Tech shows it can pave cow paths with the best of them
Ambient AI could possibly cut the work required to create clinical notes. But clinicians plead for quality, not a quicker path to note bloat.
I heard this warning early on in writing about healthcare information technology – we don’t want to just pave cow paths.
As IT was incorporated into healthcare delivery, the warning seemed appropriate. The fear was that IT would just be used to automate inefficient processes, leading to them becoming ingrained ways of doing things. That would just promulgate inefficiencies, rather than taking a fresh look at processes to see how technology could solve them.
Well, perhaps unsurprisingly, there was a lot of cow path paving in those early days. That’s because it’s difficult to envision how technology can resolve inefficiencies, particularly because inefficiencies are deeply embedded in operations and it’s hard to envision how technologies could resolve them.
But now, after several waves of implementing new technologies in healthcare, the industry should be more aware of how to eliminate inefficiency and ensuring that technology actually makes processes better, right?
Well, some recent discussion of artificial intelligent-supported note taking suggests that while AI is lightening the load for clinicians on the front end, it may be making things worse on the back end.
Note bloat, take two
The process of creating notes from clinician interactions with patients has always been a pain point, going back to before time of paper records, when thick stacks of sometimes illegible, time-sequential notes comprised a patient’s medical record.
Electronic health records systems helped eliminate piles of paper, but flummoxed clinicians in multiple other ways. Many versions of EHRs forced clinicians to type in notes (not always their strong suit), pulling their attention away from their interactions with the patient. Then, screens with check boxes sometimes sped up some of the process. But converting these menu-based boxes to verbiage for the official medical record eventually led to massive records for each patient encounter.
It’s called note bloat colloquially. A recent study described the issue this way. “Electronic health record design often prioritizes billing and administrative tasks over clinical decision-making, and care delivery, leading to clinicians spending a disproportionate amount of their workday on EHR and desk work … The phenomenon of ‘note bloat,’ where US physicians write notes that are 4 times longer than those in other countries and require too many steps to perform simple functions, is a substantial issue.”
The process of writing and finalizing patient records forced clinicians to spend inordinate amounts of time, often after work at home laughingly called, “Pajama Time,” Approaches such as charting by exception – eliminating in-range results or observations of normal functions from the record – often raised legal concerns about non-documentation of all patient observations.
So, records took too long to write, and the outputs were too long to read. This is where ambient AI was supposed to swoop in and save clinicians. Well, it did, kind of. The technology can “listen in” on a physician-patient dialogue as an AI scribe and create notes for the encounter, without the clinician having to write the bulk of the note.
On the front end, that’s been a win for clinicians, who have seen a reduction in their work needed to create the medical record. And there was hope that the nascent AI technology would eventually help bring efficiency to the entire records process, particularly that it would add more meaning and value to the end-product.
But maybe not yet
The use of ambient AI to assist in creating notes has some unintended consequences, according to an article published in the National Library of Medicine.
While it’s acknowledged that AI scribes cut down on active chart-writing time, as many as one-third of clinicians say they spend hours fixing AI drafts to catch errors or documentation inaccuracies. Worse, unedited AI notes show a tendency to replicate the copy-and-paste tendencies that bloated charts in the past. And some worry that this will degrade data quality that will negatively impact machine learning models in the future.
In fact, a recent article predicts that healthcare’s “note bloat” issues will only be made worse by AI without some fixes. “Most systems record an encounter and dump a monolithic draft that clinicians must fix line by line, trading keyboard time for proofreading time,” it contends. The article, developed by a Denmark-based company, promotes its product which applies AI reasoning in real time, “treating a consultation like a stream of evidence, not a wall of text.”
Clinicians aren’t happy
And as anecdotal proof, a single comment about AI-generated notes on a prominent social media platform provoked a torrent of disdain from those on the receiving end of those notes.
All the East Coast specialist had to post was this: “The EMR generated notes that are faxed to my office are unreal, what a s*** show.” That elicited a wave of pointed criticism of the technology.
“In a previous era, my 3-sentence post-op note with copy of pathology report was a helluva lot more informative than the crap I get now,” another doctor chimed in.
“(Nurse practitioners) with no formal training in medicine generating AI written slop with again, no formal training in medicine,” complained another clinician. “Doubling the ignorance does not improve the outcomes.”
Another clinician termed the AI-generated notes “a complete waste of time. You scan five or more pages of computer-generated nonsense to find at most 1-2 lines of human produced content that usually isn’t even that helpful.”
“I get 12-14 pages of notes for a 3-year-old who had a cold and then developed an ear infection. CAN I JUST HAVE THE CLIFF NOTES VERSION?” argued a doctor.
Several commenters pleaded for the technology to winnow out needless data and get to the point that can impact and improve care.
“Wedged into 41 pages of e-vomit will be one tiny little lonely assessment on a single page with mildly relevant information,” noted one doctor.
Can we just get to the point?
“The biggest problem is the SOAP note,” contends one clinician. “The assessment and plan must be at the very top of the first page of every note. No other attending physician cares that you can do a physical exam or take a history.”
While AI-generated notes may not yet improve care, they continue to support the business need of providing evidence useful to procure reimbursement from payers, another clinician contends. “Funnily enough, the EMR-generated notes I get faxed to my office are usually great from a prior auth perspective,” he says. They contain “the exact language the payer expects, even if it’s 100 percent boilerplate.”
And some respondents to the original post still fear that AI fabrications and hallucinations will sneak by clinicians reviewing the AI-generated notes, enabling errors to be included in a patient’s medical record.
In sum, moderation and mediation will be crucial in incorporating AI-generated notes into medical records.
“I’ve been doing some informatics for 20 years,” concludes a Midwestern physician. “The longer it goes, the more I feel like it should just be a dictated note plus labs, like it used to be. Voice to text. If AI can parse my note or H/P for billing, great. Otherwise, let billing folks do it like the old days.”
So while ambient AI has great potential to reduce clinician workloads, much work lies ahead before clinicians and patients can derive the full benefits and avoid consequences.
Fred Bazzoli is Editor Emeritus for Health Data Management.