From clinician review to measurable human oversight of AI-drafted paediatric discharge documents

A randomized study by Morquin, Bochu, Robert et al. randomly assigned 153 children discharged from a pediatric emergency department to either a conventional template group or a group drafted by a locally deployed open-source weight large language model, with both groups' documents reviewed by the discharging physician, and 117 cases were ultimately analyzed (59 in the conventional group and 58 in the model-assisted group). The paper states that the family version documents in the model-assisted group used more common vocabulary, but there was no difference in the primary outcome of parents' understanding of record consistency at follow-up (difference -0.01, 95% confidence interval -0.12 to 0.10). In the clinician-facing handoff documents, no discrepancy was judged to be a serious consequence (0/117). In exploratory comparisons, physicians modified 48 of the 58 drafts, removing 18 of 22 unsupported additions, but corrected only 6 of 22 erroneous abbreviation expansions. The authors propose that documentation elements need to be clearly defined so that human oversight of clinical AI can be measurable (ClinicalTrials.gov NCT07874659).