Truace
Subscribe

Patient representatives preferred AI-simplified geriatric reports over doctors' versions in a 64-report study

Study title: Patient-Friendly Simplifications of Outpatient Reports Using Large Language Models via Few-Shot Prompting: Blinded Comparative Evaluation

A study published in the Journal of Medical Internet Research tested whether Google's open-source Gemma-3-27B-it model could rewrite 64 deidentified German geriatric outpatient reports in patient-friendly language. Patient representatives preferred the AI versions to those written by two medical experts (P<.001), as did one of two expert reviewers (P=.001).

Read the study abstract

Background: Effective communication between clinicians and patients plays an important role in the successful delivery of health care. In geriatric medicine, where multimorbidity, polypharmacy, and functional limitations often make treatment plans complex, AI simplifications of clinical reports into patient-friendly language offer a promising way to improve communication. Objective: This study investigated whether AI could generate patient-friendly simplifications of outpatient reports that were comparable to th…

Truvace record2 min read1 sourceEvidence: Moderate · 64/100 — Controlled or validated comparison · n = 64 · 1 independent sourcePeer-reviewedTRV-2026-2427 · cite or check it
Patient-Friendly Simplifications of Outpatient Reports Using Large Language Models via Few-Shot Prompting: Blinded Comparative Evaluation"Medical laboratories and clinics in Kinwat" by Mouryan, CC BY-SA 4.0.
How the two sides scoreContested
P 54P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.

A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.

G 56G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.

In brief

The AI texts matched the geriatrician-written ones on semantic similarity to the original report (P=.02). Expert reviewers, though, judged the physician versions more complete (P<.001), and that perception correlated with text length.

Geriatric care involves multimorbidity, polypharmacy and functional limitations that make treatment plans complex, so clear communication matters. The authors conclude the approach is feasible for improving perceived comprehension, and say prospective studies must test effects on objective health outcomes, adherence and patient engagement.

Main points

  1. The study used Gemma-3-27B-it, an open-source instruction-tuned model from Google, with few-shot prompting.
  2. It covered 64 deidentified geriatric outpatient reports written in German.
  3. AI simplifications were compared with those produced by 2 medical experts in a blinded evaluation.
  4. Quality was judged by semantic similarity, review by 2 clinicians and review by patient representatives, the latter two with a Likert-scale questionnaire.
  5. Patient representatives preferred the AI versions (P<.001); one expert reviewer did too (P=.001).
  6. Semantic similarity to the original report was comparable between AI and geriatrician versions (P=.02).
  7. Expert reviewers rated physician-authored versions as more complete (P<.001), correlating with text length.

The gain

In a blinded evaluation of 64 German geriatric outpatient reports, patient representatives (P<.001) and one expert reviewer (P=.001) preferred Gemma-3 simplifications over those written by geriatricians.

The problem

In the same study, expert reviewers rated physician-written simplifications as more complete than the AI versions (P<.001), a gap that tracked with text length.

The full story

Researchers applied few-shot prompting to Gemma-3-27B-it so it would produce patient-friendly versions of 64 deidentified German geriatric outpatient reports while preserving clinical meaning. Two medical experts wrote their own simplifications of the same reports for comparison.

Evaluation was blinded and used three approaches: embedding-based semantic similarity to the original report, review by 2 clinicians, and review by patient representatives. The clinician and patient reviews used a standardized Likert-scale questionnaire.

Patient representatives preferred the AI-generated texts (P<.001), as did one expert reviewer (P=.001). On semantic similarity the AI texts were comparable to the geriatricians' versions (P=.02).

Expert reviewers found the physician-authored simplifications more complete (P<.001), which correlated with text length. The authors call for prospective studies on objective health outcomes, adherence and patient engagement.

What this doesn’t fix

The study measured perceived comprehension and preference, not objective health outcomes, adherence or patient engagement, which the authors say prospective studies must still examine.

Sources

  1. Peer-reviewedJournal of Medical Internet Research2026-11-06

The debate