Several artificial intelligence (AI) applications promise to reduce the work associated with charting patient encounters. I had the opportunity to test-drive one AI-powered medical scribe for my outpatient clinic.
Prewriting went much faster with the scribe program. I did not need to organize the information that I came across during my chart review. I could speak the details of a hospital stay, notable medication needs, or changes in lab data, and the AI scribe would summarize these and place them in the proper sections. I was able to cut my prewriting time down from 90 minutes for a full day of clinic to 20 minutes.
During the patient visit, I found it liberating to not have to type while listening. I was more conscientious of carefully outlining my reasoning to assist the tool in accuracy, and the patients noticed and appreciated the extra effort as well. The scribe took about 40 seconds to generate a note from the encounter, which I could transfer into the electronic medical record (EMR) and edit as needed. The overall visit length did not change, but I was able to spend much more time in direct conversation with my patients and less time in tedious documentation.
Regarding accuracy, I was impressed with the tool's ability to capture and organize tangents and additional pieces of information. If patients mentioned a past problem with a medication or an element of their medical history, it was reliably documented. However, I found that the application struggled with cause and effect. An event such as a medication change based on a side effect was often incorrectly summarized. I tried to counteract this by restating the information back to the patient, but this did not improve the tool's accuracy. Medication reconciliation was also a challenge; I found myself deleting a lot of the verbiage that the scribe placed in the medication section of the note, simply because it was redundant and looked messy in paragraph form.
Another area needing improvement was the capture of the small human details that go to the heart of the doctor-patient relationship. One patient shared at length how their recent bereavement had affected their motivation to care for themselves. Another talked about the joy he had from his new puppy. All of these stories, which I tend to capture in a few phrases in my written notes, were left off completely by the AI scribe. I include these details and refer back to them to develop rapport with my patients. The dog's name is important!
If I am being frank, the AI scribe is a terrible writer. It uses the passive voice all of the time, relies on jargon words like “utilize” far too much, and creates a note that is very boring to read. When I reviewed these notes weeks later, they did not help jog my memory of the patient's personality and the nature of our interactions. As someone who likes to write and appreciates good writing, I really disliked the scribe's style. My notes are one of the primary ways that I interact with colleagues; my writing serves as my professional voice. I would place significant value on an AI program that could write the way I do.
Finally, so much of the cognitive work that takes place in the EMR happens outside of the written note itself, and the tool offered no shortcuts. Medications, history elements, and diagnosis codes must all be placed in their own discrete data fields. It would be a significant leap forward to discuss the diagnosis during the visit and have the scribe code the encounter correctly or to order the labs and medications we discuss. So far, this capability is not offered.
AI scribes offer a great deal of promise to reduce the burdens of charting, although significant improvements are still needed (Table). Practicing physicians should seek to give their input along the entire pathway of AI development. The EMR was not really designed with us in mind. Let's not make the same mistake with the next generation of technology.
Advantages and disadvantages of using the AI scribe
