Will AI Smooth Out the Jagged Edges of Healthcare?

It's AI vs patient stickers and fax machines

The invisible wall of healthcare AI

I went into my first hospital volunteer position at age 14 with visions of witnessing ER-like scenes of people being dramatically brought back to life, diagnosing exotic diseases, and figuring out what kind of doctor I wanted to be. Instead, I helped the unit clerk with what she referred to as “sticker management”. Managing the stickers meant printing patient stickers out when a patient arrived, printing out more of them when a patient went to a different unit, and giving them to the many healthcare professionals who needed them. 

Nearly 30 years later, stickers have been one of the constants in medicine. Meanwhile, much of the medicine has changed: the Human Genome Project hadn’t yet released a version to the public, and now you can buy DNA for $0.30/tube. 

In all the hospitals I’ve worked at, the sticker printer is one of the few pieces of equipment that can cause every workflow to take an additional 20 minutes and make some completely impossible. Does your patient with an ankle fracture need pain medicine? The nurse needs to scan the sticker on the patient’s bracelet, so giving medication without a sticker becomes a major ordeal. Want to consent a patient for surgery? Without stickers you have to hand-write all their information on both sides of every form. 

This archaic form of medical document management is an example of the jagged frontier of healthcare, similar to the “jagged technological frontier of AI” described in a recent business paper. The authors describe that “some tasks are easily done by AI, while others, though seemingly similar in difficulty level, are outside the current capability of AI.” The paper describes an invisible wall between helpful and detrimental AI:

“The problem is that the wall is invisible, so some tasks that might logically seem to be the same distance away from the center, and therefore equally difficult – say, writing a sonnet and an exactly 50 word poem – are actually on different sides of the wall. The AI is great at the sonnet, but, because of how it conceptualizes the world in tokens, rather than words, it consistently produces poems of more or less than 50 words.  Similarly, some unexpected tasks (like idea generation) are easy for AIs while other tasks that seem to be easy for machines to do (like basic math) are challenges for LLMs.”

In the same way, some medical tasks that seem incredibly difficult are easy, like matching a single genetic defect with a targeted drug, while other tasks that seem easy, like document management without stickers, are challenges. The location of the “wall” in healthcare is just as invisible as it is in AI, but we’ve accepted it as normal. 

AI has the potential to smooth out the jagged edges of healthcare by making some of the ‘hard’ tasks easier, and moving the ‘invisible wall’ to improve efficiency and processes.

Some ML for MDs physicians were kind enough to give some additional examples of the invisible wall in healthcare. Taken together, these ridiculous inefficiencies and manual processes are a roadmap for where I hope AI can smooth out the edges of the jagged frontier of healthcare technology.

We need to talk about faxes

75% of medical communication in the US is via fax. Before the generative AI boom, I was part of a hospital leadership team that tried to get rid of faxes and a consultant to a company that tried to do the same. They were both marginally successful. Much of the problem is related to lack of interoperability of EHRs, which some standards like FHIR might help with. But AI could play a role as well. 

In the short term, I’m hopeful that Optical Character Recognition (OCR) can actually integrate all those faxes into the medical record instead of having them sit in a “Media” tab that involves scrolling through 75 pages of scanned records to find the most recent platelet count. Even if the EHRs can’t talk to each other, AI could format the information in one EHR so that another EHR could “read” it, or at a minimum format the note so the relevant information is easily visible. 

Faxes are currently an impediment to good care. Graham Walker, MD mentions “SNF transfer paperwork is either 50 pages of garbage or just a paper face sheet without literally any information about why the SNF patient was sent to the ER or  documentation about the patient's baseline mental status or baseline mobility or baseline wounds, etc”.  Clearly there’s a role for AI to produce something in between a single page with no information and 50 pages of information. Another possible role for AI in the short term would be a chatbot that could locate information in a long fax without the clinician having to scan through every page. In a task that’s reminiscent of Office Space, Dr. Walker also notes that sometimes he receives a fax via email, then has to print out the email, sign it, and fax back the paper that was previously a fax. 

As far as I know, Dr. Walker hasn’t yet been compelled to smash his computer as in Office Space

Data abstraction and aggregation

Healthcare produces a ton of data for every patient, only a fraction of which is ever used in a meaningful way. Much of the data collection is done electronically now, but as Justin Graham, MD notes, many of the data registries that determine quality of care metrics rely on humans, usually coders (41%) or nurses (27%), going through a chart and picking out the relevant details to put into a different spreadsheet. 

