Clinical AI Tool UX

This poor cardiologist is struggling despite an autonomous third hand..

I’ve recently started encountering generated replies to inbasket messages in the EHR. The workflow is as follows: you’re reviewing one of dozens of messages that come in from patients through the portal and a AI solution crafts a suggested reply based on the content of the message.

Underwhelming responses

My experience thus far has been underwhelming; it’s a hard thing to get right. To give a specific example, a patient reached out because they were recently in the emergency room and diagnosed with a new condition and started on a blood thinner. The patient was already on two other blood thinners (P2Y12 and aspirin) for stented coronary artery disease (CAD). They wanted to know if the new medication was okay to combine with their existing medications. The generated reply read something like this: “Ms. X, it’s important to follow directions given to you by the emergency room to adequately treat your new diagnosis.” This isn’t helpful and doesn’t address the core reason why the patient reached out: the safety of the new med combined with their old.

Perhaps this isn’t surprising, because to type my reply I needed to look at my last note and remind myself of the patient’s history of CAD, when they were stented, how complex their anatomy was, and their risk of bleeding and falls. I considered all of that before I decided to have them stop aspirin and continue with the P2Y12 and the new DOAC. For a different patient, perhaps with a more recently placed stent and lower risk of bleeding, management might have been different.

Instead of trying to predict the full contents of my reply without sufficient clinical context to tailor it to the patient, I would have appreciated a summary of the patient’s indication for blood thinners (new deep vein thrombosis, history of coronary PCI) along with the details I mentioned above.

Next, once I’ve come to my decision, make my reply easy. One strategy some colleagues and I have brainstormed on: autocomplete while typing. Just like when you’re typing a reply in GMail or Google Messages (it’s an Android thing), a sentence that starts with:

“You should continue your aspirin”

Might autocomplete to:

“You should continue your aspirin and stop your clopidogrel while starting the rivaroxaban for the new blood clot you were diagnosed with to reduce the risk of bleeding.”

That’s definitely an unrealistic “read my mind” suggestion, but it would still be helpful to have autocomplete save me from misspelling clopidogrel and rivaroxaban and perhaps inferring the reason for continuation once the direction is clear.

Maybe that’s overkill and there would be too much latency to make it helpful. Perhaps I just voice or text prompt what I might message back to a member of our clinical team: “continue ASA, stop clopidogrel when starting rivaroxaban” and have that transformed into a message with complete sentences that addresses the patient by name and notes indications for each medication. If the first draft is incomplete, just like Github Copilot, I can refine my prompt slightly to address specific issues: “continue ASA, stop clopidogrel when starting rivaroxaban, provide guidance on bleeding and falls”.

UX first, AI second

What I’ve also been reminded, is that messaging UX in the EHR is suboptimal (being diplomatic here) for several reasons that generative AI cannot fix.

Open your phone or your email, go to a recent message thread, now start composing a reply. Now look at your screen and imagine the following: to start your reply you had to click a reply button that shifts you to a different screen where the original message thread is either not visible or differently formatted. Now, suppose you got distracted by a coworker and walk away from your desk, did what you typed get autosaved? Nope.

Those generated replies I mentioned? Imagine the suggested replies show up elsewhere from where you might look if you were typing start a reply, not in the box where you type your message or next to it, but rather on the first screen you have been taken away from to compose your reply.

So, stepping down from my soapbox I’d summarize by saying this: generative AI tooling cannot fix bad UX at baseline. To be explicit: EHR inbasket should have autosaving and inline replies before AI reply generation.