AI, SOAP notes, and the EMR

Is longitudinal care at stake?

The ominous, ever-growing burden of clinical documentation in primary care is a problem that needs no introduction.  Is there a solution?  Enter AI to save the day!  The market is brimming with new software products, both within EMRs and as add-ons, that implement ambient listening technology, or “AI scribes”, to help clinicians document their patient visits.  The promotional copy from vendors generally reads something like, “Reduce your documentation time by 75%!  Let our AI software write your SOAP note for you!”  And clinicians try it.  And many like it.  The software listens to the clinician-patient interaction, applies AI using Large Language Models, and creates a summarization in plain (somewhat medical) English for the key narrative sections of documentation.  And clinicians truly do save time.  The AI SOAP notes, while perhaps verbose, are generally sufficient to let the clinician move on to their next patient.  And the accuracy, efficiency and discretion of these products is rapidly improving.

But on deeper analysis, I arrive at some fairly unpopular conclusions.  I portray the primary care clinician, the burden of documentation, and the advent of AI scribes using the following analogy:

A person is cursed with an itching, burning rash on their back.  This rash absorbs an inordinate amount of attention, is distracting, unpleasant, and basically ruins their day.  Along comes an inventor selling a wonderful new robot with long sharp fingers.  The robot follows the itching person around all day, intelligently scratching their back.  And it works!  They feel relieved, and their day goes better.  But their rash never improves, and may even worsen, making them ever more dependent on scratching.  As robot technology improves and is widely adopted, people increasingly focus on itching as the core problem to be solved - because it seems solvable.  Efforts to find a way to heal or prevent the rash dwindle, and are eventually forgotten.

This deserves some explanation, which requires some background on primary care documentation.  Most readers of this article will be familiar with the standard SOAP note (if not, the NIH website contains a nice summary article).  The SOAP format is especially useful for hospitalized patients, acute visits, and some specialist visits, but was never particularly well-suited to chronic care or primary care.  PCPs deal in longitudinal care, which means that the main narrative sections of their documentation - the History of Present Illness (HPI) and Assessment and Plan (A&P) - contain copious data that stays the same from one visit to the next.  This presents a challenge for quality documentation.

Primary care clinicians will understand what I mean.  The SOAP note works mostly in the “now”, meaning around a single encounter.  Imagine a patient with heart failure is seen by their PCP.  The narrative of today’s events, in the HPI and A&P, might look like this:

HPI:
Generally feeling better, with decreased swelling and less shortness of breath.  Maintaining low-salt diet and taking an extra dose of diuretic with large meals or eating out.

A & P:  Chronic systolic (congestive) heart failure (I50.22)
Continue current medications, including bumetanide 2 mg daily and 1 mg PRN.  Next echo scheduled 10/2024, appropriate f/u in place.  BMP ordered.

That is fine for the PCP today, but what about later, when the patient is sick and is seen by a covering physician?  This note has no succinct summary of the diagnosis.  The Medical History section of EMR records is seldom helpful (for many reasons).  So the clinician must scan the EMR for multiple chart notes, study the medication list, read consults, find the last echocardiogram, etc.  They need to know: how bad are this patient’s CHF exacerbations?  What do “better” and “less” mean in the previous HPI?  Better than what? Why is the patient not on aspirin?  How involved is the cardiologist?  When the patient is elderly, on many medications or has low health literacy, these questions get exponentially more difficult.

The most ambitious clinicians address this by creating complex assessments for complex problems, which serve very well in subsequent encounters.  The first part of each A&P is relatively static and is carried forward from visit to visit, whereas the second part changes each time.  Both may be formatted in bullet points that are easy to modify.  For example:

A & P: Chronic systolic (congestive) heart failure (I50.22)
   • Diagnosed 2020, related to mitral valve stenosis (non-rheumatic)
   • Lowest EF at dx 35%.  Since 2021 it is consistently > 55 %
   • Cardiologist Dr. Pumpalot, q6m, consider eventual MV replacement
   • ASA stopped 7/2024 (recurrent UGIB)
   • On ACE, HCTZ, Bumetanide (daily + PRN), Metoprolol.
Today:
   • Continue current medications, including bumetanide 2 mg daily and 1 mg PRN
   • Next echo scheduled 10/2024, appropriate f/u in place
   • BMP ordered

This looks like a sound solution, but practicing clinicians know: creating the summary takes time, and “pulling it forward” from visit to visit is difficult and error prone.  It may result in redundant orders, spurious documentation and many other inaccuracies.  When a patient’s care is shared by multiple clinicians, it gets even messier, because A&P format is arbitrary and personal to each author.  This practice is unsustainable for most full-time clinicians.  Our EMRs are not designed for it.

The long-term answer is to redesign EMRs with longitudinal documentation and planning as a core architectural feature.  The SOAP note should be something that the EMR generates as needed, not the primary data construct for encounters (a larger topic, for another time).

Back to the AI discussion: if, in light of this explanation, we view reliance on SOAP notes as a liability of today’s EMRs, it complicates the analysis of ambient listening technology.  The main output of an AI scribe is in paragraphs, distributed among the narrative sections of a SOAP note.  This certainly has advantages:

• The clinician types less, and can pay more attention to their patient.

• Completing documentation after the visit is faster, especially for clinicians who type slowly.

• If a clinician is finishing documentation and orders later (like after hours), the AI narrative makes it easy to remember what happened in the visit - a potentially huge time saver.

But other things are not so promising:

• PCPs who want the benefit of today’s AI scribes have to compromise their detailed, bulleted longitudinal plans.  AI scribes produce verbose paragraphs, not efficient summary text, and are not compatible with the pull-forward-and-modify approach.

• Charts composed of verbose, paragraph-based SOAP notes make it harder to reconstruct the patient’s longitudinal story, and harder to practice team-based care.

• AI scribes’ use of SOAP notes as their de facto supporting construct will only further solidify the role of the SOAP note in our EMRs, now and in the future.

• The short-term success of AI scribes relieves the pressure on EMR vendors to rethink existing EMR architecture, especially around longitudinal care.  

Perhaps the most demoralizing aspect for a clinical informaticist like myself is that many primary care clinicians long ago gave up on maintaining diagnosis summarization as part of their A&P.  EMRs have so utterly failed to support this practice that it is not a documentation standard.  So clinicians don’t apply pressure on their EMR vendors to facilitate it, and don’t value it as something they might lose.  They welcome AI scribes without a thought for the effects on underlying EMR architecture or its developmental direction.  They just need their backs scratched!  Can you blame them?

The day that AI produces problem-based, accurate, bulleted summarizations, the landscape will change.  That may be some way off; it requires weighted, chronological analysis of a wide array of disparate data, and sophisticated management of variations and contradictions in the medical record.  It will carry extraordinary liability.  But if and when AI can generate summary data that a clinician can sufficiently trust, it will support longitudinal care in a way that largely eludes us today.  Ambient listening technology will surely someday combine with AI record summarization, for a powerful hybrid documentation solution.  This will be AI as “Assistive Intelligence”, rather than just artificial.

In the meantime, ambient listening technology will provide its short-term relief.  The effect on the future of underlying EMR infrastructure remains to be seen, but the prognosis is worrisome.

The Functional EMR Guy: This channel offers a view of ambulatory EMR technology from my experience as a clinician, clinical informaticist and software engineer.  My driving vision is a practical approach to functional analysis and design, to help clinicians help patients by creating better EMR systems. Please subscribe! I won’t send spam or share your info.