Blog

Why Clinical AI Doesn’t Care About Your Hochdeutsch

20 January 2026 By SwissMed AI
Why Clinical AI Doesn’t Care About Your Hochdeutsch

The practical answer: use a hybrid workflow

  • Input: your clinical working language (DE/FR/IT)

  • Process instruction: “Analyze using international medical literature and standard medical terminology.”

  • Output: the language you need for documentation and handoffs

Why this works: it separates reasoning from documentation clarity and keeps your prompts auditable.

Example: 03:00 admission note in Swiss German → prompt “Summarize using standard medical terminology” → structured output in German for your clinical information system.

Four key facts

1) English is the default knowledge substrate

As of December 2023, 86.5% of PubMed-indexed publications were in English.
That matters because training and evaluation data often reflect what is most available.

2) “English is always better” is not a fact

Multilingual performance varies by model and task, so claims should be benchmark-based, not anecdotal. Medical multilingual benchmarks like MedExpQA exist specifically to measure cross-language differences.
Takeaway: English is often a reasonable default for complex reasoning, but it is not a safety guarantee.

3) The bigger risk is messy input

What reliably worsens outputs is not “German vs English”, it is ambiguity and noise:

  • unstructured histories (e.g. “pt unwell since days, maybe fever?, meds??, unclear past medical history”)

  • inconsistent terminology (e.g. mixing “Luftnot”, “dyspnea”, “can’t breathe” without onset or severity; “renal failure” without stage or creatinine)

  • copy-pasted patient messages (e.g. long chat text: “I feel weird… heart skipping… since yesterday… btw could I be pregnant??”)

  • mixed languages in one prompt (e.g. “Anamnese: seit 3 Tagen Fieber. Triage: douleur thoracique 7/10. DDx: PE vs pneumonia. Angehörige: ‘sie isch hüt mega komisch gsi’.”)

The Swiss trap: “copy-paste soup” (note fragments, patient messages, and dialect) is not a prompting strategy.
Rule: normalize, structure, and use clinical terms before asking for reasoning.

4) Swiss-specific tools are emerging, with caveats

  • AlpineAI’s SwissGPT: AlpineAI states SwissGPT can be used in all Swiss national languages and in English. But language support is not the same as clinically validated performance.

  • Apertus (EPFL/ETH/CSCS, Swiss AI Initiative): Apertus is presented as a large-scale open, multilingual Swiss model trained on 15 trillion tokens across 1,000+ languages, including Swiss German and Romansh. At the same time, Apertus is primarily a general-purpose LLM, and its medical usefulness depends on domain adaptation, clinical evaluation, and appropriate governance.

Bottom line

  • Routine work: local language is fine.

  • Complex reasoning: English is often reasonable, not guaranteed better.

  • Always required: clinical judgment, verification, and missing-data checks.

LLM outputs are decision support, not decisions.

Sources