Reviewed AI summaries shortened consultations in a 300-patient trial

Patients given a quality-checked AI explanation of their medical findings before a neurosurgical consultation rated the interaction modestly better and spent less time in the appointment in a 300-person randomised trial. Median consultation duration was 25 minutes with the summary and 33 minutes under standard care. The German study tested a reviewed explanation added to a clinician-led visit, with communication quality as its primary outcome.[1]

The AI-INFOCARE study, published in NEJM AI's November issue after an October 15 article date, randomly allocated 150 adults to receive a lay-language previsit summary and 150 to standard care. The summary was produced by the Simply Onno system and quality-checked before patients received it. The intervention therefore consisted of an AI-assisted, reviewed explanation added to the existing clinical process. Its benefit cannot be attributed to an unchecked chatbot operating independently, nor separated cleanly from the review and delivery arrangements that made the explanation safe to use.[1]

The primary outcome and the appointment clock

The primary measure was the patient-reported quality of the physician-patient interaction, evaluated with a standardised questionnaire. On its reported scale, the mean was 4.42 in the intervention group versus 4.25 in the control group: an absolute difference of 0.17 points with a 95% confidence interval of 0.05 to 0.29. The researchers reported a P value of 0.005 and a standardised effect size of 0.33. The difference is statistically supported but should be described as modest rather than transformative. An average score of 4.42 is not a percentage of patients satisfied or a probability of correct treatment.[1]

The largest component difference concerned shared decision-making, at 0.29 points, while patients also reported slightly greater autonomy support. There was no clear difference in perceived empathy. One plausible explanation is that a previsit summary frees more of the conversation for questions and choices by explaining terminology in advance. That is an interpretation of the outcome pattern; the trial did not isolate which element of the preparation caused the change.[1]

Consultation length was a secondary measure. The median duration was 25 minutes with AI-prepared summaries and 33 minutes without them; the ratio of geometric means was 0.79, with a 95% confidence interval of 0.70 to 0.88. The investigators interpreted this as roughly a 21% relative reduction under the relevant time model. The raw median difference is eight minutes, but comparing medians and geometric means as if they were the same statistic would misstate the calculation. A shorter consultation is valuable only if patients still understand their options and clinicians can assess them adequately.[1]

The outcome physicians viewed differently

Physicians viewed part of the encounter differently. Their ratings of the autonomy support they provided were 0.12 points lower in the intervention arm, with a cluster-adjusted 95% confidence interval from −0.21 to −0.04. That contrast with patients' ratings is a reason to examine how the summary changes the consultation, rather than equating shorter appointments with an easier workflow for doctors.[1]

The estimate for encounter difficulty had a confidence interval including no effect, and secondary outcomes were not adjusted for multiple comparisons. The trial did not establish better surgical outcomes or long-term understanding. Its single-centre neurosurgical setting also leaves open how the approach performs in other specialties and languages. The accessible abstract does not quantify the staff time needed to quality-check summaries, so the shorter appointment cannot by itself establish a net saving in clinical workload.[1]

The measured benefit is specific enough to guide a follow-up: provide a reviewed explanation before the visit, then assess communication, appointment time and the work of preparing and checking the summary together. Replicating that comparison elsewhere would show whether a modest patient-reported gain can survive different medical material and delivery arrangements.[1]

Report an error or suggest a correction