21 October 2026 · Medows
Safe ED AI, Abandoned: Use Fell 68% to 30%
A Nature Medicine pilot found a safe ED decision tool whose use fell from 68% to 30% over four weeks, and dropped with each hour of shift.
The tool worked. The doctors drifted away.
Picture hour one of an emergency shift. The new decision support tool is open on the screen. Someone is curious, someone is being polite, someone was told to try it.
Now picture hour nine. The queue is long. The tool is closed.
That is roughly what a new Nature Medicine paper describes. Researchers evaluated SHAKED, a clinical decision support system built on multiple large language models, in a busy tertiary emergency department. They ran it for four weeks under the DECIDE-AI stage 1 framework.
What the numbers say
According to a summary of the paper by Yesil Science, which links the original study, the safety picture was good:
- Expert reviewers rated 99 of 100 sampled AI outputs as clinically appropriate.
- No adverse events were detected.
- Length of stay was 4.9 hours in both the AI unit and the comparison unit (P = 0.99).
- Consultation cycle times were 9.4 minutes shorter in the AI unit, but that was not statistically significant (P = 0.077).
The adoption picture was not good. Clinician use fell steadily from 68% to 30% across the study period. Each additional hour on shift cut the odds of a physician opening the software by 28% (OR = 0.72, 95% CI 0.62 to 0.83). Physicians did use it selectively, most clearly for radiology consultations (OR = 2.98, 95% CI 1.58 to 5.63).
One caveat: I read the figures through a secondary summary and the paper's listing, not the full text behind Nature's paywall. Check the paper before quoting any number.
Accuracy is not the bottleneck
We spend a lot of time asking whether clinical AI is right. This study suggests a second question matters just as much: is it there when the doctor needs it, in the form the doctor can use, without another window to open?
Think about what a separate tool asks of a clinician at hour nine. Open it. Find the patient. Re-enter or paste the context. Read the output. Decide whether to trust it. Every one of those steps is cheap in a demo and expensive on a full ward.
The selective use is telling. Radiology consultations are a discrete, well-bounded question. The tool fits a moment. Most of a shift is not a series of discrete moments. It is interruptions, handovers, and half-finished thoughts.
Where this meets the workspace idea
This is the argument for building around the doctor's shift rather than adding a point tool beside it. If the AI sits inside the place where the patient list, the notes, and the plan already live, there is no tool to remember to open. The context is already there. The cost of asking drops toward zero.
That is a hypothesis, not a result. This study did not test it. What it does show is that safe and accurate output can still lose to fatigue.
What to watch
- Whether adoption curves become a standard reported outcome in clinical AI trials, next to accuracy.
- Whether larger, longer studies see the same decay across the shift.
- Whether vendors design for hour nine, not hour one.
The honest read: a decision support tool whose use fell from 68% to 30% has not failed on safety. It has failed on fit.
Sources
- Yesil Science, "Doctors stop using ED AI as shifts progress" (Oct 16, 2026): https://yesilscience.com/doctors-stop-using-ed-ai-as-shifts-progress/
- Nature Medicine, "Prospective evaluation of a large language model clinical decision support system in the emergency department": https://doi.org/10.1038/s41591-026-04601-5
- Yesil Science, The Health AI Brief, week of October 19, 2026: https://yesilscience.com/the-brief-2026-10-19/
Medows is a clinical AI workspace for the doctor on rounds. Learn more or write to alapan@medows.ai / alapanx@gmail.com.