University Medical Center Freiburg (Eye Center)
Freiburg University eye clinic trials an in-house AI that suggests discharge letter endings
Doctors at the Freiburg University eye clinic used a 7-billion-parameter language model, trained on the hospital's own letters and run on its own servers, to suggest the closing section of discharge letters. In a 12-month trial with 27 junior doctors, users improved their writing speed more than non-users, and revision requests fell slightly among users.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
39.5
± 20.2 vs. 46.7 ± 19.5 WPM
Writing speed of the clinicians who used the AI, before and after (words per minute, mean ± SD)1
41.8
± 20.3 vs. 44.4 ± 20.0 WPM
Writing speed of the control group, before and after (words per minute, mean ± SD)1
p = 0.045
Statistical test of whether efficiency improved more among AI users than controls (group exposure interaction)1
1.36
± 0.14 to 1.26 ± 0.08
Revisions requested by senior physicians per document, AI users, before and after (mean ± SD)1
The problem
Hospital doctors spend a large share of their working time in electronic health records. Much of the writing happens after hours. At the Freiburg University eye clinic, junior doctors write the closing part of each discharge letter after the consultation, and senior physicians then review it and send it back for revision if needed. The clinic wanted to cut the writing effort without adding review work.
What they built
The team built an autocomplete-style assistant inside the clinic's record system. While a doctor types the closing section of a discharge letter, it proposes the next one or two sentences as grey text. The doctor accepts it with a keystroke or keeps typing to override it. Doctors switched it on with a checkbox.
The assistant is a language model with about 7 billion parameters, called BLOOM-CLP-German, adapted to the clinic with a lightweight tuning method (QLoRA) on about 80,000 anonymized past discharge letters from the same hospital. It uses only the patient's existing record as input.
How they did it
The clinic ran a non-randomized controlled trial, meaning doctors chose for themselves whether to use the tool. A 4-month baseline was followed by a department-wide trial phase, then a comparison between clinicians who kept using it and those who did not. The study ran from January 30, 2024 to February 5, 2025 and covered 27 junior physicians and 15,615 documents. Training and use ran on the hospital's own servers. The study had ethics approval and the participants gave consent.
Writing speed was measured as active editing time in the record system, converted to words per minute. Quality was tracked through the number of revisions that senior physicians requested per document.
Results
Eleven clinicians used the AI for 44.7 to 89.5% of their documentation. Their writing speed went from 39.5 to 46.7 words per minute, against 41.8 to 44.4 in the control group. A statistical model found the difference in improvement significant (p = 0.045).
Revision requests per document fell from 1.36 to 1.26 among AI users and were 1.27 and 1.23 in the control group. The authors report no significant difference between the groups on revisions, but conclude that review burden did not rise. On average, doctors accepted 30.9% of suggestions.
What to copy
Start with one narrow, well-defined piece of a document, keep the human as author, and keep your existing review step so you can see whether quality holds. Run the trial in live workflows with the real record system and track both speed and rework.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Writing speed of the clinicians who used the AI, before and after (words per minute, mean ± SD) | 39.5 ± 20.2 vs. 46.7 ± 19.5 WPM1 | |
| Writing speed of the control group, before and after (words per minute, mean ± SD) | 41.8 ± 20.3 vs. 44.4 ± 20.0 WPM1 | |
| Statistical test of whether efficiency improved more among AI users than controls (group exposure interaction) | p = 0.0451 | |
| Revisions requested by senior physicians per document, AI users, before and after (mean ± SD) | 1.36 ± 0.14 to 1.26 ± 0.081 | |
| Revisions requested per document, control group, before and after (mean ± SD) | 1.27 ± 0.16 vs. 1.23 ± 0.121 | |
| Average share of AI suggestions accepted, fully or partially (range 0.02% to 100% across clinicians) | 30.9%1 |
Lessons for your project
Suggest, do not write. The tool only offered one or two sentences as grey text that doctors could accept or type over, so they kept control of every letter.
Keep existing review in place. Senior physicians kept checking every letter, which let the clinic watch for quality problems while testing speed.
Use the data you already have. The model needs only the patient's record as input, with no audio recordings of consultations.
Expect uneven uptake. Acceptance of suggestions ranged from 0.02% to 100% and 13 of the 27 clinicians stopped after the first trial, so measure adoption as well as speed.
Keep patient data in-house. Training and use ran on the hospital's own servers, which the authors say avoided third-party data transfers.
Caveats
- This is a peer-reviewed paper by Freiburg authors about their own clinic. No independent outlet confirms the figures, so the case is company-reported.
- Doctors chose whether to use the tool, so the groups may differ in ways the study could not measure. The authors note that AI users started with lower writing speeds and had a more positive attitude to AI.
- The reduction in revisions was seen within the AI group only. The authors report no statistically significant difference between groups for revision counts.
- The speed result has a p-value of 0.045, close to the usual 0.05 threshold. The authors say results should not be read as proof that language models are ready for unrestricted clinical use.
- The study covers only the closing section of discharge letters and only junior physicians in one clinic.
- The paper gives the study period (January 2024 to February 2025) with a 4-month baseline before AI was switched on, but no exact go-live date. The deployed date is therefore shown as the year only.
- Control group: 16 of the 27 clinicians, including 13 who tried the AI and stopped. The paper's abstract lists 17 controls and 11 AI users, while the results text lists 16 controls and 11 AI users.
- The paper refers to an earlier publication for the model's development and evaluation. This case rests only on the trial paper, which reports use in the live record system with real documents.
Sources
- Implementation outcomes of AI documentation support in routine clinical practice: A non-randomized controlled trial
PLOS Digital Health · 2026-07-23 · Research paper · journals.plos.org
Summarized in our own words from the sources above. Spotted an error? See corrections.