The common approach to clinical AI has been to take a general-purpose foundation model and add a thin layer of medical terminology. It is fast to market, but it carries trade-offs: patient data transits external model providers, the training corpus reflects clinical conventions from elsewhere, and the model’s reasoning is hard to audit at the domain level.
A Copenhagen team has shown the alternative works. In April 2026, Corti shipped Symphony for Medical Coding — a domain-specific model for ICD-10 classification trained on European patient data. On clinical-coding accuracy it outperforms the leading general-purpose models by a wide margin, and a Danish study found it identified several times more uncoded suicide-attempt events than human coders working under time pressure — a safety-critical capability that general-purpose systems miss because they lack the clinical context to recognise implicit indicators.
The deployment model is the part worth noting. Through a partnership with Dedalus, one of Europe’s largest hospital-information-system providers, Symphony runs inside existing hospital infrastructure. Patient data is processed within the institution’s own perimeter; nothing leaves the building.
Why this matters for Swiss healthcare
Two lessons follow for Swiss hospital groups, cantonal health services and MedTech companies.
The first is that domain-specific European models now match or surpass general-purpose imports on the metrics that matter clinically — accuracy, recall on safety-critical events, and auditability.
The second is architectural: superior accuracy and in-perimeter processing are not in tension. A model designed for European clinical practice can run where the data already lives, with no external API dependency and no cross-border transfer.
For workloads under both the EU AI Act and the Medical Device Regulation, that combination — domain accuracy plus physical data residency — is the cleaner compliance path. It runs through models built for European clinical practice, not adapted from elsewhere.