← Back to A German hospital pilots an on-premise Mistral Small 3.2 system from myScribe that drafts discharge summaries from FHIR patient records for clinician review

Source proof for A German hospital pilots an on-premise Mistral Small 3.2 system from myScribe that drafts discharge summaries from FHIR patient records for clinician review

Source-bound proof

Verified source excerpts for every supported field

Each colour maps a published value to the exact source passage used to support it. Only bounded excerpts are public; administrators can inspect the complete captured source.

13 fields supported

Title

A German hospital pilots an on-premise Mistral Small 3.2 system from myScribe that drafts discharge summaries from FHIR patient records for clinician review

quote · high
…s by ground- ing descriptions in the actual patient record. Deployment Context. Our study is part of an ongoing pilot deployment in a German hospital. The LLM runs fully on-premise within the hospi- tal’s secure environment under…

Description

An ongoing pilot in a German hospital runs Mistral Small 3.2 fully on-premise under GDPR constraints to draft discharge summaries from structured FHIR patient data. Clinicians edit and verify every draft before it enters official documentation. The EACL 2026 industry-track paper compares four input encodings on 31 patients from the Internal Medicine department, with clinician ratings; JSON encoding performed best.

quote · high
…Context. Our study is part of an ongoing pilot deployment in a German hospital. The LLM runs fully on-premise within the hospi- tal’s secure environment under GDPR1 constraints. The available hardware and privacy requirements rule out cloud-based or closed-…

Company

Undisclosed German hospital (system by myScribe GmbH)

quote · medium
…and Analysis of a Clinical Summarization System for a German Hospital Moiz Rauf myScribe Gmbh, Germany m.rauf@myscribe.de Sean Papay Fundamentals of Natural Language Processing Unive…

Country

Germany

quote · high
…tions in the actual patient record. Deployment Context. Our study is part of an ongoing pilot deployment in a German hospital. The LLM runs fully on-premise within the hospi- tal’s secure environment under…

Industry

Healthcare

classification · high
…early and cohesively. In this paper, we present the design and development of a clinical summarization system integrated into a live German hospital workflow to help with the generation of discharge summaries. We first describe the syste…

Problem

At the end of hospital treatment the many electronic health records created for a patient must be summarised into a discharge summary, which the source calls a time-intensive bottleneck for doctors.

quote · high
…rrative that details inpatient treatment and helps in outpatient care. However, manually writ- ing detailed hospital course summaries remains a time-intensive bottleneck for doctors (Arndt et al., 2017; Overhage and McCallie Jr, 2020; Alissa et al., 2022). This…

Solution

A zero-shot system feeds structured FHIR patient data, encoded as text, to an out-of-the-box Mistral Small 3.2 running fully on-premise in the hospital to draft German-language discharge summaries. Clinicians edit and verify every draft before it enters official documentation.

quote · high
…d within the hospital. As such, based on work from (Sivarajkumar et al., 2024), we develop a zero-shot inference system, where a fixed natural-language prompt and the encoded patient record are provided directly to an out-of-the-box Mistral Small 3.2. This allows us to perform local inference on modest hardware without any initia…

Business value

The source frames this as an ongoing pilot and reports no cost or time savings. In an exploratory evaluation on 31 patients, JSON input encoding gave the strongest overall clinician-rated summary quality; the clinicians also found the system often failed to formulate correct follow-up steps, producing hallucinated text.

quote · high
…g events relevant to current medical discourse remain a challenge. Sim- ilarly, the system often failed to formulate correct follow-up steps; this information was not provided in the FHIR source data, often resulting in halluci- nated text. 6 Conclusion In this work, we presented a case study of a de- ployed EHR summar…

AI capabilities

Generative AI

classification · high
…much higher than for general-domain tasks. In this work, we present and analyze an LLM- based clinical summarization system as currently deployed in a German hospital. We make contribu- tions of two kinds: Firstly, we present our deployed model a…

Technology

Mistral Small 3.2

classification · high
…are detailed in Appendix B.3 4.4 Summarization Model As text-generation model, our system uses Mistral Small 3.2, a multilingual decoder-only transformer language model with 128K token context. For zero- shot inference, we sample texts from our model using parameters detai…

Deployment options

On-premise

classification · high
…Context. Our study is part of an ongoing pilot deployment in a German hospital. The LLM runs fully on-premise within the hospi- tal’s secure environment under GDPR1 constraints. The available hardware and privacy requirements rule out cloud-based or closed-…

Headline outcome

derived · medium
…emingly minor structural changes result in sub- stantial performance variation. In terms of overall performance, JSON formatting is highly effective, presumably as its explicit hierarchical organiza- tion that aligns well with th…

Use case type

Content generation

classification · high
…e summary, describing the course of care clearly and cohesively. In this paper, we present the design and development of a clinical summarization system integrated into a live German hospital workflow to help with the generation of discharge summaries. We first describe the system, its components, and its context of use within a h…
Capture details
Captured
19 Nov 2026, 06:03 UTC
Extractor
unpdf@1
Snapshot hash
b898fca1441aadb4736245bbad4ae38f264fc19e4728ae8bbc59e101b5c22bba