The University of Florida highlighted the study on 26 October as guidance for clinical AI developers. The JAMIA paper was first published online on 5 September.[1][2]
Researchers analysed 22,254 hospital admissions from one Boston hospital’s MIMIC datasets. Inputs covered six hours before admission through 24 hours afterwards; patients were kept separate across training, validation and test sets.[2]
Risk-prediction scores in the main analysis
| Task / AUROC | Structured records | Records, reports and X-rays |
|---|---|---|
| In-hospital death | 0.849 · GPT-oss | 0.853 · MedGemma |
| Length-of-stay category | 0.868 · GPT-oss | 0.861 · MedGemma |
AUROC measures discrimination across decision thresholds, rather than the percentage of patients correctly diagnosed. The added inputs made little difference in this comparison.[2]
For discharge diagnoses and subsequent medication use, three-input MedGemma reached micro-F1 scores of 0.589 and 0.405: relative improvements of 21.4% and 18.4% over the best single-input approach. Micro-F1 combines precision and recall across labels.[2]
The medication task predicts recorded use during hours 24–48 after admission. It does not identify the best treatment. The study evaluated predictions, without testing a bedside intervention.[2]
Sources & further reading
- UF research sharpens clinical AI design University of Florida CTSI · 26 Oct 2026 · Accessed 28 Oct 2026
- Less can be better: decomposing clinical data modalities Peng et al. / JAMIA · 5 Sep 2026 · Accessed 28 Oct 2026