What Might Be Lost When AI Takes Over Physician Notes?

One of the most exciting elements of our future healthcare system is the reduction of administrative load, and particularly, documentation burden. It is not difficult to see why: 21% of physicians report spending more than 8 hours per week on the electronic health record (EHR) outside of normal working hours. Rather than relaxing or sleeping during this “pajama time,” many physicians are instead curling up with the radiant red aurora of Epic.

It is not just physicians who are impacted by this problem — patients also have cause to be interested. One large study from Massachusetts General Brigham found that primary care physicians spend a median of 36.2 minutes on the EHR for every 30-minute appointment. A 6.2-minute prolongation may not seem substantial, but over the course of a 16-appointment day, that adds up to nearly 2 hours of delay. And while much of that delay is pushed to pajama time, patients certainly feel the manifestation of this burden in waiting room delay frustration and the sometimes-dissociated conversations of triadic interviewing.

Manual documentation can also make the clinical work itself more difficult. In recent years, “note bloat” has emerged as an oft-cited phrase — referring to the practice of using dot phrases and copy-pasted templates in notes even when that information is not needed or may be medically incorrect. I saw a patient chart recently that said “prednisone started in March” — a reasonable statement, yet the “March” in question was that of 2018, and copy-forwarding resulted in incorrect information that changed management.

Thus, it is not difficult to understand why health systems and companies are investing millions in combatting the need for manual documentation. These new technologies have the potential to make our health system more efficient. This is a good thing. However, as we leave note writing behind, there may be some unforeseen consequences.

The practice of writing has cognitive benefits that improve quality of thinking and problem solving. A large study of California students found that those randomly assigned to write as part of their biology education as opposed to traditional quiz-based pedagogy increased from the 45th to 52nd percentile of critical thinking on a standardized assessment, while those in the standard group decreased from the 42nd percentile to below the 40th. Intellectuals have also affirmed this power of writing: when asked by a historian about his writings as records of his work, eminent physicist Richard Feynman, PhD, objected, saying “they aren’t a record of my thinking process. They are my thinking process. I actually did the work on paper.”

So, it may be important for various areas of medicine to still include some writing. For guidance here, we can look to the psychology literature.

Nobel-prize winning psychologist Daniel Kahneman, PhD, famously wrote of the two types of thinking humans rely upon. In his 2011 book “Thinking, Fast and Slow,” Kahneman defines these approaches: System 1 “operates automatically and quickly, with little or no effort and no sense of voluntary control”; System 2 “allocates attention to the effortful mental activities that demand it, including complex computations.” In medicine, both System 1 and System 2 can be found.

A physician hearing “elephant sitting on my chest” or “worst headache of my life” doesn’t have to actively think about the next step in the clinical algorithms for these conditions (workups to rule out acute myocardial infarction and subarachnoid hemorrhage, respectively). These may not be the ultimate diagnoses, but the first response is rapid and subconscious. System 1 thinking does not require writing.

System 2 thinking, however, might. A patient presenting with gram-negative sepsis who has a past medical history of multiple myeloma, heart failure secondary to cardiac amyloidosis, diabetes mellitus, and chronic kidney disease cannot be easily managed (by all but the most exceptional physicians) through rapid thought. Building an assessment and plan for this patient takes deliberate thinking — weighing risks and benefits of corticosteroids that may palliate neoplastic symptoms but also cause immunosuppressive consequences; balancing the tradeoffs of specific diuretics that may treat the heart failure but worsen the kidney disease. Writing — including in the form of EHR documentation — can be a powerful tool for this type of care.

So, how should we balance reduction of the administrative burden via excessive writing with maintaining the myriad benefits of writing when needed?

To start, perhaps ambient listening and medical documentation technologies should be focused on specialties best served by System 1 thinking, where writing is truly just an administrative burden and not a tool for problem-solving. Emergency medicine — a field of experts in rapid assessment and clinical gestalts — comes to mind. In those areas where writing is a fundamental tool — hospital medicine, infectious disease, oncology — perhaps writing can remain, with technology used to make the process more efficient but not removed altogether.

Along a different dimension, evidence shows that thinking can transform from System 2 to System 1 with practice. That is why a clinical constellation that an attending physician can understand, diagnose, and build a management plan for in 30 seconds might take me, a medical student, an hour or more to solve. Thus, perhaps different levels of training can have different writing requirements (this is already in practice in a more analogue fashion: the practice of attendings co-signing/attesting resident notes rather than having to write their own from scratch).

While AI will inevitably take over many administrative tasks, including writing, the stakes in medicine are especially high. Medicine depends on careful judgment. This judgment is forged through the deliberate process of organizing complexities into coherent narratives — history and physicals, progress notes, consult notes, discharge summaries. If writing is reduced to a purely clerical task, or delegated entirely to AI systems, then we may lose one of the final bastions of the intellectual tradition that has enabled so much innovation and humanity in our field.