ESI-matched held 6 hours; the stigma sentence moved the queue
A 19 September npj Digital Medicine experiment held ESI-matched pairs and a six-hour deterioration door while three open-weight models ranked neutral and stigmatizing sentences. The stigmatizing frequent-ED-use line moved rank in every model-dataset cell, up to 9.4 percent.
Science··Midday
The six-hour door
The npj Digital Medicine team built matched Emergency Severity Index pairs at two university hospitals: within each pair, one chart worsened inside six hours while a twin did not. Gemma, Qwen and DeepSeek reordered those pairs before and after a single demographic, social or stigma cue was added. Outcome was six-hour deterioration, not model preference. Eighteen wording grids compared neutral and stigmatizing sentences about the same idea, yielding 221,556 model comparisons across pairs, engines and phrasings rather than one headline patient tally.[1]
Which sentence moved the rank
When the frequent-visit stigma line was tested, every model-and-dataset cell showed harmful reprioritization, topping out near 9.4 percent, and beat the neutral wording by between 1.4 and 7.1 points inside each cell. Holding the clinical idea fixed while swapping only the sentence still moved rank. Psychiatric-history cues also shifted queues, up to about 9.5 percent, though only one engine showed a significant stigma-versus-neutral gap there. Race, language and insurance labels did not move consistently, so the harm was not uniform across every attribute tested.[1]
What a repeat would have to match
This is an experiment on three open-weight models, with ESI-matched and a 6-hour deterioration door. Closed commercial models, another Emergency Severity Index (ESI) cut or a different deterioration clock may not return the same percentages. Treating input selection as a safety bound is a reading of the shift measured in this setup.[1]
