actually the best way to practice. I can see what I said last time, include, dlete or edit it, and with minimal typing have an updated note.
And yes- there are anectdotes about how docs can copy forward garbage and make the note look ridiculous, but those stories are about hospital notes with multiple authors (many whom are in training) working on inpatients with quickly changing factors… very different than an established physician taking care of a long term patient over many years…
But either way, I agree NLP will be a much greater player in the future than structured documentation.
Just as funny are transcriptionist errors. I once read a discharge summary that said the patient was sent home on an “h tube clocker”.
Received yesterday a (four page! – shortened here) child’s ER visit dictation gaffe…
CC: Cold symptoms with fever
Diagnosis: Urinary tract infection
Plan: Script for Amoxil given to be filled if child develops ear pain
(At least I hope it was just a dictation gaffe.)
I, too, hope NLP moves from the academics to the trenches soon.
How about a combination of two technologies- the ease of copying forward the previous visit coupled with speech recognition (by which we all mean Dragon) to dictate the variations in real time. We do this with Epic which, btw, has NLP to some degree.
I think that NLP engines tied to dictated reports can be powerful tools. Particularly if they can create CDA documents as outputs. This is a bright future, however I haven’t yet seen this reality in a production setting yet. Structured templates are also a bright future, with the great efficiency gains that features like carry-forward and dropdowns promise, they also introduce a risk we now refer to as ‘documentation hygiene’ risks. That is that poor documentation or temporally subjective information (ex ‘do something tomorrow’ or ‘did something today’) can be propagated and replicated over time. We keep an eye on this in production and focus on education of good documentation practices to fix.
I agree that NLP shows tremendous promise, but is only as strong as the output format. With the slowly adopted (and then discarded) standards for data-transmission into EHRs, it is time-consuming to keep re-inventing the output file format. I believe that if we take the EHR vendor out of the equation, we can utilize NLP for a host of other very cool and lucrative ends. Querying otherwise flat data from a transcript could lead to better reporting, more analysis, PQRI and possibly drug trials…all things that the small practice is otherwise precluded from participating.
If we think larger about what NLP can offer and stop focusing all of our HIT energy on EHR as the final solution, I believe we can start to really create some powerful tools for the practice of medicine that doesn’t intrude on the physician’s care of patients.
Why does “unstructured” data equate to “unusable” data?
In my field, anthropology, all the data is unstructured. The art and science of anthropology is to see structure in it. A checklist won’t do. Since in medicine you’re dealing not only in symptoms but real people with real, individual histories, and ways of seeing the world, doesn’t it seem to you that trying to understand a whole person, however difficult it is to do, may in fact strengthen the healing process in that person?
When we last looked at NLP via nuance, it was licensed per report, and seemed cost prohibitive (not to mention retrospective). having the function baked in to the application would be helpful. i’m optimistic for the future uses, still.
I use Dragon for HPI and Plan. I use tools like drop down lists, templated text and right left click on postitive/negative lists for everything else. The lists then transform to standard English. I use active links to bring med lists, problems lists, diagnoses and data into a note or letter. I use Dragon for commands and to do email. I have Dragon on the system server and use the same processes for inpatient work. We have no paper in the hospital.
BTW, struggling with CPOE…
anyone successfully using Dragon for ED?
Dragon in ED:
We have eliminated outside transcription in favor of Dragon in our ED’s at Kaiser in the Sacramento Valley. Since we have Epic, most docs in our ED use Templates, Dragon or both.
In fact, as of April 15th, 2009 all outpatient transcription was eliminated and Dragon is offered to any doc who desires it.