2026-10-23
Dresden neurosurgery randomized 300 adults to a checked AI lay summary or usual care. Median visit time fell from 33 to 25 minutes; mean interaction score rose from 4.25 to 4.42.
Imaging reports and clinic letters are written for the next clinician. A first neurosurgical visit has to fit the finding, the options, and a decision into one slot. When health literacy and jargon get in the way, understanding often arrives on the way home.
The trial asks what changes if that explanation moves to before the patient sits down. Does a checked plain-language summary improve the conversation, and does the visit get shorter? Most earlier evidence stops at model scores or at clinicians saying a draft looks usable. What was missing was a randomized test that hands the summary to real patients and then measures the visit itself.
The NEJM AI full text sat behind an access check. Everything below comes from the published abstract. Prompts, model version, the checklist, the baseline table, and subgroups are not in the text that was retrieved.
The trial ran in outpatient neurosurgery at University Hospital Carl Gustav Carus, TUD Dresden University of Technology. Three hundred adults were randomized 1:1 to usual care, or to usual care plus a quality-checked lay summary of their own findings before the visit. The product is named Simply Onno. The trial is AI-INFOCARE, published in NEJM AI in October 2026.
The primary end point is patient-rated quality of the physician-patient interaction on the FAPI (Fragebogen zur Arzt-Patient-Interaktion), filled in right after the visit. Secondary end points are autonomy support on the Health Care Climate Questionnaire (patients rate it, and physicians rate their own support), encounter difficulty on the 10-item Difficult Doctor-Patient Relationship Questionnaire, and consultation length in minutes.
The abstract calls the summary quality-checked, so a review step sits between generation and the patient. It does not say who reviewed, how many drafts were sent back, or whether patients actually read them.
The preregistered protocol for the same trial, published in JMIR Research Protocols in 2025, describes the generator as a Claude-based Simply Onno system and a checklist comparison against the source document. That is the protocol, not the results paper. The results abstract keeps the product name and the phrase "quality-checked."
| End point | Intervention vs control | Difference |
| Mean FAPI | 4.42±0.51 vs 4.25±0.54 | +0.17 (95% CI 0.05 to 0.29, P=0.005, Hedges' g=0.33) |
| Shared decision-making | arm means not in the abstract | +0.29 (95% CI 0.15 to 0.42) |
| Perceived empathy | arm means not in the abstract | no evidence of a difference |
| Patient-rated autonomy support | arm means not in the abstract | +0.14 (95% CI 0.06 to 0.23) |
| Physician-rated autonomy support | arm means not in the abstract | −0.12 (cluster-robust 95% CI −0.21 to −0.04) |
| Median consultation time | 25 vs 33 minutes | geometric-mean ratio 0.79 (95% CI 0.70 to 0.88), about 21% shorter |
| Encounter difficulty, DDPRQ-10 | arm means not in the abstract | +0.09 (cluster-robust 95% CI −0.03 to 0.21) |
The FAPI gap is 0.17 points. Both arms already sit above 4.2, with standard deviations near 0.5, so 0.17 is about a third of that spread and lines up with the reported g of 0.33. The gain is in shared decision-making. Empathy does not separate.
Time is the harder number. The median is 8 minutes shorter. The "about 21%" in the abstract is the geometric-mean ratio of 0.79, whose interval excludes 1. It is not the same figure as 25/33. The abstract gives no P value for duration. Secondary confidence intervals are not adjusted for multiplicity.
Patients rated autonomy support higher. Physicians rated their own autonomy support lower. Encounter difficulty ticks up, and the interval crosses zero. The abstract's own reading is that evidence for a harder visit is limited and compatible with no difference.
Eight minutes off the median is a scheduling number, not a questionnaire decimal. A first neurosurgical visit spends much of its time on the scan and on whether to operate. That time can become another patient, or it can stay on the decision.
The quality gain sits in whether the patient was pulled into the decision, not in whether the physician felt more empathic. Empathy was high in both arms. Visits got shorter while shared decision-making scores went up. A prepared patient was expected to ask more and hold the room. On the median here, that did not happen.
The usable setting is a plain-language pass over the patient's own report, with a check after generation. The result is incremental. The abstract says clinical outcomes and clinician workload still need confirmation before routine adoption.
The abstract names three boundaries. Clinical outcomes were not measured, so a better interaction score is not evidence of a better treatment decision. Total clinician workload was not shown to fall. A shorter visit does not include drafting, checking, and answering what the summary stirred up. Secondary intervals are unadjusted, so shared decision-making and the HCCQ should not be read with the same confidence as the primary P value.
The setting is one outpatient neurosurgery clinic. Nothing in the abstract carries the result to primary care, discharge conversations, or other languages. Both FAPI means already exceed 4.2, and the abstract never states the scale maximum. If the instrument is near its ceiling, 0.17 means less clinically than P=0.005 suggests.
The full text was not retrieved, so several design points cannot be checked in the results paper: whether patients finished the summary, whether allocation was concealed, and whether physicians knew the assignment. If they knew, a shorter visit can mix in skipping the basics. Baseline balance and intention-to-treat versus per-protocol analysis are absent from the abstract.
The protocol says neither side can be blinded, that ratings may carry a novelty effect from an extra document, and that consultation time was the objective check. Time did fall, in the same direction as the subjective scores. The abstract still cannot rule novelty out, and it does not report a factual error rate for the generated text.
Witold H. Polanski is a coauthor. Crossref still lists his affiliation as the same neurosurgery department. The protocol discloses a relationship with Simply Onno, says potential bias cannot be excluded, and plans for the vendor to stay out of randomization, allocation, and outcome assessment. The results abstract does not restate the conflict or say whether those safeguards held.