Generative AI excels at finding information in large texts, and it’s hard to imagine why we would keep having humans do this. There will likely need to be a human in the loop to resolve questions or unclear categorization, but that should be the exception instead of the rule. And some of this manual data abstraction happens real-time also. As Justin Graham, MD also notes, many case managers have two computers side by side to abstract data from the EHR (computer 1) and compare it to insurance data (computer 2). This is clearly ridiculous, but hospitals’ willingness to pay for this arrangement shows that there’s not an easily available alternative. 

The difficulty with large datasets also affects clinical care. As Graham Walker, MD notes, it’s really hard to get an understanding of a patient’s glucose levels from glucometers or continuous glucose monitors without printing out an entire spreadsheet. There are many similar devices like pacemakers, arrhythmia detection devices, and others that print out huge amounts of data that might be easily visible to the specialist who always deals with them but is much harder for the many other kinds of physicians like emergency physicians, anesthesiologists, and primary care docs who also need to access that information. Having AI chatbots for these devices that a physician could query for things like estimated HgbA1c and frequency of Vtach runs would make it much simpler for physicians to care for patients with these devices.

Aggregating information via pre-rounding, or gathering all the data about a patient and putting it into a standardized formats, is the time-honored job of medical students and interns. As Ajay Perumbeti, MD mentions, somehow pre-rounding in the era of EHRs has actually decreased the time physicians spend with patients. I have vivid memories of starting to pre-round at midnight at Santa Clara Valley Medical Center as a medical student, stopping at each patient’s room to write down the range of blood pressures and heart rates. Somehow the transition to EHRs hasn’t actually decreased the time medical students and interns spend pre-rounding or combing through records when a patient is admitted. Not only would data aggregation by AI be faster, it would be more accurate than the sleep-deprived transcription of numbers that I used to do.

What else is on the other side of the invisible wall of healthcare? 

AI isn’t going to smooth out all the jagged edges of healthcare. I can’t think of obvious ways that AI might help with some of the issues below, though I’d love to know if you have ideas! Below are some other areas that ML for MDs physicians mentioned as being anachronistically analog in a digital age:

• Communication methods were mentioned by ML for MDs physicians multiple times. 

  • Ethan Goh, MD notes that the NHS in the UK still uses pagers. These are just a monumental waste of time, as doctors page someone then have to step away from the phone and miss the call back, or nurses page a doctor who then calls back the number and has no idea who paged her. I haven’t used a pager in 7 years and don’t miss them at all. What I remember most about them is having to use 5 different pagers and them being so heavy on my scrubs that my pants would fall off when I ran to an emergency intubation.

  • Our hospital seems unable to figure out a way to communicate as effectively as a cellphone, but also doesn't want people to use cell phones. In practice, this means that in the OR I have a portable phone (not cell phone) that takes a full two minutes to “turn on”. I am pretty sure this is a technology issue that we solved in 1987. This is less than ideal in an operating room, where you really don’t want to wait two minutes for the phone to turn on when you need to activate a massive transfusion protocol. It also has no saved numbers in it, so I have to look up the number for the blood bank on my cell phone then dial it into the portable phone.

This phone takes 2 full minutes to turn on

• Paper forms, in all their forms

  • Graham Walker, MD mentions the number of paper forms he has to manually sign for low-risk therapies like outpatient oxygen, even though we electronically sign for high-risk interventions all the time.

  • Justin Graham, MD mentions paper forms for vaccines and school clearance forms

  • Graham Walker, MD also mentions “printing out ER discharge paperwork because it's required by law to provide paper even though the CURES Act requires all patients to have immediate access to it digitally as well”

Conclusion

We all navigate the jagged edges of healthcare every day, sometimes filling out three different paper forms for a patient’s cutting-edge medical device. We’ve taken these edges as immovable for too long, and I hope that AI can at least allow us to think of how we can smooth out the sharpest edges that affect patients and clinicians. AI will eliminate some of the roughest edges while others will require regulatory action or creatively revamped processes. But I hope we can all start to at least notice the invisible wall of healthcare, how it overlaps with the invisible wall of AI, and where we can use the strengths of the new technology to improve our processes. Or at a minimum, we can start to imagine a future without patient stickers.